Kubernetes

14 results back to index


pages: 355 words: 81,788

Monolith to Microservices: Evolutionary Patterns to Transform Your Monolith by Sam Newman

Airbnb, business process, continuous integration, database schema, DevOps, fault tolerance, ghettoisation, inventory management, Jeff Bezos, Kubernetes, loose coupling, microservices, MVC pattern, price anchoring, pull request, single page application, software as a service, source of truth, telepresence

Not all problem spaces are equal, obviously, but I feel that if you’re already on the public cloud, you may not always need the complexity of a container-based platform like Kubernetes. Warning I do see people reaching for Kubernetes and the like a bit too early in the process of adopting microservices, often assuming it is a prerequisite. Far from it—platforms like Kubernetes excel at helping you manage multiple processes, but you should wait until you have enough processes that your current approach and technology are starting to strain. You might find that you need only five microservices, and that you can happily handle this with your existing solutions—in which case, great! Don’t adopt a Kubernetes-based platform just because you see everyone else doing it, which can also be said for microservices!

Potential Solutions You want a tool that allows for a high degree of automation, that can allow developers ideally to self-service provision deployments, and that handles automated desired state management. For microservices, Kubernetes has emerged as the tool of choice in this space. It requires that you containerize your services, but once you do, you can use Kubernetes to manage the deployment of your service instances across multiple machines, ensuring you can scale to improve robustness and handle load (assuming you have enough hardware). Vanilla Kubernetes isn’t what I would consider developer-friendly. A multitude of people are working on higher-order, more developer-friendly abstractions, and I expect that work to continue.

A multitude of people are working on higher-order, more developer-friendly abstractions, and I expect that work to continue. In the future, I expect that many developers who are running software on Kubernetes won’t even realize, as it will just become an implementation detail. I tend to see larger organizations adopt a packaged version of Kubernetes, such as OpenShift from RedHat, which bundles Kubernetes with tooling that makes it easier to work with within a corporate environment—perhaps handling corporate identity and access management controls. Some of these packaged versions also provide simplified abstractions for developers to work with. If you’re lucky enough to be on the public cloud, you could use the many different options there to handle deployments of your microservice architecture, including managed Kubernetes offerings.


pages: 302 words: 73,946

People Powered: How Communities Can Supercharge Your Business, Brand, and Teams by Jono Bacon

Airbnb, barriers to entry, blockchain, bounce rate, Cass Sunstein, Charles Lindbergh, Debian, Firefox, if you build it, they will come, IKEA effect, Internet Archive, Jono Bacon, Kickstarter, Kubernetes, lateral thinking, Mark Shuttleworth, Minecraft, minimum viable product, more computing power than Apollo, planetary scale, pull request, Richard Stallman, Richard Thaler, sexual politics, Silicon Valley, Travis Kalanick, Y Combinator

“Market Brief—2018 Digital Games & Interactive Entertainment Industry Year in Review,” SuperData Research, accessed May 25, 2018, https://www.superdataresearch.com/market-data/market-brief-year-in-review/. 11. IGN Boards, accessed May 25, 2018, http://www.ign.com/boards/. 12. Alexander van Engelen, interview with Jono Bacon via forum private message, May 10, 2018. 13. “Kubernetes/Kubernetes: Production-Grade Container Scheduling and Management,” GitHub, accessed March 2, 2019, https://github.com/kubernetes/kubernetes. 14. Gilbert Schacter et al., Psychology (New York: Worth Publishers, 2011), 295. 15. L. Chapman, and E. Newcomer, “Software Maker Docker Is Raising Funding at $1.3 Billion Valuation,” Bloomberg, accessed May 10, 2018, https://www.bloomberg.com/news/articles/2017-08-09/docker-is-said-to-be-raising-funding-at-1-3-billion-valuation; Red Hat Fact Sheet, https://investors.redhat.com/~/media/Files/R/Red-Hat-IR/documents/q218-fact-sheet.pdf; “Red Hat Reports Fourth Quarter and Fiscal Year 2018 Results,” RedHat, March 26, 2018, https://investors.redhat.com/news-and-events/press-releases/2018/03-26-2018-211600973. 16.

Here enthusiastic participants don’t just add independent pieces of work to the stockpile; they actively work together as a team on shared projects. This can unlock some quite literally world-changing opportunity. On June 7, 2014, a new open-source project called Kubernetes was announced. It was a piece of software that could be used for managing how software services run on the cloud. I won’t bore you too much with what Kubernetes does, but safe to say, it rocked the tech and enterprise world. A critical element of why Kubernetes succeeded is that it is open-source. This means that its code is freely available and when there are gaps in functionality, or bugs that cause problems for users, there is a way in which anyone (who meets certain guidelines) can fill in these gaps and create these additional features or fixes.

“Our method of working across multiple open-source communities hasn’t just allowed us to survive; it’s enabled us to actually thrive as new technological shifts have occurred. Our innovative technologies are an output of our organizational culture—our people—who give us the ability to adapt and rebound in the wake of disruptive change.”29 Red Hat doesn’t operate alone. The global open-source community has produced tools such as Linux, Kubernetes, OpenStack, Apache, Debian, Jenkins, GNOME, and others that have had a profound impact on various industries, powering clouds, devices, vehicles, space shuttles, and more. Why do these companies invest in creating code that is shared freely? There are many reasons. Open code means you can attract new contributors who produce features, fix bugs, and improve the overall security of a product.


pages: 1,380 words: 190,710

Building Secure and Reliable Systems: Best Practices for Designing, Implementing, and Maintaining Systems by Heather Adkins, Betsy Beyer, Paul Blankinship, Ana Oprea, Piotr Lewandowski, Adam Stubblefield

anti-pattern, barriers to entry, bash_history, business continuity plan, business process, Cass Sunstein, cloud computing, continuous integration, correlation does not imply causation, create, read, update, delete, cryptocurrency, cyber-physical system, database schema, Debian, defense in depth, DevOps, Edward Snowden, fault tolerance, fear of failure, general-purpose programming language, Google Chrome, Internet of things, Kubernetes, load shedding, margin call, microservices, MITM: man-in-the-middle, performance metric, pull request, ransomware, revision control, Richard Thaler, risk tolerance, self-driving car, Skype, slashdot, software as a service, source of truth, Stuxnet, Turing test, undersea cable, uranium enrichment, Valgrind, web application, Y2K, zero day

Adversaries can bypass deployment decisions that don’t occur at choke points. Consider Kubernetes as an example for setting up deployment choke points, as shown in Figure 14-7. Suppose you want to verify all deployments to the pods in a specific Kubernetes cluster. The master node would make a good choke point because all deployments are supposed to flow through it. To make this a proper choke point, configure the worker nodes to accept requests only from the master node. This way, adversaries cannot deploy directly to worker nodes.19 Figure 14-7. Kubernetes architecture—all deployments must flow through the master Ideally, the choke point performs the policy decision, either directly or via an RPC.

Kubernetes architecture—all deployments must flow through the master Ideally, the choke point performs the policy decision, either directly or via an RPC. Kubernetes offers an Admission Controller webhook for this exact purpose. If you use Google Kubernetes Engine, Binary Authorization offers a hosted admission controller and many additional features. And even if you don’t use Kubernetes, you may be able to modify your “admission” point to perform the deployment decision. Alternatively, you can place a “proxy” in front of the choke point and perform the policy decision in the proxy, as shown in Figure 14-8. This approach requires configuring your “admission” point to allow access only via the proxy.

A deployment is any assignment of some artifact to some environment. You can consider each of the following to be a deployment: Pushing code: Issuing a command to cause a server to download and run a new binary Updating a Kubernetes Deployment object to pick up a new Docker image Booting a VM or physical machine, which loads initial software or firmware Updating configuration: Running a SQL command to change a database schema Updating a Kubernetes Deployment object to change a command-line flag Publishing a package or other data, which will be consumed by other users: Uploading a deb package to an apt repository Uploading a Docker image to a container registry Uploading an APK to the Google Play Store Post-deployment changes are out of scope for this chapter.


pages: 461 words: 106,027

Zero to Sold: How to Start, Run, and Sell a Bootstrapped Business by Arvid Kahl

"side hustle", business process, centre right, Chuck Templeton: OpenTable:, continuous integration, coronavirus, COVID-19, Covid-19, crowdsourcing, domain-specific language, financial independence, Google Chrome, if you build it, they will come, information asymmetry, information retrieval, inventory management, Jeff Bezos, job automation, Kubernetes, minimum viable product, Network effects, performance metric, post-work, premature optimization, risk tolerance, Ruby on Rails, sentiment analysis, Silicon Valley, software as a service, source of truth, statistical model, subscription business, supply-chain management, trickle-down economics, web application

Message Queues are useful to decouple the request from the actual work. Using containers also allows you to test and quickly deploy new versions of your product without much hassle. In a Google Kubernetes Engine setup, updating the running version of your product is a single change to a configuration file. Reverting to the previous version is just as easy. When I set up the infrastructure for FeedbackPanda on the Google Cloud, I liked that part the most. With a simple edit to a config file and a shell command, the Kubernetes cluster would fetch the new container from the container repository, start it, and reroute traffic to the new instance. If there were any errors in the container, it would revert to the previously running system.

Three main things are important to create a system capable of automated rollbacks: artifacts, versioning, and bidirectional migrations. Package your releases into easy-to-deploy artifacts. An artifact is a bundle, a package that can be easily copied and run on a server, either as some sort of executable or as a container that systems such as orchestration tools like Kubernetes can manage. The idea is that everything is well-specified, and you don't need to move around files or assets. Usually, this involves a build process, where your whole application gets compiled, optimized, and packed up so that a final artifact can be created. Artifacts are usually idempotent with the sources that created them: compiling your artifact from the same code twice will result in the exact same artifact.

The more it looks like a collection of simple building blocks, the better. I've fallen into the specialist trap before. When we started FeedbackPanda, I was looking for a cloud hosting provider capable of orchestrating Docker containers. I thought the big cloud providers were not specific enough with their hosted Kubernetes offerings, so I picked a small infrastructure startup that promised to allow interacting with Docker containers through their dashboards and GUIs instead of having to use command-line calls. I was thrilled and set up our production system on their platform. For a while, it went well, but then, their service experienced a few problems.


Seeking SRE: Conversations About Running Production Systems at Scale by David N. Blank-Edelman

Affordable Care Act / Obamacare, algorithmic trading, Amazon Web Services, bounce rate, business continuity plan, business process, cloud computing, cognitive bias, cognitive dissonance, commoditize, continuous integration, crowdsourcing, dark matter, database schema, Debian, defense in depth, DevOps, domain-specific language, en.wikipedia.org, fault tolerance, fear of failure, friendly fire, game design, Grace Hopper, information retrieval, Infrastructure as a Service, Internet of things, invisible hand, iterative process, Kubernetes, loose coupling, Lyft, Marc Andreessen, microservices, minimum viable product, MVC pattern, performance metric, platform as a service, pull request, RAND corporation, remote working, Richard Feynman, risk tolerance, Ruby on Rails, search engine result page, self-driving car, sentiment analysis, Silicon Valley, single page application, Snapchat, software as a service, software is eating the world, source of truth, the scientific method, Toyota Production System, web application, WebSocket, zero day

Instead of spending time on tasks that don’t give us a competitive advantage, like data center management and hardware configuration, we can shift that problem to cloud providers and benefit from their economies of scale. The second part is around adopting cloud primitives, shifting from bespoke solutions to open source products with vibrant communities. An example of this is our planned move from our homegrown container orchestration system, Helios, to a managed Kubernetes services (Google Kubernetes Engine). In adopting Kubernetes instead of further investing in our own container orchestration system, we can benefit from the many contributions of the open source community. Making these shifts allows the ops teams to focus on higher-level problems facing the organization, thereby delivering more value.

Back then, there was no Docker, containers were fairly arcane stuff to most, and there was obviously no established container orchestration platform available in the open source space. Hence, building Bazooka was a necessity and not a result of a not-invented-here syndrome. Later, we migrated completely to Kubernetes. Bazooka and the Platforms team are history now. Closing the Loop: Take Your Own Pager Now that it was easy and quick to deploy changes and create new services, a new bottleneck became apparent: even though developers could deploy at any time, they had to be mindful and coordinate with those carrying the pager for their systems.

With every company out there scrambling desperately to move to cloud providers and drastically change their cost structure and therefore valuation (cloud providers are variable operational costs, not fixed), and all companies trying desperately to catch up with the latest infrastructure-as-code breakthroughs (Docker and Kubernetes), operational complexity is increasing exponentially. Most companies lack the skill set and engineering culture treats that enable for a healthy balance of forces to release innovation and operate it reliably. I believe SRE is just Google’s, Microsoft’s, LinkedIn’s, Facebook’s and other huge companies’ efforts to tackle the same issue.


pages: 1,136 words: 73,489

Working in Public: The Making and Maintenance of Open Source Software by Nadia Eghbal

Amazon Web Services, barriers to entry, Benevolent Dictator For Life (BDFL), bitcoin, Clayton Christensen, cloud computing, commoditize, continuous integration, crowdsourcing, cryptocurrency, David Heinemeier Hansson, death of newspapers, Debian, disruptive innovation, en.wikipedia.org, Ethereum, Firefox, Guido van Rossum, Hacker Ethic, Induced demand, informal economy, Jane Jacobs, Jean Tirole, Kevin Kelly, Kickstarter, Kubernetes, Mark Zuckerberg, Menlo Park, Network effects, node package manager, Norbert Wiener, pirate software, pull request, RFC: Request For Comment, Richard Stallman, Ronald Coase, Ruby on Rails, side project, Silicon Valley, Snapchat, social graph, software as a service, Steve Jobs, Steve Wozniak, Steven Levy, Stewart Brand, The Death and Life of Great American Cities, The Nature of the Firm, transaction costs, two-sided market, urban planning, web application, wikimedia commons, Zimmermann PGP

They can perform a variety of services, whether it’s being the first to respond to an issue, checking that contributions pass tests and comply with requirements, or updating dependencies. Rust’s Highfive bot, for example, finds the best reviewer for a given pull request then tags them,282 while Kubernetes’s Prow Robot reviews, triages, and merges pull requests.283 Researcher Mairieli Wessel et al. found that out of 351 popular, active projects sampled on GitHub, 26% use bots to manage their workflow.284 While some maintainers worry about bots coming off as impersonal to new contributors, this form of automation seems to be an increasingly common practice: Wessel et al. identify a “boom” in bot adoption starting after 2013.

Kraut and Paul Resnick, Building Successful Online Communities: Evidence-Based Social Design (Cambridge, MA: The MIT Press, 2016), 112. 279 “Dear GitHub,” Dear-Github Code, GitHub, January 10, 2016, https://github.com/dear-github/dear-github. 280 Brandon Keepers (bkeepers), “Dear Open Source Maintainers,” Dear-Github Pull Requests, GitHub, February 12, 2016, https://github.com/dear-github/dear-github/pull/115. 281 Isaacs / Github Code, GitHub, accessed April 25, 2020, https://github.com/isaacs/github. 282 “Rust Highfive Robot,” GitHub, n.d., https://github.com/rust-highfive. 283 “Kubernetes Prow Robot,” GitHub, n.d., https://github.com/k8s-ci-robot. 284 Mairieli Wessel, Bruno Mendes de Souza, Igor Steinmacher, Igor S. Wiese, Ivanilton Polato, Ana Paula Chaves, and Marco A. Gerosa, “The Power of Bots: Characterizing and Understanding Bots in OSS Projects,” Proceedings of the ACM on Human-Computer Interaction 2 (November 2018): 1–19, https://doi.org/10.1145/3274451. 285 “The Configuration Issue to End All Configuration Issues ,” Probot Issues, GitHub screenshot, November 20, 2017, https://github.com/probot/probot/issues/258#issuecomment-345739177. 286 Nadia Eghbal, “Understanding User Support Systems in Open Source,” Nadia Eghbal, September 27, 2018, https://nadiaeghbal.com/user-support. 287 “README,” Youtube-dl, GitHub, accessed March 13, 2020, https://github.com/rg3/youtube-dl. 288 React-native-firebase Issues, GitHub, accessed March 29, 2020, https://raw.githubusercontent.com/invertase/react-native-firebase/d6db2601f62fa35e79957a6f73454e62e85f9714/.github/ISSUE_TEMPLATE/Bug_report.md. 289 Philip Guo, “PG Vlog #75 - Python Tutor Software Development Philosophy,” October 23, 2017, YouTube video, 10:30, https://www.youtube.com/watch?


pages: 719 words: 181,090

Site Reliability Engineering: How Google Runs Production Systems by Betsy Beyer, Chris Jones, Jennifer Petoff, Niall Richard Murphy

Air France Flight 447, anti-pattern, barriers to entry, business intelligence, business process, Checklist Manifesto, cloud computing, combinatorial explosion, continuous integration, correlation does not imply causation, crowdsourcing, database schema, defense in depth, DevOps, en.wikipedia.org, fault tolerance, Flash crash, George Santayana, Google Chrome, Google Earth, information asymmetry, job automation, job satisfaction, Kubernetes, linear programming, load shedding, loose coupling, meta analysis, meta-analysis, microservices, minimum viable product, MVC pattern, performance metric, platform as a service, revision control, risk tolerance, side project, six sigma, the scientific method, Toyota Production System, trickle-down economics, web application, zero day

Some datacenters end up with multiple generations of compute hardware, and sometimes we augment datacenters after they are built. But for the most part, our datacenter hardware is homogeneous. 2 Some readers may be more familiar with Borg’s descendant, Kubernetes—an open source Container Cluster orchestration framework started by Google in 2014; see http://kubernetes.io and [Bur16]. For more details on the similarities between Borg and Apache Mesos, see [Ver15]. 3 See http://grpc.io. 4 Protocol buffers are a language-neutral, platform-neutral extensible mechanism for serializing structured data. For more details, see https://developers.google.com/protocol-buffers/. 5 We assume the probability of two simultaneous task failures in our environment is low enough to be negligible.

Burgess, Principles of Network and System Administration: Wiley, 1999. [Bur06] M. Burrows, “The Chubby Lock Service for Loosely-Coupled Distributed Systems”, in OSDI ’06: Seventh Symposium on Operating System Design and Implementation, November 2006. [Bur16] B. Burns, B. Grant, D. Oppenheimer, E. Brewer, and J. Wilkes, “Borg, Omega, and Kubernetes” in ACM Queue, vol. 14, no. 1, 2016. [Cas99] M. Castro and B. Liskov, “Practical Byzantine Fault Tolerance”, in Proc. OSDI 1999, 1999. [Cha10] C. Chambers, A. Raniwala, F. Perry, S. Adams, R. Henry, R. Bradshaw, and N. Weizenbaum, “FlumeJava: Easy, Efficient Data-Parallel Pipelines”, in ACM SIGPLAN Conference on Programming Language Design and Implementation, 2010.


Martin Kleppmann-Designing Data-Intensive Applications. The Big Ideas Behind Reliable, Scalable and Maintainable Systems-O’Reilly (2017) by Unknown

active measures, Amazon Web Services, bitcoin, blockchain, business intelligence, business process, c2.com, cloud computing, collaborative editing, commoditize, conceptual framework, cryptocurrency, database schema, DevOps, distributed ledger, Donald Knuth, Edward Snowden, Ethereum, ethereum blockchain, fault tolerance, finite state, Flash crash, full text search, general-purpose programming language, informal economy, information retrieval, Internet of things, iterative process, John von Neumann, Kubernetes, loose coupling, Marc Andreessen, microservices, natural language processing, Network effects, packet switching, peer-to-peer, performance metric, place-making, premature optimization, recommendation engine, Richard Feynman, self-driving car, semantic web, Shoshana Zuboff, social graph, social web, software as a service, software is eating the world, sorting algorithm, source of truth, SPARQL, speech recognition, statistical model, undersea cable, web application, WebSocket, wikimedia commons

And this is why MapReduce is designed to tolerate frequent unexpected task termina‐ tion: it’s not because the hardware is particularly unreliable, it’s because the freedom to arbitrarily terminate processes enables better resource utilization in a computing cluster. Among open source cluster schedulers, preemption is less widely used. YARN’s CapacityScheduler supports preemption for balancing the resource allocation of dif‐ ferent queues [58], but general priority preemption is not supported in YARN, Mesos, or Kubernetes at the time of writing [60]. In an environment where tasks are not so often terminated, the design decisions of MapReduce make less sense. In the next section, we will look at some alternatives to MapReduce that make different design decisions. 418 | Chapter 10: Batch Processing Beyond MapReduce Although MapReduce became very popular and received a lot of hype in the late 2000s, it is just one among many possible programming models for distributed sys‐ tems.

They do not fit well with the requirements of modern application development, such as dependency and package management, version control, rolling upgrades, evolvability, monitoring, metrics, calls to network services, and integration with external systems. On the other hand, deployment and cluster management tools such as Mesos, YARN, Docker, Kubernetes, and others are designed specifically for the purpose of running application code. By focusing on doing one thing well, they are able to do it much better than a database that provides execution of user-defined functions as one of its many features. I think it makes sense to have some parts of a system that specialize in durable data storage, and other parts that specialize in running application code.

The opposite of bounded. 558 | Glossary Index A aborts (transactions), 222, 224 in two-phase commit, 356 performance of optimistic concurrency con‐ trol, 266 retrying aborted transactions, 231 abstraction, 21, 27, 222, 266, 321 access path (in network model), 37, 60 accidental complexity, removing, 21 accountability, 535 ACID properties (transactions), 90, 223 atomicity, 223, 228 consistency, 224, 529 durability, 226 isolation, 225, 228 acknowledgements (messaging), 445 active/active replication (see multi-leader repli‐ cation) active/passive replication (see leader-based rep‐ lication) ActiveMQ (messaging), 137, 444 distributed transaction support, 361 ActiveRecord (object-relational mapper), 30, 232 actor model, 138 (see also message-passing) comparison to Pregel model, 425 comparison to stream processing, 468 Advanced Message Queuing Protocol (see AMQP) aerospace systems, 6, 10, 305, 372 aggregation data cubes and materialized views, 101 in batch processes, 406 in stream processes, 466 aggregation pipeline query language, 48 Agile, 22 minimizing irreversibility, 414, 497 moving faster with confidence, 532 Unix philosophy, 394 agreement, 365 (see also consensus) Airflow (workflow scheduler), 402 Ajax, 131 Akka (actor framework), 139 algorithms algorithm correctness, 308 B-trees, 79-83 for distributed systems, 306 hash indexes, 72-75 mergesort, 76, 402, 405 red-black trees, 78 SSTables and LSM-trees, 76-79 all-to-all replication topologies, 175 AllegroGraph (database), 50 ALTER TABLE statement (SQL), 40, 111 Amazon Dynamo (database), 177 Amazon Web Services (AWS), 8 Kinesis Streams (messaging), 448 network reliability, 279 postmortems, 9 RedShift (database), 93 S3 (object storage), 398 checking data integrity, 530 amplification of bias, 534 of failures, 364, 495 Index | 559 of tail latency, 16, 207 write amplification, 84 AMQP (Advanced Message Queuing Protocol), 444 (see also messaging systems) comparison to log-based messaging, 448, 451 message ordering, 446 analytics, 90 comparison to transaction processing, 91 data warehousing (see data warehousing) parallel query execution in MPP databases, 415 predictive (see predictive analytics) relation to batch processing, 411 schemas for, 93-95 snapshot isolation for queries, 238 stream analytics, 466 using MapReduce, analysis of user activity events (example), 404 anti-caching (in-memory databases), 89 anti-entropy, 178 Apache ActiveMQ (see ActiveMQ) Apache Avro (see Avro) Apache Beam (see Beam) Apache BookKeeper (see BookKeeper) Apache Cassandra (see Cassandra) Apache CouchDB (see CouchDB) Apache Curator (see Curator) Apache Drill (see Drill) Apache Flink (see Flink) Apache Giraph (see Giraph) Apache Hadoop (see Hadoop) Apache HAWQ (see HAWQ) Apache HBase (see HBase) Apache Helix (see Helix) Apache Hive (see Hive) Apache Impala (see Impala) Apache Jena (see Jena) Apache Kafka (see Kafka) Apache Lucene (see Lucene) Apache MADlib (see MADlib) Apache Mahout (see Mahout) Apache Oozie (see Oozie) Apache Parquet (see Parquet) Apache Qpid (see Qpid) Apache Samza (see Samza) Apache Solr (see Solr) Apache Spark (see Spark) 560 | Index Apache Storm (see Storm) Apache Tajo (see Tajo) Apache Tez (see Tez) Apache Thrift (see Thrift) Apache ZooKeeper (see ZooKeeper) Apama (stream analytics), 466 append-only B-trees, 82, 242 append-only files (see logs) Application Programming Interfaces (APIs), 5, 27 for batch processing, 403 for change streams, 456 for distributed transactions, 361 for graph processing, 425 for services, 131-136 (see also services) evolvability, 136 RESTful, 133 SOAP, 133 application state (see state) approximate search (see similarity search) archival storage, data from databases, 131 arcs (see edges) arithmetic mean, 14 ASCII text, 119, 395 ASN.1 (schema language), 127 asynchronous networks, 278, 553 comparison to synchronous networks, 284 formal model, 307 asynchronous replication, 154, 553 conflict detection, 172 data loss on failover, 157 reads from asynchronous follower, 162 Asynchronous Transfer Mode (ATM), 285 atomic broadcast (see total order broadcast) atomic clocks (caesium clocks), 294, 295 (see also clocks) atomicity (concurrency), 553 atomic increment-and-get, 351 compare-and-set, 245, 327 (see also compare-and-set operations) replicated operations, 246 write operations, 243 atomicity (transactions), 223, 228, 553 atomic commit, 353 avoiding, 523, 528 blocking and nonblocking, 359 in stream processing, 360, 477 maintaining derived data, 453 for multi-object transactions, 229 for single-object writes, 230 auditability, 528-533 designing for, 531 self-auditing systems, 530 through immutability, 460 tools for auditable data systems, 532 availability, 8 (see also fault tolerance) in CAP theorem, 337 in service level agreements (SLAs), 15 Avro (data format), 122-127 code generation, 127 dynamically generated schemas, 126 object container files, 125, 131, 414 reader determining writer’s schema, 125 schema evolution, 123 use in Hadoop, 414 awk (Unix tool), 391 AWS (see Amazon Web Services) Azure (see Microsoft) B B-trees (indexes), 79-83 append-only/copy-on-write variants, 82, 242 branching factor, 81 comparison to LSM-trees, 83-85 crash recovery, 82 growing by splitting a page, 81 optimizations, 82 similarity to dynamic partitioning, 212 backpressure, 441, 553 in TCP, 282 backups database snapshot for replication, 156 integrity of, 530 snapshot isolation for, 238 use for ETL processes, 405 backward compatibility, 112 BASE, contrast to ACID, 223 bash shell (Unix), 70, 395, 503 batch processing, 28, 389-431, 553 combining with stream processing lambda architecture, 497 unifying technologies, 498 comparison to MPP databases, 414-418 comparison to stream processing, 464 comparison to Unix, 413-414 dataflow engines, 421-423 fault tolerance, 406, 414, 422, 442 for data integration, 494-498 graphs and iterative processing, 424-426 high-level APIs and languages, 403, 426-429 log-based messaging and, 451 maintaining derived state, 495 MapReduce and distributed filesystems, 397-413 (see also MapReduce) measuring performance, 13, 390 outputs, 411-413 key-value stores, 412 search indexes, 411 using Unix tools (example), 391-394 Bayou (database), 522 Beam (dataflow library), 498 bias, 534 big ball of mud, 20 Bigtable data model, 41, 99 binary data encodings, 115-128 Avro, 122-127 MessagePack, 116-117 Thrift and Protocol Buffers, 117-121 binary encoding based on schemas, 127 by network drivers, 128 binary strings, lack of support in JSON and XML, 114 BinaryProtocol encoding (Thrift), 118 Bitcask (storage engine), 72 crash recovery, 74 Bitcoin (cryptocurrency), 532 Byzantine fault tolerance, 305 concurrency bugs in exchanges, 233 bitmap indexes, 97 blockchains, 532 Byzantine fault tolerance, 305 blocking atomic commit, 359 Bloom (programming language), 504 Bloom filter (algorithm), 79, 466 BookKeeper (replicated log), 372 Bottled Water (change data capture), 455 bounded datasets, 430, 439, 553 (see also batch processing) bounded delays, 553 in networks, 285 process pauses, 298 broadcast hash joins, 409 Index | 561 brokerless messaging, 442 Brubeck (metrics aggregator), 442 BTM (transaction coordinator), 356 bulk synchronous parallel (BSP) model, 425 bursty network traffic patterns, 285 business data processing, 28, 90, 390 byte sequence, encoding data in, 112 Byzantine faults, 304-306, 307, 553 Byzantine fault-tolerant systems, 305, 532 Byzantine Generals Problem, 304 consensus algorithms and, 366 C caches, 89, 553 and materialized views, 101 as derived data, 386, 499-504 database as cache of transaction log, 460 in CPUs, 99, 338, 428 invalidation and maintenance, 452, 467 linearizability, 324 CAP theorem, 336-338, 554 Cascading (batch processing), 419, 427 hash joins, 409 workflows, 403 cascading failures, 9, 214, 281 Cascalog (batch processing), 60 Cassandra (database) column-family data model, 41, 99 compaction strategy, 79 compound primary key, 204 gossip protocol, 216 hash partitioning, 203-205 last-write-wins conflict resolution, 186, 292 leaderless replication, 177 linearizability, lack of, 335 log-structured storage, 78 multi-datacenter support, 184 partitioning scheme, 213 secondary indexes, 207 sloppy quorums, 184 cat (Unix tool), 391 causal context, 191 (see also causal dependencies) causal dependencies, 186-191 capturing, 191, 342, 494, 514 by total ordering, 493 causal ordering, 339 in transactions, 262 sending message to friends (example), 494 562 | Index causality, 554 causal ordering, 339-343 linearizability and, 342 total order consistent with, 344, 345 consistency with, 344-347 consistent snapshots, 340 happens-before relationship, 186 in serializable transactions, 262-265 mismatch with clocks, 292 ordering events to capture, 493 violations of, 165, 176, 292, 340 with synchronized clocks, 294 CEP (see complex event processing) certificate transparency, 532 chain replication, 155 linearizable reads, 351 change data capture, 160, 454 API support for change streams, 456 comparison to event sourcing, 457 implementing, 454 initial snapshot, 455 log compaction, 456 changelogs, 460 change data capture, 454 for operator state, 479 generating with triggers, 455 in stream joins, 474 log compaction, 456 maintaining derived state, 452 Chaos Monkey, 7, 280 checkpointing in batch processors, 422, 426 in high-performance computing, 275 in stream processors, 477, 523 chronicle data model, 458 circuit-switched networks, 284 circular buffers, 450 circular replication topologies, 175 clickstream data, analysis of, 404 clients calling services, 131 pushing state changes to, 512 request routing, 214 stateful and offline-capable, 170, 511 clocks, 287-299 atomic (caesium) clocks, 294, 295 confidence interval, 293-295 for global snapshots, 294 logical (see logical clocks) skew, 291-294, 334 slewing, 289 synchronization and accuracy, 289-291 synchronization using GPS, 287, 290, 294, 295 time-of-day versus monotonic clocks, 288 timestamping events, 471 cloud computing, 146, 275 need for service discovery, 372 network glitches, 279 shared resources, 284 single-machine reliability, 8 Cloudera Impala (see Impala) clustered indexes, 86 CODASYL model, 36 (see also network model) code generation with Avro, 127 with Thrift and Protocol Buffers, 118 with WSDL, 133 collaborative editing multi-leader replication and, 170 column families (Bigtable), 41, 99 column-oriented storage, 95-101 column compression, 97 distinction between column families and, 99 in batch processors, 428 Parquet, 96, 131, 414 sort order in, 99-100 vectorized processing, 99, 428 writing to, 101 comma-separated values (see CSV) command query responsibility segregation (CQRS), 462 commands (event sourcing), 459 commits (transactions), 222 atomic commit, 354-355 (see also atomicity; transactions) read committed isolation, 234 three-phase commit (3PC), 359 two-phase commit (2PC), 355-359 commutative operations, 246 compaction of changelogs, 456 (see also log compaction) for stream operator state, 479 of log-structured storage, 73 issues with, 84 size-tiered and leveled approaches, 79 CompactProtocol encoding (Thrift), 119 compare-and-set operations, 245, 327 implementing locks, 370 implementing uniqueness constraints, 331 implementing with total order broadcast, 350 relation to consensus, 335, 350, 352, 374 relation to transactions, 230 compatibility, 112, 128 calling services, 136 properties of encoding formats, 139 using databases, 129-131 using message-passing, 138 compensating transactions, 355, 461, 526 complex event processing (CEP), 465 complexity distilling in theoretical models, 310 hiding using abstraction, 27 of software systems, managing, 20 composing data systems (see unbundling data‐ bases) compute-intensive applications, 3, 275 concatenated indexes, 87 in Cassandra, 204 Concord (stream processor), 466 concurrency actor programming model, 138, 468 (see also message-passing) bugs from weak transaction isolation, 233 conflict resolution, 171, 174 detecting concurrent writes, 184-191 dual writes, problems with, 453 happens-before relationship, 186 in replicated systems, 161-191, 324-338 lost updates, 243 multi-version concurrency control (MVCC), 239 optimistic concurrency control, 261 ordering of operations, 326, 341 reducing, through event logs, 351, 462, 507 time and relativity, 187 transaction isolation, 225 write skew (transaction isolation), 246-251 conflict-free replicated datatypes (CRDTs), 174 conflicts conflict detection, 172 causal dependencies, 186, 342 in consensus algorithms, 368 in leaderless replication, 184 Index | 563 in log-based systems, 351, 521 in nonlinearizable systems, 343 in serializable snapshot isolation (SSI), 264 in two-phase commit, 357, 364 conflict resolution automatic conflict resolution, 174 by aborting transactions, 261 by apologizing, 527 convergence, 172-174 in leaderless systems, 190 last write wins (LWW), 186, 292 using atomic operations, 246 using custom logic, 173 determining what is a conflict, 174, 522 in multi-leader replication, 171-175 avoiding conflicts, 172 lost updates, 242-246 materializing, 251 relation to operation ordering, 339 write skew (transaction isolation), 246-251 congestion (networks) avoidance, 282 limiting accuracy of clocks, 293 queueing delays, 282 consensus, 321, 364-375, 554 algorithms, 366-368 preventing split brain, 367 safety and liveness properties, 365 using linearizable operations, 351 cost of, 369 distributed transactions, 352-375 in practice, 360-364 two-phase commit, 354-359 XA transactions, 361-364 impossibility of, 353 membership and coordination services, 370-373 relation to compare-and-set, 335, 350, 352, 374 relation to replication, 155, 349 relation to uniqueness constraints, 521 consistency, 224, 524 across different databases, 157, 452, 462, 492 causal, 339-348, 493 consistent prefix reads, 165-167 consistent snapshots, 156, 237-242, 294, 455, 500 (see also snapshots) 564 | Index crash recovery, 82 enforcing constraints (see constraints) eventual, 162, 322 (see also eventual consistency) in ACID transactions, 224, 529 in CAP theorem, 337 linearizability, 324-338 meanings of, 224 monotonic reads, 164-165 of secondary indexes, 231, 241, 354, 491, 500 ordering guarantees, 339-352 read-after-write, 162-164 sequential, 351 strong (see linearizability) timeliness and integrity, 524 using quorums, 181, 334 consistent hashing, 204 consistent prefix reads, 165 constraints (databases), 225, 248 asynchronously checked, 526 coordination avoidance, 527 ensuring idempotence, 519 in log-based systems, 521-524 across multiple partitions, 522 in two-phase commit, 355, 357 relation to consensus, 374, 521 relation to event ordering, 347 requiring linearizability, 330 Consul (service discovery), 372 consumers (message streams), 137, 440 backpressure, 441 consumer offsets in logs, 449 failures, 445, 449 fan-out, 11, 445, 448 load balancing, 444, 448 not keeping up with producers, 441, 450, 502 context switches, 14, 297 convergence (conflict resolution), 172-174, 322 coordination avoidance, 527 cross-datacenter, 168, 493 cross-partition ordering, 256, 294, 348, 523 services, 330, 370-373 coordinator (in 2PC), 356 failure, 358 in XA transactions, 361-364 recovery, 363 copy-on-write (B-trees), 82, 242 CORBA (Common Object Request Broker Architecture), 134 correctness, 6 auditability, 528-533 Byzantine fault tolerance, 305, 532 dealing with partial failures, 274 in log-based systems, 521-524 of algorithm within system model, 308 of compensating transactions, 355 of consensus, 368 of derived data, 497, 531 of immutable data, 461 of personal data, 535, 540 of time, 176, 289-295 of transactions, 225, 515, 529 timeliness and integrity, 524-528 corruption of data detecting, 519, 530-533 due to pathological memory access, 529 due to radiation, 305 due to split brain, 158, 302 due to weak transaction isolation, 233 formalization in consensus, 366 integrity as absence of, 524 network packets, 306 on disks, 227 preventing using write-ahead logs, 82 recovering from, 414, 460 Couchbase (database) durability, 89 hash partitioning, 203-204, 211 rebalancing, 213 request routing, 216 CouchDB (database) B-tree storage, 242 change feed, 456 document data model, 31 join support, 34 MapReduce support, 46, 400 replication, 170, 173 covering indexes, 86 CPUs cache coherence and memory barriers, 338 caching and pipelining, 99, 428 increasing parallelism, 43 CRDTs (see conflict-free replicated datatypes) CREATE INDEX statement (SQL), 85, 500 credit rating agencies, 535 Crunch (batch processing), 419, 427 hash joins, 409 sharded joins, 408 workflows, 403 cryptography defense against attackers, 306 end-to-end encryption and authentication, 519, 543 proving integrity of data, 532 CSS (Cascading Style Sheets), 44 CSV (comma-separated values), 70, 114, 396 Curator (ZooKeeper recipes), 330, 371 curl (Unix tool), 135, 397 cursor stability, 243 Cypher (query language), 52 comparison to SPARQL, 59 D data corruption (see corruption of data) data cubes, 102 data formats (see encoding) data integration, 490-498, 543 batch and stream processing, 494-498 lambda architecture, 497 maintaining derived state, 495 reprocessing data, 496 unifying, 498 by unbundling databases, 499-515 comparison to federated databases, 501 combining tools by deriving data, 490-494 derived data versus distributed transac‐ tions, 492 limits of total ordering, 493 ordering events to capture causality, 493 reasoning about dataflows, 491 need for, 385 data lakes, 415 data locality (see locality) data models, 27-64 graph-like models, 49-63 Datalog language, 60-63 property graphs, 50 RDF and triple-stores, 55-59 query languages, 42-48 relational model versus document model, 28-42 data protection regulations, 542 data systems, 3 about, 4 Index | 565 concerns when designing, 5 future of, 489-544 correctness, constraints, and integrity, 515-533 data integration, 490-498 unbundling databases, 499-515 heterogeneous, keeping in sync, 452 maintainability, 18-22 possible faults in, 221 reliability, 6-10 hardware faults, 7 human errors, 9 importance of, 10 software errors, 8 scalability, 10-18 unreliable clocks, 287-299 data warehousing, 91-95, 554 comparison to data lakes, 415 ETL (extract-transform-load), 92, 416, 452 keeping data systems in sync, 452 schema design, 93 slowly changing dimension (SCD), 476 data-intensive applications, 3 database triggers (see triggers) database-internal distributed transactions, 360, 364, 477 databases archival storage, 131 comparison of message brokers to, 443 dataflow through, 129 end-to-end argument for, 519-520 checking integrity, 531 inside-out, 504 (see also unbundling databases) output from batch workflows, 412 relation to event streams, 451-464 (see also changelogs) API support for change streams, 456, 506 change data capture, 454-457 event sourcing, 457-459 keeping systems in sync, 452-453 philosophy of immutable events, 459-464 unbundling, 499-515 composing data storage technologies, 499-504 designing applications around dataflow, 504-509 566 | Index observing derived state, 509-515 datacenters geographically distributed, 145, 164, 278, 493 multi-tenancy and shared resources, 284 network architecture, 276 network faults, 279 replication across multiple, 169 leaderless replication, 184 multi-leader replication, 168, 335 dataflow, 128-139, 504-509 correctness of dataflow systems, 525 differential, 504 message-passing, 136-139 reasoning about, 491 through databases, 129 through services, 131-136 dataflow engines, 421-423 comparison to stream processing, 464 directed acyclic graphs (DAG), 424 partitioning, approach to, 429 support for declarative queries, 427 Datalog (query language), 60-63 datatypes binary strings in XML and JSON, 114 conflict-free, 174 in Avro encodings, 122 in Thrift and Protocol Buffers, 121 numbers in XML and JSON, 114 Datomic (database) B-tree storage, 242 data model, 50, 57 Datalog query language, 60 excision (deleting data), 463 languages for transactions, 255 serial execution of transactions, 253 deadlocks detection, in two-phase commit (2PC), 364 in two-phase locking (2PL), 258 Debezium (change data capture), 455 declarative languages, 42, 554 Bloom, 504 CSS and XSL, 44 Cypher, 52 Datalog, 60 for batch processing, 427 recursive SQL queries, 53 relational algebra and SQL, 42 SPARQL, 59 delays bounded network delays, 285 bounded process pauses, 298 unbounded network delays, 282 unbounded process pauses, 296 deleting data, 463 denormalization (data representation), 34, 554 costs, 39 in derived data systems, 386 materialized views, 101 updating derived data, 228, 231, 490 versus normalization, 462 derived data, 386, 439, 554 from change data capture, 454 in event sourcing, 458-458 maintaining derived state through logs, 452-457, 459-463 observing, by subscribing to streams, 512 outputs of batch and stream processing, 495 through application code, 505 versus distributed transactions, 492 deterministic operations, 255, 274, 554 accidental nondeterminism, 423 and fault tolerance, 423, 426 and idempotence, 478, 492 computing derived data, 495, 526, 531 in state machine replication, 349, 452, 458 joins, 476 DevOps, 394 differential dataflow, 504 dimension tables, 94 dimensional modeling (see star schemas) directed acyclic graphs (DAGs), 424 dirty reads (transaction isolation), 234 dirty writes (transaction isolation), 235 discrimination, 534 disks (see hard disks) distributed actor frameworks, 138 distributed filesystems, 398-399 decoupling from query engines, 417 indiscriminately dumping data into, 415 use by MapReduce, 402 distributed systems, 273-312, 554 Byzantine faults, 304-306 cloud versus supercomputing, 275 detecting network faults, 280 faults and partial failures, 274-277 formalization of consensus, 365 impossibility results, 338, 353 issues with failover, 157 limitations of distributed transactions, 363 multi-datacenter, 169, 335 network problems, 277-286 quorums, relying on, 301 reasons for using, 145, 151 synchronized clocks, relying on, 291-295 system models, 306-310 use of clocks and time, 287 distributed transactions (see transactions) Django (web framework), 232 DNS (Domain Name System), 216, 372 Docker (container manager), 506 document data model, 30-42 comparison to relational model, 38-42 document references, 38, 403 document-oriented databases, 31 many-to-many relationships and joins, 36 multi-object transactions, need for, 231 versus relational model convergence of models, 41 data locality, 41 document-partitioned indexes, 206, 217, 411 domain-driven design (DDD), 457 DRBD (Distributed Replicated Block Device), 153 drift (clocks), 289 Drill (query engine), 93 Druid (database), 461 Dryad (dataflow engine), 421 dual writes, problems with, 452, 507 duplicates, suppression of, 517 (see also idempotence) using a unique ID, 518, 522 durability (transactions), 226, 554 duration (time), 287 measurement with monotonic clocks, 288 dynamic partitioning, 212 dynamically typed languages analogy to schema-on-read, 40 code generation and, 127 Dynamo-style databases (see leaderless replica‐ tion) E edges (in graphs), 49, 403 property graph model, 50 edit distance (full-text search), 88 effectively-once semantics, 476, 516 Index | 567 (see also exactly-once semantics) preservation of integrity, 525 elastic systems, 17 Elasticsearch (search server) document-partitioned indexes, 207 partition rebalancing, 211 percolator (stream search), 467 usage example, 4 use of Lucene, 79 ElephantDB (database), 413 Elm (programming language), 504, 512 encodings (data formats), 111-128 Avro, 122-127 binary variants of JSON and XML, 115 compatibility, 112 calling services, 136 using databases, 129-131 using message-passing, 138 defined, 113 JSON, XML, and CSV, 114 language-specific formats, 113 merits of schemas, 127 representations of data, 112 Thrift and Protocol Buffers, 117-121 end-to-end argument, 277, 519-520 checking integrity, 531 publish/subscribe streams, 512 enrichment (stream), 473 Enterprise JavaBeans (EJB), 134 entities (see vertices) epoch (consensus algorithms), 368 epoch (Unix timestamps), 288 equi-joins, 403 erasure coding (error correction), 398 Erlang OTP (actor framework), 139 error handling for network faults, 280 in transactions, 231 error-correcting codes, 277, 398 Esper (CEP engine), 466 etcd (coordination service), 370-373 linearizable operations, 333 locks and leader election, 330 quorum reads, 351 service discovery, 372 use of Raft algorithm, 349, 353 Ethereum (blockchain), 532 Ethernet (networks), 276, 278, 285 packet checksums, 306, 519 568 | Index Etherpad (collaborative editor), 170 ethics, 533-543 code of ethics and professional practice, 533 legislation and self-regulation, 542 predictive analytics, 533-536 amplifying bias, 534 feedback loops, 536 privacy and tracking, 536-543 consent and freedom of choice, 538 data as assets and power, 540 meaning of privacy, 539 surveillance, 537 respect, dignity, and agency, 543, 544 unintended consequences, 533, 536 ETL (extract-transform-load), 92, 405, 452, 554 use of Hadoop for, 416 event sourcing, 457-459 commands and events, 459 comparison to change data capture, 457 comparison to lambda architecture, 497 deriving current state from event log, 458 immutability and auditability, 459, 531 large, reliable data systems, 519, 526 Event Store (database), 458 event streams (see streams) events, 440 deciding on total order of, 493 deriving views from event log, 461 difference to commands, 459 event time versus processing time, 469, 477, 498 immutable, advantages of, 460, 531 ordering to capture causality, 493 reads as, 513 stragglers, 470, 498 timestamp of, in stream processing, 471 EventSource (browser API), 512 eventual consistency, 152, 162, 308, 322 (see also conflicts) and perpetual inconsistency, 525 evolvability, 21, 111 calling services, 136 graph-structured data, 52 of databases, 40, 129-131, 461, 497 of message-passing, 138 reprocessing data, 496, 498 schema evolution in Avro, 123 schema evolution in Thrift and Protocol Buffers, 120 schema-on-read, 39, 111, 128 exactly-once semantics, 360, 476, 516 parity with batch processors, 498 preservation of integrity, 525 exclusive mode (locks), 258 eXtended Architecture transactions (see XA transactions) extract-transform-load (see ETL) F Facebook Presto (query engine), 93 React, Flux, and Redux (user interface libra‐ ries), 512 social graphs, 49 Wormhole (change data capture), 455 fact tables, 93 failover, 157, 554 (see also leader-based replication) in leaderless replication, absence of, 178 leader election, 301, 348, 352 potential problems, 157 failures amplification by distributed transactions, 364, 495 failure detection, 280 automatic rebalancing causing cascading failures, 214 perfect failure detectors, 359 timeouts and unbounded delays, 282, 284 using ZooKeeper, 371 faults versus, 7 partial failures in distributed systems, 275-277, 310 fan-out (messaging systems), 11, 445 fault tolerance, 6-10, 555 abstractions for, 321 formalization in consensus, 365-369 use of replication, 367 human fault tolerance, 414 in batch processing, 406, 414, 422, 425 in log-based systems, 520, 524-526 in stream processing, 476-479 atomic commit, 477 idempotence, 478 maintaining derived state, 495 microbatching and checkpointing, 477 rebuilding state after a failure, 478 of distributed transactions, 362-364 transaction atomicity, 223, 354-361 faults, 6 Byzantine faults, 304-306 failures versus, 7 handled by transactions, 221 handling in supercomputers and cloud computing, 275 hardware, 7 in batch processing versus distributed data‐ bases, 417 in distributed systems, 274-277 introducing deliberately, 7, 280 network faults, 279-281 asymmetric faults, 300 detecting, 280 tolerance of, in multi-leader replication, 169 software errors, 8 tolerating (see fault tolerance) federated databases, 501 fence (CPU instruction), 338 fencing (preventing split brain), 158, 302-304 generating fencing tokens, 349, 370 properties of fencing tokens, 308 stream processors writing to databases, 478, 517 Fibre Channel (networks), 398 field tags (Thrift and Protocol Buffers), 119-121 file descriptors (Unix), 395 financial data, 460 Firebase (database), 456 Flink (processing framework), 421-423 dataflow APIs, 427 fault tolerance, 422, 477, 479 Gelly API (graph processing), 425 integration of batch and stream processing, 495, 498 machine learning, 428 query optimizer, 427 stream processing, 466 flow control, 282, 441, 555 FLP result (on consensus), 353 FlumeJava (dataflow library), 403, 427 followers, 152, 555 (see also leader-based replication) foreign keys, 38, 403 forward compatibility, 112 forward decay (algorithm), 16 Index | 569 Fossil (version control system), 463 shunning (deleting data), 463 FoundationDB (database) serializable transactions, 261, 265, 364 fractal trees, 83 full table scans, 403 full-text search, 555 and fuzzy indexes, 88 building search indexes, 411 Lucene storage engine, 79 functional reactive programming (FRP), 504 functional requirements, 22 futures (asynchronous operations), 135 fuzzy search (see similarity search) G garbage collection immutability and, 463 process pauses for, 14, 296-299, 301 (see also process pauses) genome analysis, 63, 429 geographically distributed datacenters, 145, 164, 278, 493 geospatial indexes, 87 Giraph (graph processing), 425 Git (version control system), 174, 342, 463 GitHub, postmortems, 157, 158, 309 global indexes (see term-partitioned indexes) GlusterFS (distributed filesystem), 398 GNU Coreutils (Linux), 394 GoldenGate (change data capture), 161, 170, 455 (see also Oracle) Google Bigtable (database) data model (see Bigtable data model) partitioning scheme, 199, 202 storage layout, 78 Chubby (lock service), 370 Cloud Dataflow (stream processor), 466, 477, 498 (see also Beam) Cloud Pub/Sub (messaging), 444, 448 Docs (collaborative editor), 170 Dremel (query engine), 93, 96 FlumeJava (dataflow library), 403, 427 GFS (distributed file system), 398 gRPC (RPC framework), 135 MapReduce (batch processing), 390 570 | Index (see also MapReduce) building search indexes, 411 task preemption, 418 Pregel (graph processing), 425 Spanner (see Spanner) TrueTime (clock API), 294 gossip protocol, 216 government use of data, 541 GPS (Global Positioning System) use for clock synchronization, 287, 290, 294, 295 GraphChi (graph processing), 426 graphs, 555 as data models, 49-63 example of graph-structured data, 49 property graphs, 50 RDF and triple-stores, 55-59 versus the network model, 60 processing and analysis, 424-426 fault tolerance, 425 Pregel processing model, 425 query languages Cypher, 52 Datalog, 60-63 recursive SQL queries, 53 SPARQL, 59-59 Gremlin (graph query language), 50 grep (Unix tool), 392 GROUP BY clause (SQL), 406 grouping records in MapReduce, 406 handling skew, 407 H Hadoop (data infrastructure) comparison to distributed databases, 390 comparison to MPP databases, 414-418 comparison to Unix, 413-414, 499 diverse processing models in ecosystem, 417 HDFS distributed filesystem (see HDFS) higher-level tools, 403 join algorithms, 403-410 (see also MapReduce) MapReduce (see MapReduce) YARN (see YARN) happens-before relationship, 340 capturing, 187 concurrency and, 186 hard disks access patterns, 84 detecting corruption, 519, 530 faults in, 7, 227 sequential write throughput, 75, 450 hardware faults, 7 hash indexes, 72-75 broadcast hash joins, 409 partitioned hash joins, 409 hash partitioning, 203-205, 217 consistent hashing, 204 problems with hash mod N, 210 range queries, 204 suitable hash functions, 203 with fixed number of partitions, 210 HAWQ (database), 428 HBase (database) bug due to lack of fencing, 302 bulk loading, 413 column-family data model, 41, 99 dynamic partitioning, 212 key-range partitioning, 202 log-structured storage, 78 request routing, 216 size-tiered compaction, 79 use of HDFS, 417 use of ZooKeeper, 370 HDFS (Hadoop Distributed File System), 398-399 (see also distributed filesystems) checking data integrity, 530 decoupling from query engines, 417 indiscriminately dumping data into, 415 metadata about datasets, 410 NameNode, 398 use by Flink, 479 use by HBase, 212 use by MapReduce, 402 HdrHistogram (numerical library), 16 head (Unix tool), 392 head vertex (property graphs), 51 head-of-line blocking, 15 heap files (databases), 86 Helix (cluster manager), 216 heterogeneous distributed transactions, 360, 364 heuristic decisions (in 2PC), 363 Hibernate (object-relational mapper), 30 hierarchical model, 36 high availability (see fault tolerance) high-frequency trading, 290, 299 high-performance computing (HPC), 275 hinted handoff, 183 histograms, 16 Hive (query engine), 419, 427 for data warehouses, 93 HCatalog and metastore, 410 map-side joins, 409 query optimizer, 427 skewed joins, 408 workflows, 403 Hollerith machines, 390 hopping windows (stream processing), 472 (see also windows) horizontal scaling (see scaling out) HornetQ (messaging), 137, 444 distributed transaction support, 361 hot spots, 201 due to celebrities, 205 for time-series data, 203 in batch processing, 407 relieving, 205 hot standbys (see leader-based replication) HTTP, use in APIs (see services) human errors, 9, 279, 414 HyperDex (database), 88 HyperLogLog (algorithm), 466 I I/O operations, waiting for, 297 IBM DB2 (database) distributed transaction support, 361 recursive query support, 54 serializable isolation, 242, 257 XML and JSON support, 30, 42 electromechanical card-sorting machines, 390 IMS (database), 36 imperative query APIs, 46 InfoSphere Streams (CEP engine), 466 MQ (messaging), 444 distributed transaction support, 361 System R (database), 222 WebSphere (messaging), 137 idempotence, 134, 478, 555 by giving operations unique IDs, 518, 522 idempotent operations, 517 immutability advantages of, 460, 531 Index | 571 deriving state from event log, 459-464 for crash recovery, 75 in B-trees, 82, 242 in event sourcing, 457 inputs to Unix commands, 397 limitations of, 463 Impala (query engine) for data warehouses, 93 hash joins, 409 native code generation, 428 use of HDFS, 417 impedance mismatch, 29 imperative languages, 42 setting element styles (example), 45 in doubt (transaction status), 358 holding locks, 362 orphaned transactions, 363 in-memory databases, 88 durability, 227 serial transaction execution, 253 incidents cascading failures, 9 crashes due to leap seconds, 290 data corruption and financial losses due to concurrency bugs, 233 data corruption on hard disks, 227 data loss due to last-write-wins, 173, 292 data on disks unreadable, 309 deleted items reappearing, 174 disclosure of sensitive data due to primary key reuse, 157 errors in transaction serializability, 529 gigabit network interface with 1 Kb/s throughput, 311 network faults, 279 network interface dropping only inbound packets, 279 network partitions and whole-datacenter failures, 275 poor handling of network faults, 280 sending message to ex-partner, 494 sharks biting undersea cables, 279 split brain due to 1-minute packet delay, 158, 279 vibrations in server rack, 14 violation of uniqueness constraint, 529 indexes, 71, 555 and snapshot isolation, 241 as derived data, 386, 499-504 572 | Index B-trees, 79-83 building in batch processes, 411 clustered, 86 comparison of B-trees and LSM-trees, 83-85 concatenated, 87 covering (with included columns), 86 creating, 500 full-text search, 88 geospatial, 87 hash, 72-75 index-range locking, 260 multi-column, 87 partitioning and secondary indexes, 206-209, 217 secondary, 85 (see also secondary indexes) problems with dual writes, 452, 491 SSTables and LSM-trees, 76-79 updating when data changes, 452, 467 Industrial Revolution, 541 InfiniBand (networks), 285 InfiniteGraph (database), 50 InnoDB (storage engine) clustered index on primary key, 86 not preventing lost updates, 245 preventing write skew, 248, 257 serializable isolation, 257 snapshot isolation support, 239 inside-out databases, 504 (see also unbundling databases) integrating different data systems (see data integration) integrity, 524 coordination-avoiding data systems, 528 correctness of dataflow systems, 525 in consensus formalization, 365 integrity checks, 530 (see also auditing) end-to-end, 519, 531 use of snapshot isolation, 238 maintaining despite software bugs, 529 Interface Definition Language (IDL), 117, 122 intermediate state, materialization of, 420-423 internet services, systems for implementing, 275 invariants, 225 (see also constraints) inversion of control, 396 IP (Internet Protocol) unreliability of, 277 ISDN (Integrated Services Digital Network), 284 isolation (in transactions), 225, 228, 555 correctness and, 515 for single-object writes, 230 serializability, 251-266 actual serial execution, 252-256 serializable snapshot isolation (SSI), 261-266 two-phase locking (2PL), 257-261 violating, 228 weak isolation levels, 233-251 preventing lost updates, 242-246 read committed, 234-237 snapshot isolation, 237-242 iterative processing, 424-426 J Java Database Connectivity (JDBC) distributed transaction support, 361 network drivers, 128 Java Enterprise Edition (EE), 134, 356, 361 Java Message Service (JMS), 444 (see also messaging systems) comparison to log-based messaging, 448, 451 distributed transaction support, 361 message ordering, 446 Java Transaction API (JTA), 355, 361 Java Virtual Machine (JVM) bytecode generation, 428 garbage collection pauses, 296 process reuse in batch processors, 422 JavaScript in MapReduce querying, 46 setting element styles (example), 45 use in advanced queries, 48 Jena (RDF framework), 57 Jepsen (fault tolerance testing), 515 jitter (network delay), 284 joins, 555 by index lookup, 403 expressing as relational operators, 427 in relational and document databases, 34 MapReduce map-side joins, 408-410 broadcast hash joins, 409 merge joins, 410 partitioned hash joins, 409 MapReduce reduce-side joins, 403-408 handling skew, 407 sort-merge joins, 405 parallel execution of, 415 secondary indexes and, 85 stream joins, 472-476 stream-stream join, 473 stream-table join, 473 table-table join, 474 time-dependence of, 475 support in document databases, 42 JOTM (transaction coordinator), 356 JSON Avro schema representation, 122 binary variants, 115 for application data, issues with, 114 in relational databases, 30, 42 representing a résumé (example), 31 Juttle (query language), 504 K k-nearest neighbors, 429 Kafka (messaging), 137, 448 Kafka Connect (database integration), 457, 461 Kafka Streams (stream processor), 466, 467 fault tolerance, 479 leader-based replication, 153 log compaction, 456, 467 message offsets, 447, 478 request routing, 216 transaction support, 477 usage example, 4 Ketama (partitioning library), 213 key-value stores, 70 as batch process output, 412 hash indexes, 72-75 in-memory, 89 partitioning, 201-205 by hash of key, 203, 217 by key range, 202, 217 dynamic partitioning, 212 skew and hot spots, 205 Kryo (Java), 113 Kubernetes (cluster manager), 418, 506 L lambda architecture, 497 Lamport timestamps, 345 Index | 573 Large Hadron Collider (LHC), 64 last write wins (LWW), 173, 334 discarding concurrent writes, 186 problems with, 292 prone to lost updates, 246 late binding, 396 latency instability under two-phase locking, 259 network latency and resource utilization, 286 response time versus, 14 tail latency, 15, 207 leader-based replication, 152-161 (see also replication) failover, 157, 301 handling node outages, 156 implementation of replication logs change data capture, 454-457 (see also changelogs) statement-based, 158 trigger-based replication, 161 write-ahead log (WAL) shipping, 159 linearizability of operations, 333 locking and leader election, 330 log sequence number, 156, 449 read-scaling architecture, 161 relation to consensus, 367 setting up new followers, 155 synchronous versus asynchronous, 153-155 leaderless replication, 177-191 (see also replication) detecting concurrent writes, 184-191 capturing happens-before relationship, 187 happens-before relationship and concur‐ rency, 186 last write wins, 186 merging concurrently written values, 190 version vectors, 191 multi-datacenter, 184 quorums, 179-182 consistency limitations, 181-183, 334 sloppy quorums and hinted handoff, 183 read repair and anti-entropy, 178 leap seconds, 8, 290 in time-of-day clocks, 288 leases, 295 implementation with ZooKeeper, 370 574 | Index need for fencing, 302 ledgers, 460 distributed ledger technologies, 532 legacy systems, maintenance of, 18 less (Unix tool), 397 LevelDB (storage engine), 78 leveled compaction, 79 Levenshtein automata, 88 limping (partial failure), 311 linearizability, 324-338, 555 cost of, 335-338 CAP theorem, 336 memory on multi-core CPUs, 338 definition, 325-329 implementing with total order broadcast, 350 in ZooKeeper, 370 of derived data systems, 492, 524 avoiding coordination, 527 of different replication methods, 332-335 using quorums, 334 relying on, 330-332 constraints and uniqueness, 330 cross-channel timing dependencies, 331 locking and leader election, 330 stronger than causal consistency, 342 using to implement total order broadcast, 351 versus serializability, 329 LinkedIn Azkaban (workflow scheduler), 402 Databus (change data capture), 161, 455 Espresso (database), 31, 126, 130, 153, 216 Helix (cluster manager) (see Helix) profile (example), 30 reference to company entity (example), 34 Rest.li (RPC framework), 135 Voldemort (database) (see Voldemort) Linux, leap second bug, 8, 290 liveness properties, 308 LMDB (storage engine), 82, 242 load approaches to coping with, 17 describing, 11 load testing, 16 load balancing (messaging), 444 local indexes (see document-partitioned indexes) locality (data access), 32, 41, 555 in batch processing, 400, 405, 421 in stateful clients, 170, 511 in stream processing, 474, 478, 508, 522 location transparency, 134 in the actor model, 138 locks, 556 deadlock, 258 distributed locking, 301-304, 330 fencing tokens, 303 implementation with ZooKeeper, 370 relation to consensus, 374 for transaction isolation in snapshot isolation, 239 in two-phase locking (2PL), 257-261 making operations atomic, 243 performance, 258 preventing dirty writes, 236 preventing phantoms with index-range locks, 260, 265 read locks (shared mode), 236, 258 shared mode and exclusive mode, 258 in two-phase commit (2PC) deadlock detection, 364 in-doubt transactions holding locks, 362 materializing conflicts with, 251 preventing lost updates by explicit locking, 244 log sequence number, 156, 449 logic programming languages, 504 logical clocks, 293, 343, 494 for read-after-write consistency, 164 logical logs, 160 logs (data structure), 71, 556 advantages of immutability, 460 compaction, 73, 79, 456, 460 for stream operator state, 479 creating using total order broadcast, 349 implementing uniqueness constraints, 522 log-based messaging, 446-451 comparison to traditional messaging, 448, 451 consumer offsets, 449 disk space usage, 450 replaying old messages, 451, 496, 498 slow consumers, 450 using logs for message storage, 447 log-structured storage, 71-79 log-structured merge tree (see LSMtrees) replication, 152, 158-161 change data capture, 454-457 (see also changelogs) coordination with snapshot, 156 logical (row-based) replication, 160 statement-based replication, 158 trigger-based replication, 161 write-ahead log (WAL) shipping, 159 scalability limits, 493 loose coupling, 396, 419, 502 lost updates (see updates) LSM-trees (indexes), 78-79 comparison to B-trees, 83-85 Lucene (storage engine), 79 building indexes in batch processes, 411 similarity search, 88 Luigi (workflow scheduler), 402 LWW (see last write wins) M machine learning ethical considerations, 534 (see also ethics) iterative processing, 424 models derived from training data, 505 statistical and numerical algorithms, 428 MADlib (machine learning toolkit), 428 magic scaling sauce, 18 Mahout (machine learning toolkit), 428 maintainability, 18-22, 489 defined, 23 design principles for software systems, 19 evolvability (see evolvability) operability, 19 simplicity and managing complexity, 20 many-to-many relationships in document model versus relational model, 39 modeling as graphs, 49 many-to-one and many-to-many relationships, 33-36 many-to-one relationships, 34 MapReduce (batch processing), 390, 399-400 accessing external services within job, 404, 412 comparison to distributed databases designing for frequent faults, 417 diversity of processing models, 416 diversity of storage, 415 Index | 575 comparison to stream processing, 464 comparison to Unix, 413-414 disadvantages and limitations of, 419 fault tolerance, 406, 414, 422 higher-level tools, 403, 426 implementation in Hadoop, 400-403 the shuffle, 402 implementation in MongoDB, 46-48 machine learning, 428 map-side processing, 408-410 broadcast hash joins, 409 merge joins, 410 partitioned hash joins, 409 mapper and reducer functions, 399 materialization of intermediate state, 419-423 output of batch workflows, 411-413 building search indexes, 411 key-value stores, 412 reduce-side processing, 403-408 analysis of user activity events (exam‐ ple), 404 grouping records by same key, 406 handling skew, 407 sort-merge joins, 405 workflows, 402 marshalling (see encoding) massively parallel processing (MPP), 216 comparison to composing storage technolo‐ gies, 502 comparison to Hadoop, 414-418, 428 master-master replication (see multi-leader replication) master-slave replication (see leader-based repli‐ cation) materialization, 556 aggregate values, 101 conflicts, 251 intermediate state (batch processing), 420-423 materialized views, 101 as derived data, 386, 499-504 maintaining, using stream processing, 467, 475 Maven (Java build tool), 428 Maxwell (change data capture), 455 mean, 14 media monitoring, 467 median, 14 576 | Index meeting room booking (example), 249, 259, 521 membership services, 372 Memcached (caching server), 4, 89 memory in-memory databases, 88 durability, 227 serial transaction execution, 253 in-memory representation of data, 112 random bit-flips in, 529 use by indexes, 72, 77 memory barrier (CPU instruction), 338 MemSQL (database) in-memory storage, 89 read committed isolation, 236 memtable (in LSM-trees), 78 Mercurial (version control system), 463 merge joins, MapReduce map-side, 410 mergeable persistent data structures, 174 merging sorted files, 76, 402, 405 Merkle trees, 532 Mesos (cluster manager), 418, 506 message brokers (see messaging systems) message-passing, 136-139 advantages over direct RPC, 137 distributed actor frameworks, 138 evolvability, 138 MessagePack (encoding format), 116 messages exactly-once semantics, 360, 476 loss of, 442 using total order broadcast, 348 messaging systems, 440-451 (see also streams) backpressure, buffering, or dropping mes‐ sages, 441 brokerless messaging, 442 event logs, 446-451 comparison to traditional messaging, 448, 451 consumer offsets, 449 replaying old messages, 451, 496, 498 slow consumers, 450 message brokers, 443-446 acknowledgements and redelivery, 445 comparison to event logs, 448, 451 multiple consumers of same topic, 444 reliability, 442 uniqueness in log-based messaging, 522 Meteor (web framework), 456 microbatching, 477, 495 microservices, 132 (see also services) causal dependencies across services, 493 loose coupling, 502 relation to batch/stream processors, 389, 508 Microsoft Azure Service Bus (messaging), 444 Azure Storage, 155, 398 Azure Stream Analytics, 466 DCOM (Distributed Component Object Model), 134 MSDTC (transaction coordinator), 356 Orleans (see Orleans) SQL Server (see SQL Server) migrating (rewriting) data, 40, 130, 461, 497 modulus operator (%), 210 MongoDB (database) aggregation pipeline, 48 atomic operations, 243 BSON, 41 document data model, 31 hash partitioning (sharding), 203-204 key-range partitioning, 202 lack of join support, 34, 42 leader-based replication, 153 MapReduce support, 46, 400 oplog parsing, 455, 456 partition splitting, 212 request routing, 216 secondary indexes, 207 Mongoriver (change data capture), 455 monitoring, 10, 19 monotonic clocks, 288 monotonic reads, 164 MPP (see massively parallel processing) MSMQ (messaging), 361 multi-column indexes, 87 multi-leader replication, 168-177 (see also replication) handling write conflicts, 171 conflict avoidance, 172 converging toward a consistent state, 172 custom conflict resolution logic, 173 determining what is a conflict, 174 linearizability, lack of, 333 replication topologies, 175-177 use cases, 168 clients with offline operation, 170 collaborative editing, 170 multi-datacenter replication, 168, 335 multi-object transactions, 228 need for, 231 Multi-Paxos (total order broadcast), 367 multi-table index cluster tables (Oracle), 41 multi-tenancy, 284 multi-version concurrency control (MVCC), 239, 266 detecting stale MVCC reads, 263 indexes and snapshot isolation, 241 mutual exclusion, 261 (see also locks) MySQL (database) binlog coordinates, 156 binlog parsing for change data capture, 455 circular replication topology, 175 consistent snapshots, 156 distributed transaction support, 361 InnoDB storage engine (see InnoDB) JSON support, 30, 42 leader-based replication, 153 performance of XA transactions, 360 row-based replication, 160 schema changes in, 40 snapshot isolation support, 242 (see also InnoDB) statement-based replication, 159 Tungsten Replicator (multi-leader replica‐ tion), 170 conflict detection, 177 N nanomsg (messaging library), 442 Narayana (transaction coordinator), 356 NATS (messaging), 137 near-real-time (nearline) processing, 390 (see also stream processing) Neo4j (database) Cypher query language, 52 graph data model, 50 Nephele (dataflow engine), 421 netcat (Unix tool), 397 Netflix Chaos Monkey, 7, 280 Network Attached Storage (NAS), 146, 398 network model, 36 Index | 577 graph databases versus, 60 imperative query APIs, 46 Network Time Protocol (see NTP) networks congestion and queueing, 282 datacenter network topologies, 276 faults (see faults) linearizability and network delays, 338 network partitions, 279, 337 timeouts and unbounded delays, 281 next-key locking, 260 nodes (in graphs) (see vertices) nodes (processes), 556 handling outages in leader-based replica‐ tion, 156 system models for failure, 307 noisy neighbors, 284 nonblocking atomic commit, 359 nondeterministic operations accidental nondeterminism, 423 partial failures in distributed systems, 275 nonfunctional requirements, 22 nonrepeatable reads, 238 (see also read skew) normalization (data representation), 33, 556 executing joins, 39, 42, 403 foreign key references, 231 in systems of record, 386 versus denormalization, 462 NoSQL, 29, 499 transactions and, 223 Notation3 (N3), 56 npm (package manager), 428 NTP (Network Time Protocol), 287 accuracy, 289, 293 adjustments to monotonic clocks, 289 multiple server addresses, 306 numbers, in XML and JSON encodings, 114 O object-relational mapping (ORM) frameworks, 30 error handling and aborted transactions, 232 unsafe read-modify-write cycle code, 244 object-relational mismatch, 29 observer pattern, 506 offline systems, 390 (see also batch processing) 578 | Index stateful, offline-capable clients, 170, 511 offline-first applications, 511 offsets consumer offsets in partitioned logs, 449 messages in partitioned logs, 447 OLAP (online analytic processing), 91, 556 data cubes, 102 OLTP (online transaction processing), 90, 556 analytics queries versus, 411 workload characteristics, 253 one-to-many relationships, 30 JSON representation, 32 online systems, 389 (see also services) Oozie (workflow scheduler), 402 OpenAPI (service definition format), 133 OpenStack Nova (cloud infrastructure) use of ZooKeeper, 370 Swift (object storage), 398 operability, 19 operating systems versus databases, 499 operation identifiers, 518, 522 operational transformation, 174 operators, 421 flow of data between, 424 in stream processing, 464 optimistic concurrency control, 261 Oracle (database) distributed transaction support, 361 GoldenGate (change data capture), 161, 170, 455 lack of serializability, 226 leader-based replication, 153 multi-table index cluster tables, 41 not preventing write skew, 248 partitioned indexes, 209 PL/SQL language, 255 preventing lost updates, 245 read committed isolation, 236 Real Application Clusters (RAC), 330 recursive query support, 54 snapshot isolation support, 239, 242 TimesTen (in-memory database), 89 WAL-based replication, 160 XML support, 30 ordering, 339-352 by sequence numbers, 343-348 causal ordering, 339-343 partial order, 341 limits of total ordering, 493 total order broadcast, 348-352 Orleans (actor framework), 139 outliers (response time), 14 Oz (programming language), 504 P package managers, 428, 505 packet switching, 285 packets corruption of, 306 sending via UDP, 442 PageRank (algorithm), 49, 424 paging (see virtual memory) ParAccel (database), 93 parallel databases (see massively parallel pro‐ cessing) parallel execution of graph analysis algorithms, 426 queries in MPP databases, 216 Parquet (data format), 96, 131 (see also column-oriented storage) use in Hadoop, 414 partial failures, 275, 310 limping, 311 partial order, 341 partitioning, 199-218, 556 and replication, 200 in batch processing, 429 multi-partition operations, 514 enforcing constraints, 522 secondary index maintenance, 495 of key-value data, 201-205 by key range, 202 skew and hot spots, 205 rebalancing partitions, 209-214 automatic or manual rebalancing, 213 problems with hash mod N, 210 using dynamic partitioning, 212 using fixed number of partitions, 210 using N partitions per node, 212 replication and, 147 request routing, 214-216 secondary indexes, 206-209 document-based partitioning, 206 term-based partitioning, 208 serial execution of transactions and, 255 Paxos (consensus algorithm), 366 ballot number, 368 Multi-Paxos (total order broadcast), 367 percentiles, 14, 556 calculating efficiently, 16 importance of high percentiles, 16 use in service level agreements (SLAs), 15 Percona XtraBackup (MySQL tool), 156 performance describing, 13 of distributed transactions, 360 of in-memory databases, 89 of linearizability, 338 of multi-leader replication, 169 perpetual inconsistency, 525 pessimistic concurrency control, 261 phantoms (transaction isolation), 250 materializing conflicts, 251 preventing, in serializability, 259 physical clocks (see clocks) pickle (Python), 113 Pig (dataflow language), 419, 427 replicated joins, 409 skewed joins, 407 workflows, 403 Pinball (workflow scheduler), 402 pipelined execution, 423 in Unix, 394 point in time, 287 polyglot persistence, 29 polystores, 501 PostgreSQL (database) BDR (multi-leader replication), 170 causal ordering of writes, 177 Bottled Water (change data capture), 455 Bucardo (trigger-based replication), 161, 173 distributed transaction support, 361 foreign data wrappers, 501 full text search support, 490 leader-based replication, 153 log sequence number, 156 MVCC implementation, 239, 241 PL/pgSQL language, 255 PostGIS geospatial indexes, 87 preventing lost updates, 245 preventing write skew, 248, 261 read committed isolation, 236 recursive query support, 54 representing graphs, 51 Index | 579 serializable snapshot isolation (SSI), 261 snapshot isolation support, 239, 242 WAL-based replication, 160 XML and JSON support, 30, 42 pre-splitting, 212 Precision Time Protocol (PTP), 290 predicate locks, 259 predictive analytics, 533-536 amplifying bias, 534 ethics of (see ethics) feedback loops, 536 preemption of datacenter resources, 418 of threads, 298 Pregel processing model, 425 primary keys, 85, 556 compound primary key (Cassandra), 204 primary-secondary replication (see leaderbased replication) privacy, 536-543 consent and freedom of choice, 538 data as assets and power, 540 deleting data, 463 ethical considerations (see ethics) legislation and self-regulation, 542 meaning of, 539 surveillance, 537 tracking behavioral data, 536 probabilistic algorithms, 16, 466 process pauses, 295-299 processing time (of events), 469 producers (message streams), 440 programming languages dataflow languages, 504 for stored procedures, 255 functional reactive programming (FRP), 504 logic programming, 504 Prolog (language), 61 (see also Datalog) promises (asynchronous operations), 135 property graphs, 50 Cypher query language, 52 Protocol Buffers (data format), 117-121 field tags and schema evolution, 120 provenance of data, 531 publish/subscribe model, 441 publishers (message streams), 440 punch card tabulating machines, 390 580 | Index pure functions, 48 putting computation near data, 400 Q Qpid (messaging), 444 quality of service (QoS), 285 Quantcast File System (distributed filesystem), 398 query languages, 42-48 aggregation pipeline, 48 CSS and XSL, 44 Cypher, 52 Datalog, 60 Juttle, 504 MapReduce querying, 46-48 recursive SQL queries, 53 relational algebra and SQL, 42 SPARQL, 59 query optimizers, 37, 427 queueing delays (networks), 282 head-of-line blocking, 15 latency and response time, 14 queues (messaging), 137 quorums, 179-182, 556 for leaderless replication, 179 in consensus algorithms, 368 limitations of consistency, 181-183, 334 making decisions in distributed systems, 301 monitoring staleness, 182 multi-datacenter replication, 184 relying on durability, 309 sloppy quorums and hinted handoff, 183 R R-trees (indexes), 87 RabbitMQ (messaging), 137, 444 leader-based replication, 153 race conditions, 225 (see also concurrency) avoiding with linearizability, 331 caused by dual writes, 452 dirty writes, 235 in counter increments, 235 lost updates, 242-246 preventing with event logs, 462, 507 preventing with serializable isolation, 252 write skew, 246-251 Raft (consensus algorithm), 366 sensitivity to network problems, 369 term number, 368 use in etcd, 353 RAID (Redundant Array of Independent Disks), 7, 398 railways, schema migration on, 496 RAMCloud (in-memory storage), 89 ranking algorithms, 424 RDF (Resource Description Framework), 57 querying with SPARQL, 59 RDMA (Remote Direct Memory Access), 276 read committed isolation level, 234-237 implementing, 236 multi-version concurrency control (MVCC), 239 no dirty reads, 234 no dirty writes, 235 read path (derived data), 509 read repair (leaderless replication), 178 for linearizability, 335 read replicas (see leader-based replication) read skew (transaction isolation), 238, 266 as violation of causality, 340 read-after-write consistency, 163, 524 cross-device, 164 read-modify-write cycle, 243 read-scaling architecture, 161 reads as events, 513 real-time collaborative editing, 170 near-real-time processing, 390 (see also stream processing) publish/subscribe dataflow, 513 response time guarantees, 298 time-of-day clocks, 288 rebalancing partitions, 209-214, 556 (see also partitioning) automatic or manual rebalancing, 213 dynamic partitioning, 212 fixed number of partitions, 210 fixed number of partitions per node, 212 problems with hash mod N, 210 recency guarantee, 324 recommendation engines batch process outputs, 412 batch workflows, 403, 420 iterative processing, 424 statistical and numerical algorithms, 428 records, 399 events in stream processing, 440 recursive common table expressions (SQL), 54 redelivery (messaging), 445 Redis (database) atomic operations, 243 durability, 89 Lua scripting, 255 single-threaded execution, 253 usage example, 4 redundancy hardware components, 7 of derived data, 386 (see also derived data) Reed–Solomon codes (error correction), 398 refactoring, 22 (see also evolvability) regions (partitioning), 199 register (data structure), 325 relational data model, 28-42 comparison to document model, 38-42 graph queries in SQL, 53 in-memory databases with, 89 many-to-one and many-to-many relation‐ ships, 33 multi-object transactions, need for, 231 NoSQL as alternative to, 29 object-relational mismatch, 29 relational algebra and SQL, 42 versus document model convergence of models, 41 data locality, 41 relational databases eventual consistency, 162 history, 28 leader-based replication, 153 logical logs, 160 philosophy compared to Unix, 499, 501 schema changes, 40, 111, 130 statement-based replication, 158 use of B-tree indexes, 80 relationships (see edges) reliability, 6-10, 489 building a reliable system from unreliable components, 276 defined, 6, 22 hardware faults, 7 human errors, 9 importance of, 10 of messaging systems, 442 Index | 581 software errors, 8 Remote Method Invocation (Java RMI), 134 remote procedure calls (RPCs), 134-136 (see also services) based on futures, 135 data encoding and evolution, 136 issues with, 134 using Avro, 126, 135 using Thrift, 135 versus message brokers, 137 repeatable reads (transaction isolation), 242 replicas, 152 replication, 151-193, 556 and durability, 227 chain replication, 155 conflict resolution and, 246 consistency properties, 161-167 consistent prefix reads, 165 monotonic reads, 164 reading your own writes, 162 in distributed filesystems, 398 leaderless, 177-191 detecting concurrent writes, 184-191 limitations of quorum consistency, 181-183, 334 sloppy quorums and hinted handoff, 183 monitoring staleness, 182 multi-leader, 168-177 across multiple datacenters, 168, 335 handling write conflicts, 171-175 replication topologies, 175-177 partitioning and, 147, 200 reasons for using, 145, 151 single-leader, 152-161 failover, 157 implementation of replication logs, 158-161 relation to consensus, 367 setting up new followers, 155 synchronous versus asynchronous, 153-155 state machine replication, 349, 452 using erasure coding, 398 with heterogeneous data systems, 453 replication logs (see logs) reprocessing data, 496, 498 (see also evolvability) from log-based messaging, 451 request routing, 214-216 582 | Index approaches to, 214 parallel query execution, 216 resilient systems, 6 (see also fault tolerance) response time as performance metric for services, 13, 389 guarantees on, 298 latency versus, 14 mean and percentiles, 14 user experience, 15 responsibility and accountability, 535 REST (Representational State Transfer), 133 (see also services) RethinkDB (database) document data model, 31 dynamic partitioning, 212 join support, 34, 42 key-range partitioning, 202 leader-based replication, 153 subscribing to changes, 456 Riak (database) Bitcask storage engine, 72 CRDTs, 174, 191 dotted version vectors, 191 gossip protocol, 216 hash partitioning, 203-204, 211 last-write-wins conflict resolution, 186 leaderless replication, 177 LevelDB storage engine, 78 linearizability, lack of, 335 multi-datacenter support, 184 preventing lost updates across replicas, 246 rebalancing, 213 search feature, 209 secondary indexes, 207 siblings (concurrently written values), 190 sloppy quorums, 184 ring buffers, 450 Ripple (cryptocurrency), 532 rockets, 10, 36, 305 RocksDB (storage engine), 78 leveled compaction, 79 rollbacks (transactions), 222 rolling upgrades, 8, 112 routing (see request routing) row-oriented storage, 96 row-based replication, 160 rowhammer (memory corruption), 529 RPCs (see remote procedure calls) Rubygems (package manager), 428 rules (Datalog), 61 S safety and liveness properties, 308 in consensus algorithms, 366 in transactions, 222 sagas (see compensating transactions) Samza (stream processor), 466, 467 fault tolerance, 479 streaming SQL support, 466 sandboxes, 9 SAP HANA (database), 93 scalability, 10-18, 489 approaches for coping with load, 17 defined, 22 describing load, 11 describing performance, 13 partitioning and, 199 replication and, 161 scaling up versus scaling out, 146 scaling out, 17, 146 (see also shared-nothing architecture) scaling up, 17, 146 scatter/gather approach, querying partitioned databases, 207 SCD (slowly changing dimension), 476 schema-on-read, 39 comparison to evolvable schema, 128 in distributed filesystems, 415 schema-on-write, 39 schemaless databases (see schema-on-read) schemas, 557 Avro, 122-127 reader determining writer’s schema, 125 schema evolution, 123 dynamically generated, 126 evolution of, 496 affecting application code, 111 compatibility checking, 126 in databases, 129-131 in message-passing, 138 in service calls, 136 flexibility in document model, 39 for analytics, 93-95 for JSON and XML, 115 merits of, 127 schema migration on railways, 496 Thrift and Protocol Buffers, 117-121 schema evolution, 120 traditional approach to design, fallacy in, 462 searches building search indexes in batch processes, 411 k-nearest neighbors, 429 on streams, 467 partitioned secondary indexes, 206 secondaries (see leader-based replication) secondary indexes, 85, 557 partitioning, 206-209, 217 document-partitioned, 206 index maintenance, 495 term-partitioned, 208 problems with dual writes, 452, 491 updating, transaction isolation and, 231 secondary sorts, 405 sed (Unix tool), 392 self-describing files, 127 self-joins, 480 self-validating systems, 530 semantic web, 57 semi-synchronous replication, 154 sequence number ordering, 343-348 generators, 294, 344 insufficiency for enforcing constraints, 347 Lamport timestamps, 345 use of timestamps, 291, 295, 345 sequential consistency, 351 serializability, 225, 233, 251-266, 557 linearizability versus, 329 pessimistic versus optimistic concurrency control, 261 serial execution, 252-256 partitioning, 255 using stored procedures, 253, 349 serializable snapshot isolation (SSI), 261-266 detecting stale MVCC reads, 263 detecting writes that affect prior reads, 264 distributed execution, 265, 364 performance of SSI, 265 preventing write skew, 262-265 two-phase locking (2PL), 257-261 index-range locks, 260 performance, 258 Serializable (Java), 113 Index | 583 serialization, 113 (see also encoding) service discovery, 135, 214, 372 using DNS, 216, 372 service level agreements (SLAs), 15 service-oriented architecture (SOA), 132 (see also services) services, 131-136 microservices, 132 causal dependencies across services, 493 loose coupling, 502 relation to batch/stream processors, 389, 508 remote procedure calls (RPCs), 134-136 issues with, 134 similarity to databases, 132 web services, 132, 135 session windows (stream processing), 472 (see also windows) sessionization, 407 sharding (see partitioning) shared mode (locks), 258 shared-disk architecture, 146, 398 shared-memory architecture, 146 shared-nothing architecture, 17, 146-147, 557 (see also replication) distributed filesystems, 398 (see also distributed filesystems) partitioning, 199 use of network, 277 sharks biting undersea cables, 279 counting (example), 46-48 finding (example), 42 website about (example), 44 shredding (in relational model), 38 siblings (concurrent values), 190, 246 (see also conflicts) similarity search edit distance, 88 genome data, 63 k-nearest neighbors, 429 single-leader replication (see leader-based rep‐ lication) single-threaded execution, 243, 252 in batch processing, 406, 421, 426 in stream processing, 448, 463, 522 size-tiered compaction, 79 skew, 557 584 | Index clock skew, 291-294, 334 in transaction isolation read skew, 238, 266 write skew, 246-251, 262-265 (see also write skew) meanings of, 238 unbalanced workload, 201 compensating for, 205 due to celebrities, 205 for time-series data, 203 in batch processing, 407 slaves (see leader-based replication) sliding windows (stream processing), 472 (see also windows) sloppy quorums, 183 (see also quorums) lack of linearizability, 334 slowly changing dimension (data warehouses), 476 smearing (leap seconds adjustments), 290 snapshots (databases) causal consistency, 340 computing derived data, 500 in change data capture, 455 serializable snapshot isolation (SSI), 261-266, 329 setting up a new replica, 156 snapshot isolation and repeatable read, 237-242 implementing with MVCC, 239 indexes and MVCC, 241 visibility rules, 240 synchronized clocks for global snapshots, 294 snowflake schemas, 95 SOAP, 133 (see also services) evolvability, 136 software bugs, 8 maintaining integrity, 529 solid state drives (SSDs) access patterns, 84 detecting corruption, 519, 530 faults in, 227 sequential write throughput, 75 Solr (search server) building indexes in batch processes, 411 document-partitioned indexes, 207 request routing, 216 usage example, 4 use of Lucene, 79 sort (Unix tool), 392, 394, 395 sort-merge joins (MapReduce), 405 Sorted String Tables (see SSTables) sorting sort order in column storage, 99 source of truth (see systems of record) Spanner (database) data locality, 41 snapshot isolation using clocks, 295 TrueTime API, 294 Spark (processing framework), 421-423 bytecode generation, 428 dataflow APIs, 427 fault tolerance, 422 for data warehouses, 93 GraphX API (graph processing), 425 machine learning, 428 query optimizer, 427 Spark Streaming, 466 microbatching, 477 stream processing on top of batch process‐ ing, 495 SPARQL (query language), 59 spatial algorithms, 429 split brain, 158, 557 in consensus algorithms, 352, 367 preventing, 322, 333 using fencing tokens to avoid, 302-304 spreadsheets, dataflow programming capabili‐ ties, 504 SQL (Structured Query Language), 21, 28, 43 advantages and limitations of, 416 distributed query execution, 48 graph queries in, 53 isolation levels standard, issues with, 242 query execution on Hadoop, 416 résumé (example), 30 SQL injection vulnerability, 305 SQL on Hadoop, 93 statement-based replication, 158 stored procedures, 255 SQL Server (database) data warehousing support, 93 distributed transaction support, 361 leader-based replication, 153 preventing lost updates, 245 preventing write skew, 248, 257 read committed isolation, 236 recursive query support, 54 serializable isolation, 257 snapshot isolation support, 239 T-SQL language, 255 XML support, 30 SQLstream (stream analytics), 466 SSDs (see solid state drives) SSTables (storage format), 76-79 advantages over hash indexes, 76 concatenated index, 204 constructing and maintaining, 78 making LSM-Tree from, 78 staleness (old data), 162 cross-channel timing dependencies, 331 in leaderless databases, 178 in multi-version concurrency control, 263 monitoring for, 182 of client state, 512 versus linearizability, 324 versus timeliness, 524 standbys (see leader-based replication) star replication topologies, 175 star schemas, 93-95 similarity to event sourcing, 458 Star Wars analogy (event time versus process‐ ing time), 469 state derived from log of immutable events, 459 deriving current state from the event log, 458 interplay between state changes and appli‐ cation code, 507 maintaining derived state, 495 maintenance by stream processor in streamstream joins, 473 observing derived state, 509-515 rebuilding after stream processor failure, 478 separation of application code and, 505 state machine replication, 349, 452 statement-based replication, 158 statically typed languages analogy to schema-on-write, 40 code generation and, 127 statistical and numerical algorithms, 428 StatsD (metrics aggregator), 442 stdin, stdout, 395, 396 Stellar (cryptocurrency), 532 Index | 585 stock market feeds, 442 STONITH (Shoot The Other Node In The Head), 158 stop-the-world (see garbage collection) storage composing data storage technologies, 499-504 diversity of, in MapReduce, 415 Storage Area Network (SAN), 146, 398 storage engines, 69-104 column-oriented, 95-101 column compression, 97-99 defined, 96 distinction between column families and, 99 Parquet, 96, 131 sort order in, 99-100 writing to, 101 comparing requirements for transaction processing and analytics, 90-96 in-memory storage, 88 durability, 227 row-oriented, 70-90 B-trees, 79-83 comparing B-trees and LSM-trees, 83-85 defined, 96 log-structured, 72-79 stored procedures, 161, 253-255, 557 and total order broadcast, 349 pros and cons of, 255 similarity to stream processors, 505 Storm (stream processor), 466 distributed RPC, 468, 514 Trident state handling, 478 straggler events, 470, 498 stream processing, 464-481, 557 accessing external services within job, 474, 477, 478, 517 combining with batch processing lambda architecture, 497 unifying technologies, 498 comparison to batch processing, 464 complex event processing (CEP), 465 fault tolerance, 476-479 atomic commit, 477 idempotence, 478 microbatching and checkpointing, 477 rebuilding state after a failure, 478 for data integration, 494-498 586 | Index maintaining derived state, 495 maintenance of materialized views, 467 messaging systems (see messaging systems) reasoning about time, 468-472 event time versus processing time, 469, 477, 498 knowing when window is ready, 470 types of windows, 472 relation to databases (see streams) relation to services, 508 search on streams, 467 single-threaded execution, 448, 463 stream analytics, 466 stream joins, 472-476 stream-stream join, 473 stream-table join, 473 table-table join, 474 time-dependence of, 475 streams, 440-451 end-to-end, pushing events to clients, 512 messaging systems (see messaging systems) processing (see stream processing) relation to databases, 451-464 (see also changelogs) API support for change streams, 456 change data capture, 454-457 derivative of state by time, 460 event sourcing, 457-459 keeping systems in sync, 452-453 philosophy of immutable events, 459-464 topics, 440 strict serializability, 329 strong consistency (see linearizability) strong one-copy serializability, 329 subjects, predicates, and objects (in triplestores), 55 subscribers (message streams), 440 (see also consumers) supercomputers, 275 surveillance, 537 (see also privacy) Swagger (service definition format), 133 swapping to disk (see virtual memory) synchronous networks, 285, 557 comparison to asynchronous networks, 284 formal model, 307 synchronous replication, 154, 557 chain replication, 155 conflict detection, 172 system models, 300, 306-310 assumptions in, 528 correctness of algorithms, 308 mapping to the real world, 309 safety and liveness, 308 systems of record, 386, 557 change data capture, 454, 491 treating event log as, 460 systems thinking, 536 T t-digest (algorithm), 16 table-table joins, 474 Tableau (data visualization software), 416 tail (Unix tool), 447 tail vertex (property graphs), 51 Tajo (query engine), 93 Tandem NonStop SQL (database), 200 TCP (Transmission Control Protocol), 277 comparison to circuit switching, 285 comparison to UDP, 283 connection failures, 280 flow control, 282, 441 packet checksums, 306, 519, 529 reliability and duplicate suppression, 517 retransmission timeouts, 284 use for transaction sessions, 229 telemetry (see monitoring) Teradata (database), 93, 200 term-partitioned indexes, 208, 217 termination (consensus), 365 Terrapin (database), 413 Tez (dataflow engine), 421-423 fault tolerance, 422 support by higher-level tools, 427 thrashing (out of memory), 297 threads (concurrency) actor model, 138, 468 (see also message-passing) atomic operations, 223 background threads, 73, 85 execution pauses, 286, 296-298 memory barriers, 338 preemption, 298 single (see single-threaded execution) three-phase commit, 359 Thrift (data format), 117-121 BinaryProtocol, 118 CompactProtocol, 119 field tags and schema evolution, 120 throughput, 13, 390 TIBCO, 137 Enterprise Message Service, 444 StreamBase (stream analytics), 466 time concurrency and, 187 cross-channel timing dependencies, 331 in distributed systems, 287-299 (see also clocks) clock synchronization and accuracy, 289 relying on synchronized clocks, 291-295 process pauses, 295-299 reasoning about, in stream processors, 468-472 event time versus processing time, 469, 477, 498 knowing when window is ready, 470 timestamp of events, 471 types of windows, 472 system models for distributed systems, 307 time-dependence in stream joins, 475 time-of-day clocks, 288 timeliness, 524 coordination-avoiding data systems, 528 correctness of dataflow systems, 525 timeouts, 279, 557 dynamic configuration of, 284 for failover, 158 length of, 281 timestamps, 343 assigning to events in stream processing, 471 for read-after-write consistency, 163 for transaction ordering, 295 insufficiency for enforcing constraints, 347 key range partitioning by, 203 Lamport, 345 logical, 494 ordering events, 291, 345 Titan (database), 50 tombstones, 74, 191, 456 topics (messaging), 137, 440 total order, 341, 557 limits of, 493 sequence numbers or timestamps, 344 total order broadcast, 348-352, 493, 522 consensus algorithms and, 366-368 Index | 587 implementation in ZooKeeper and etcd, 370 implementing with linearizable storage, 351 using, 349 using to implement linearizable storage, 350 tracking behavioral data, 536 (see also privacy) transaction coordinator (see coordinator) transaction manager (see coordinator) transaction processing, 28, 90-95 comparison to analytics, 91 comparison to data warehousing, 93 transactions, 221-267, 558 ACID properties of, 223 atomicity, 223 consistency, 224 durability, 226 isolation, 225 compensating (see compensating transac‐ tions) concept of, 222 distributed transactions, 352-364 avoiding, 492, 502, 521-528 failure amplification, 364, 495 in doubt/uncertain status, 358, 362 two-phase commit, 354-359 use of, 360-361 XA transactions, 361-364 OLTP versus analytics queries, 411 purpose of, 222 serializability, 251-266 actual serial execution, 252-256 pessimistic versus optimistic concur‐ rency control, 261 serializable snapshot isolation (SSI), 261-266 two-phase locking (2PL), 257-261 single-object and multi-object, 228-232 handling errors and aborts, 231 need for multi-object transactions, 231 single-object writes, 230 snapshot isolation (see snapshots) weak isolation levels, 233-251 preventing lost updates, 242-246 read committed, 234-238 transitive closure (graph algorithm), 424 trie (data structure), 88 triggers (databases), 161, 441 implementing change data capture, 455 implementing replication, 161 588 | Index triple-stores, 55-59 SPARQL query language, 59 tumbling windows (stream processing), 472 (see also windows) in microbatching, 477 tuple spaces (programming model), 507 Turtle (RDF data format), 56 Twitter constructing home timelines (example), 11, 462, 474, 511 DistributedLog (event log), 448 Finagle (RPC framework), 135 Snowflake (sequence number generator), 294 Summingbird (processing library), 497 two-phase commit (2PC), 353, 355-359, 558 confusion with two-phase locking, 356 coordinator failure, 358 coordinator recovery, 363 how it works, 357 issues in practice, 363 performance cost, 360 transactions holding locks, 362 two-phase locking (2PL), 257-261, 329, 558 confusion with two-phase commit, 356 index-range locks, 260 performance of, 258 type checking, dynamic versus static, 40 U UDP (User Datagram Protocol) comparison to TCP, 283 multicast, 442 unbounded datasets, 439, 558 (see also streams) unbounded delays, 558 in networks, 282 process pauses, 296 unbundling databases, 499-515 composing data storage technologies, 499-504 federation versus unbundling, 501 need for high-level language, 503 designing applications around dataflow, 504-509 observing derived state, 509-515 materialized views and caching, 510 multi-partition data processing, 514 pushing state changes to clients, 512 uncertain (transaction status) (see in doubt) uniform consensus, 365 (see also consensus) uniform interfaces, 395 union type (in Avro), 125 uniq (Unix tool), 392 uniqueness constraints asynchronously checked, 526 requiring consensus, 521 requiring linearizability, 330 uniqueness in log-based messaging, 522 Unix philosophy, 394-397 command-line batch processing, 391-394 Unix pipes versus dataflow engines, 423 comparison to Hadoop, 413-414 comparison to relational databases, 499, 501 comparison to stream processing, 464 composability and uniform interfaces, 395 loose coupling, 396 pipes, 394 relation to Hadoop, 499 UPDATE statement (SQL), 40 updates preventing lost updates, 242-246 atomic write operations, 243 automatically detecting lost updates, 245 compare-and-set operations, 245 conflict resolution and replication, 246 using explicit locking, 244 preventing write skew, 246-251 V validity (consensus), 365 vBuckets (partitioning), 199 vector clocks, 191 (see also version vectors) vectorized processing, 99, 428 verification, 528-533 avoiding blind trust, 530 culture of, 530 designing for auditability, 531 end-to-end integrity checks, 531 tools for auditable data systems, 532 version control systems, reliance on immutable data, 463 version vectors, 177, 191 capturing causal dependencies, 343 versus vector clocks, 191 Vertica (database), 93 handling writes, 101 replicas using different sort orders, 100 vertical scaling (see scaling up) vertices (in graphs), 49 property graph model, 50 Viewstamped Replication (consensus algo‐ rithm), 366 view number, 368 virtual machines, 146 (see also cloud computing) context switches, 297 network performance, 282 noisy neighbors, 284 reliability in cloud services, 8 virtualized clocks in, 290 virtual memory process pauses due to page faults, 14, 297 versus memory management by databases, 89 VisiCalc (spreadsheets), 504 vnodes (partitioning), 199 Voice over IP (VoIP), 283 Voldemort (database) building read-only stores in batch processes, 413 hash partitioning, 203-204, 211 leaderless replication, 177 multi-datacenter support, 184 rebalancing, 213 reliance on read repair, 179 sloppy quorums, 184 VoltDB (database) cross-partition serializability, 256 deterministic stored procedures, 255 in-memory storage, 89 output streams, 456 secondary indexes, 207 serial execution of transactions, 253 statement-based replication, 159, 479 transactions in stream processing, 477 W WAL (write-ahead log), 82 web services (see services) Web Services Description Language (WSDL), 133 webhooks, 443 webMethods (messaging), 137 WebSocket (protocol), 512 Index | 589 windows (stream processing), 466, 468-472 infinite windows for changelogs, 467, 474 knowing when all events have arrived, 470 stream joins within a window, 473 types of windows, 472 winners (conflict resolution), 173 WITH RECURSIVE syntax (SQL), 54 workflows (MapReduce), 402 outputs, 411-414 key-value stores, 412 search indexes, 411 with map-side joins, 410 working set, 393 write amplification, 84 write path (derived data), 509 write skew (transaction isolation), 246-251 characterizing, 246-251, 262 examples of, 247, 249 materializing conflicts, 251 occurrence in practice, 529 phantoms, 250 preventing in snapshot isolation, 262-265 in two-phase locking, 259-261 options for, 248 write-ahead log (WAL), 82, 159 writes (database) atomic write operations, 243 detecting writes affecting prior reads, 264 preventing dirty writes with read commit‐ ted, 235 WS-* framework, 133 (see also services) WS-AtomicTransaction (2PC), 355 590 | Index X XA transactions, 355, 361-364 heuristic decisions, 363 limitations of, 363 xargs (Unix tool), 392, 396 XML binary variants, 115 encoding RDF data, 57 for application data, issues with, 114 in relational databases, 30, 41 XSL/XPath, 45 Y Yahoo!


pages: 1,237 words: 227,370

Designing Data-Intensive Applications: The Big Ideas Behind Reliable, Scalable, and Maintainable Systems by Martin Kleppmann

active measures, Amazon Web Services, bitcoin, blockchain, business intelligence, business process, c2.com, cloud computing, collaborative editing, commoditize, conceptual framework, cryptocurrency, database schema, DevOps, distributed ledger, Donald Knuth, Edward Snowden, Ethereum, ethereum blockchain, fault tolerance, finite state, Flash crash, full text search, general-purpose programming language, informal economy, information retrieval, Infrastructure as a Service, Internet of things, iterative process, John von Neumann, Kubernetes, loose coupling, Marc Andreessen, microservices, natural language processing, Network effects, packet switching, peer-to-peer, performance metric, place-making, premature optimization, recommendation engine, Richard Feynman, self-driving car, semantic web, Shoshana Zuboff, social graph, social web, software as a service, software is eating the world, sorting algorithm, source of truth, SPARQL, speech recognition, statistical model, undersea cable, web application, WebSocket, wikimedia commons

And this is why MapReduce is designed to tolerate frequent unexpected task termination: it’s not because the hardware is particularly unreliable, it’s because the freedom to arbitrarily terminate processes enables better resource utilization in a computing cluster. Among open source cluster schedulers, preemption is less widely used. YARN’s CapacityScheduler supports preemption for balancing the resource allocation of different queues [58], but general priority preemption is not supported in YARN, Mesos, or Kubernetes at the time of writing [60]. In an environment where tasks are not so often terminated, the design decisions of MapReduce make less sense. In the next section, we will look at some alternatives to MapReduce that make different design decisions. Beyond MapReduce Although MapReduce became very popular and received a lot of hype in the late 2000s, it is just one among many possible programming models for distributed systems.

They do not fit well with the requirements of modern application development, such as dependency and package management, version control, rolling upgrades, evolvability, monitoring, metrics, calls to network services, and integration with external systems. On the other hand, deployment and cluster management tools such as Mesos, YARN, Docker, Kubernetes, and others are designed specifically for the purpose of running application code. By focusing on doing one thing well, they are able to do it much better than a database that provides execution of user-defined functions as one of its many features. I think it makes sense to have some parts of a system that specialize in durable data storage, and other parts that specialize in running application code.

InfiniteGraph (database), Graph-Like Data Models InnoDB (storage engine)clustered index on primary key, Storing values within the index not preventing lost updates, Automatically detecting lost updates preventing write skew, Characterizing write skew, Implementation of two-phase locking serializable isolation, Implementation of two-phase locking snapshot isolation support, Snapshot Isolation and Repeatable Read inside-out databases, Designing Applications Around Dataflow(see also unbundling databases) integrating different data systems (see data integration) integrity, Timeliness and Integritycoordination-avoiding data systems, Coordination-avoiding data systems correctness of dataflow systems, Correctness of dataflow systems in consensus formalization, Fault-Tolerant Consensus integrity checks, Don’t just blindly trust what they promise(see also auditing) end-to-end, The end-to-end argument, The end-to-end argument again use of snapshot isolation, Snapshot Isolation and Repeatable Read maintaining despite software bugs, Maintaining integrity in the face of software bugs Interface Definition Language (IDL), Thrift and Protocol Buffers, Avro intermediate state, materialization of, Materialization of Intermediate State-Discussion of materialization internet services, systems for implementing, Cloud Computing and Supercomputing invariants, Consistency(see also constraints) inversion of control, Separation of logic and wiring IP (Internet Protocol)unreliability of, Cloud Computing and Supercomputing ISDN (Integrated Services Digital Network), Synchronous Versus Asynchronous Networks isolation (in transactions), Isolation, Single-Object and Multi-Object Operations, Glossarycorrectness and, Aiming for Correctness for single-object writes, Single-object writes serializability, Serializability-Performance of serializable snapshot isolationactual serial execution, Actual Serial Execution-Summary of serial execution serializable snapshot isolation (SSI), Serializable Snapshot Isolation (SSI)-Performance of serializable snapshot isolation two-phase locking (2PL), Two-Phase Locking (2PL)-Index-range locks violating, Single-Object and Multi-Object Operations weak isolation levels, Weak Isolation Levels-Materializing conflictspreventing lost updates, Preventing Lost Updates-Conflict resolution and replication read committed, Read Committed-Implementing read committed snapshot isolation, Snapshot Isolation and Repeatable Read-Repeatable read and naming confusion iterative processing, Graphs and Iterative Processing-Parallel execution J Java Database Connectivity (JDBC)distributed transaction support, XA transactions network drivers, The Merits of Schemas Java Enterprise Edition (EE), The problems with remote procedure calls (RPCs), Introduction to two-phase commit, XA transactions Java Message Service (JMS), Message brokers compared to databases(see also messaging systems) comparison to log-based messaging, Logs compared to traditional messaging, Replaying old messages distributed transaction support, XA transactions message ordering, Acknowledgments and redelivery Java Transaction API (JTA), Introduction to two-phase commit, XA transactions Java Virtual Machine (JVM)bytecode generation, The move toward declarative query languages garbage collection pauses, Process Pauses process reuse in batch processors, Dataflow engines JavaScriptin MapReduce querying, MapReduce Querying setting element styles (example), Declarative Queries on the Web use in advanced queries, MapReduce Querying Jena (RDF framework), The RDF data model Jepsen (fault tolerance testing), Aiming for Correctness jitter (network delay), Network congestion and queueing joins, Glossaryby index lookup, Reduce-Side Joins and Grouping expressing as relational operators, The move toward declarative query languages in relational and document databases, Many-to-One and Many-to-Many Relationships MapReduce map-side joins, Map-Side Joins-MapReduce workflows with map-side joinsbroadcast hash joins, Broadcast hash joins merge joins, Map-side merge joins partitioned hash joins, Partitioned hash joins MapReduce reduce-side joins, Reduce-Side Joins and Grouping-Handling skewhandling skew, Handling skew sort-merge joins, Sort-merge joins parallel execution of, Comparing Hadoop to Distributed Databases secondary indexes and, Other Indexing Structures stream joins, Stream Joins-Time-dependence of joinsstream-stream join, Stream-stream join (window join) stream-table join, Stream-table join (stream enrichment) table-table join, Table-table join (materialized view maintenance) time-dependence of, Time-dependence of joins support in document databases, Convergence of document and relational databases JOTM (transaction coordinator), Introduction to two-phase commit JSONAvro schema representation, Avro binary variants, Binary encoding for application data, issues with, JSON, XML, and Binary Variants in relational databases, The Object-Relational Mismatch, Convergence of document and relational databases representing a résumé (example), The Object-Relational Mismatch Juttle (query language), Designing Applications Around Dataflow K k-nearest neighbors, Specialization for different domains Kafka (messaging), Message brokers, Using logs for message storageKafka Connect (database integration), API support for change streams, Deriving several views from the same event log Kafka Streams (stream processor), Stream analytics, Maintaining materialized viewsfault tolerance, Rebuilding state after a failure leader-based replication, Leaders and Followers log compaction, Log compaction, Maintaining materialized views message offsets, Using logs for message storage, Idempotence request routing, Request Routing transaction support, Atomic commit revisited usage example, Thinking About Data Systems Ketama (partitioning library), Partitioning proportionally to nodes key-value stores, Data Structures That Power Your Databaseas batch process output, Key-value stores as batch process output hash indexes, Hash Indexes-Hash Indexes in-memory, Keeping everything in memory partitioning, Partitioning of Key-Value Data-Skewed Workloads and Relieving Hot Spotsby hash of key, Partitioning by Hash of Key, Summary by key range, Partitioning by Key Range, Summary dynamic partitioning, Dynamic partitioning skew and hot spots, Skewed Workloads and Relieving Hot Spots Kryo (Java), Language-Specific Formats Kubernetes (cluster manager), Designing for frequent faults, Separation of application code and state L lambda architecture, The lambda architecture Lamport timestamps, Lamport timestamps Large Hadron Collider (LHC), Summary last write wins (LWW), Converging toward a consistent state, Implementing Linearizable Systemsdiscarding concurrent writes, Last write wins (discarding concurrent writes) problems with, Timestamps for ordering events prone to lost updates, Conflict resolution and replication late binding, Separation of logic and wiring latencyinstability under two-phase locking, Performance of two-phase locking network latency and resource utilization, Can we not simply make network delays predictable?


pages: 333 words: 64,581

Clean Agile: Back to Basics by Robert C. Martin

Alan Turing: On Computable Numbers, with an Application to the Entscheidungsproblem, c2.com, continuous integration, DevOps, double entry bookkeeping, en.wikipedia.org, failed state, Frederick Winslow Taylor, index card, iterative process, Kubernetes, loose coupling, microservices, remote working, revision control, Turing machine

Various data formats (JSON, XML, YAML, etc.) and markup languages (including HTML) Command-line and script-based interaction with the operating system Source repository tools (Git. Is there another choice?) Continuous integration/build tools (Jenkins, TeamCity, GoCD, etc.) Deployment/server management tools (Docker, Kubernetes, Ansible, Chef, Puppet, etc.) Communication tools: email, Slack, the English language (!) Testing tools (unit testing frameworks, Cucumber, Selenium, etc.) These categories of tools are essential to building software. Without them, it is impossible to deliver anything in today’s world.


pages: 540 words: 103,101

Building Microservices by Sam Newman

airport security, Amazon Web Services, anti-pattern, business process, call centre, continuous integration, create, read, update, delete, defense in depth, don't repeat yourself, Edward Snowden, fault tolerance, index card, information retrieval, Infrastructure as a Service, inventory management, job automation, Kubernetes, load shedding, loose coupling, microservices, MITM: man-in-the-middle, platform as a service, premature optimization, pull request, recommendation engine, social graph, software as a service, source of truth, the built environment, web application, WebSocket

If you want tools to help you manage services across multiple Docker instances across multiple machines, you’ll need to look at other software that adds these capabilities. There is a key need for a scheduling layer that lets you request a container and then finds a Docker container that can run it for you. In this space, Google’s recently open sourced Kubernetes and CoreOS’s cluster technology can help, and it seems every month there is a new entrant in this space. Deis is another interesting tool based on Docker, which is attempting to provide a Heroku-like PaaS on top of Docker. I talked earlier about PaaS solutions. My struggle with them has always been that they often get the abstraction level wrong, and that self-hosted solutions lag significantly behind hosted solutions like Heroku.


pages: 514 words: 111,012

The Art of Monitoring by James Turnbull

Amazon Web Services, anti-pattern, cloud computing, continuous integration, correlation does not imply causation, Debian, DevOps, domain-specific language, failed state, Kickstarter, Kubernetes, microservices, performance metric, pull request, Ruby on Rails, software as a service, source of truth, web application, WebSocket

There are not many standalone open-source tools available yet, but there are some notable initial entrants. cAdvisor — A Google-developed tool that analyzes the resource usage and performance characteristics of running containers. Can run as a container on a Docker server. Heapster — If you're using Kubernetes, Heapster does resource analysis and monitoring of container clusters. collectd cgroups — Docker uses cgroups or control groups to help allocate resources like CPU. You can use the collectd cgroups plugin to gather this information. Summary In this chapter we've looked at Docker and monitoring containers.


pages: 409 words: 112,055

The Fifth Domain: Defending Our Country, Our Companies, and Ourselves in the Age of Cyber Threats by Richard A. Clarke, Robert K. Knake

A Declaration of the Independence of Cyberspace, Affordable Care Act / Obamacare, Airbnb, Albert Einstein, Amazon Web Services, autonomous vehicles, barriers to entry, bitcoin, Black Swan, blockchain, borderless world, business cycle, business intelligence, call centre, Cass Sunstein, cloud computing, cognitive bias, commoditize, computer vision, corporate governance, cryptocurrency, data acquisition, DevOps, don't be evil, Donald Trump, Edward Snowden, Exxon Valdez, global village, immigration reform, Infrastructure as a Service, Internet of things, Jeff Bezos, Julian Assange, Kubernetes, Mark Zuckerberg, Metcalfe’s law, MITM: man-in-the-middle, move fast and break things, move fast and break things, Network effects, open borders, platform as a service, Ponzi scheme, ransomware, Richard Thaler, Sand Hill Road, Schrödinger's Cat, self-driving car, shareholder value, Silicon Valley, Silicon Valley startup, Skype, smart cities, Snapchat, software as a service, Steven Levy, Stuxnet, technoutopianism, Tim Cook: Apple, undersea cable, WikiLeaks, Y2K, zero day

Just what they sound like, containers run on top of virtual machines and keep applications running in the cloud from interacting with other applications running in the same virtual machine. Thus, a vulnerability in one piece of software running inside a container would not let an attacker gain access to data in another. Instead of installing lots of software in system libraries throughout a computer, containers like Docker or Kubernetes keep all the files for the software to execute within the container. For advanced threat actors, compromising the security of a cloud provider to get to your data may be something they are more than willing and able to do. Thus, the cloud may not be the right solution for all applications. The U.S. government, while taking a “cloud-first” approach in the Obama administration, has been noticeably slow to make the transition, instead consolidating its own data centers.


pages: 509 words: 132,327

Rise of the Machines: A Cybernetic History by Thomas Rid

1960s counterculture, A Declaration of the Independence of Cyberspace, agricultural Revolution, Albert Einstein, Alistair Cooke, Apple II, Apple's 1984 Super Bowl advert, back-to-the-land, Berlin Wall, British Empire, Brownian motion, Buckminster Fuller, business intelligence, Charles Lindbergh, Claude Shannon: information theory, conceptual framework, connected car, domain-specific language, Douglas Engelbart, Douglas Engelbart, dumpster diving, Extropian, full employment, game design, global village, Haight Ashbury, Howard Rheingold, Jaron Lanier, job automation, John Markoff, John von Neumann, Kevin Kelly, Kubernetes, Marshall McLuhan, Menlo Park, Mitch Kapor, Mother of all demos, new economy, New Journalism, Norbert Wiener, offshore financial centre, oil shale / tar sands, pattern recognition, RAND corporation, Silicon Valley, Simon Singh, speech recognition, Steve Jobs, Steve Wozniak, Steven Levy, Stewart Brand, technoutopianism, Telecommunications Act of 1996, telepresence, The Hackers Conference, Vernor Vinge, Whole Earth Catalog, Whole Earth Review, Y2K, Yom Kippur War, Zimmermann PGP

Norbert Wiener’s initial cybernetic research was a $2,325 defense contract approved in December 1940. Here, Wiener (center) is pictured withtwo senior army officers: Brigadier General Leonard Greely (left) and Colonel Donald B. Diehl (right). In the summer of 1946, Wiener coined the term “cybernetics” from the Greek kubernetes, for “steersman,” inspired by observing man-machine servomechanical systems at war. W. Ross Ashby was an early British cyberneticist and innovator. He built the homeostat in 1946, a contraption then touted as the world’s first “thinking machine,” based on Royal Air Force bomb switches. Ashby’s homeostat was not just self-adaptive, but both system and environment at the same time.