Designs, Lessons and Advice from Building Large Distributed Systems
6 likes752 views
- MapReduce is a programming model for processing large datasets in a distributed manner across clusters of machines. It handles parallelization, load balancing, and hardware failures automatically. - In MapReduce, the input data is mapped to intermediate key-value pairs, shuffled and sorted by the keys, then reduced to produce the final output. This pattern applies to many large-scale computing problems. - Google uses MapReduce for tasks like generating map tiles, processing web search query logs, and mining user activity data from clickstreams. It hides the complex distributed systems details from programmers.
Designs, Lessons and Advice from Building Large Distributed Systems
- 1. Designs, Lessons and Advice from Building Large Distributed Systems Jeff Dean Google Fellow jeff@google.com
- 2. Computing shifting to really small and really big devices UI-centric devices Large consolidated computing farms
- 3. Google’s data center at The Dalles, OR
- 4. The Machinery Servers • CPUs • DRAM • Disks Clusters Racks • 40-80 servers • Ethernet switch
- 5. Architectural view of the storage hierarchy P P P P One server L1$ … L1$ L1$ … L1$ DRAM: 16GB, 100ns, 20GB/s L2$ … L2$ Disk: 2TB, 10ms, 200MB/s Disk Local DRAM
- 6. Architectural view of the storage hierarchy P P P P One server L1$ … L1$ L1$ … L1$ DRAM: 16GB, 100ns, 20GB/s L2$ … L2$ Disk: 2TB, 10ms, 200MB/s Disk Local DRAM Rack Switch Local rack (80 servers) Disk Disk DRAM DRAM DRAM Disk Disk … DRAM DRAM DRAM Disk Disk DRAM: 1TB, 300us, 100MB/s Disk Disk Disk: 160TB, 11ms, 100MB/s DRAM DRAM
- 7. Architectural view of the storage hierarchy P P P P One server L1$ … L1$ L1$ … L1$ DRAM: 16GB, 100ns, 20GB/s L2$ … L2$ Disk: 2TB, 10ms, 200MB/s Disk Local DRAM Rack Switch Local rack (80 servers) Disk Disk DRAM DRAM DRAM Disk Disk … DRAM DRAM DRAM Disk Disk DRAM: 1TB, 300us, 100MB/s Disk Disk Disk: 160TB, 11ms, 100MB/s DRAM DRAM Cluster Switch Cluster (30+ racks) … DRAM: 30TB, 500us, 10MB/s Disk: 4.80PB, 12ms, 10MB/s
- 8. Storage hierarchy: a different view A bumpy ride that has been getting bumpier over time
- 9. Reliability & Availability • Things will crash. Deal with it! – Assume you could start with super reliable servers (MTBF of 30 years) – Build computing system with 10 thousand of those – Watch one fail per day • Fault-tolerant software is inevitable • Typical yearly flakiness metrics – 1-5% of your disk drives will die – Servers will crash at least twice (2-4% failure rate)
- 10. The Joys of Real Hardware Typical first year for a new cluster: ~0.5 overheating (power down most machines in <5 mins, ~1-2 days to recover) ~1 PDU failure (~500-1000 machines suddenly disappear, ~6 hours to come back) ~1 rack-move (plenty of warning, ~500-1000 machines powered down, ~6 hours) ~1 network rewiring (rolling ~5% of machines down over 2-day span) ~20 rack failures (40-80 machines instantly disappear, 1-6 hours to get back) ~5 racks go wonky (40-80 machines see 50% packetloss) ~8 network maintenances (4 might cause ~30-minute random connectivity losses) ~12 router reloads (takes out DNS and external vips for a couple minutes) ~3 router failures (have to immediately pull traffic for an hour) ~dozens of minor 30-second blips for dns ~1000 individual machine failures ~thousands of hard drive failures slow disks, bad memory, misconfigured machines, flaky machines, etc. Long distance links: wild dogs, sharks, dead horses, drunken hunters, etc.
- 11. Understanding downtime behavior matters
- 12. Understanding downtime behavior matters
- 13. Understanding downtime behavior matters
- 14. Understanding fault statistics matters
- 15. Understanding fault statistics matters Can we expect faults to be independent or correlated? Are there common failure patterns we should program around?
- 16. Google Cluster Environment • Cluster is 1000s of machines, typically one or handful of configurations • File system (GFS) + Cluster scheduling system are core services • Typically 100s to 1000s of active jobs (some w/1 task, some w/1000s) • mix of batch and low-latency, user-facing production jobs scheduling master job 1 job 3 task task ... job 12 task job 7 job 3 task task job 5 task GFS chunk server scheduling slave ... chunk server scheduling slave master Linux Linux Chubby Commodity HW Commodity HW lock service Machine 1 Machine N
- 17. GFS Design GFS Master Misc. servers Replicas Client Masters GFS Master Client Client C0 C1 C1 C0 C5 C5 C2 C5 C3 … C2 Chunkserver 1 Chunkserver 2 Chunkserver N • Master manages metadata • Data transfers happen directly between clients/ chunkservers • Files broken into chunks (typically 64 MB)
- 18. GFS Usage @ Google • 200+ clusters • Many clusters of 1000s of machines • Pools of 1000s of clients • 4+ PB Filesystems • 40 GB/s read/write load – (in the presence of frequent HW failures)
- 19. Google: Most Systems are Distributed Systems • Distributed systems are a must: – data, request volume or both are too large for single machine • careful design about how to partition problems • need high capacity systems even within a single datacenter –multiple datacenters, all around the world • almost all products deployed in multiple locations –services used heavily even internally • a web search touches 50+ separate services, 1000s machines
- 20. Many Internal Services • Simpler from a software engineering standpoint – few dependencies, clearly specified (Protocol Buffers) – easy to test new versions of individual services – ability to run lots of experiments • Development cycles largely decoupled – lots of benefits: small teams can work independently – easier to have many engineering offices around the world
- 21. Protocol Buffers • Good protocol description language is vital • Desired attributes: – self-describing, multiple language support – efficient to encode/decode (200+ MB/s), compact serialized form Our solution: Protocol Buffers (in active use since 2000) message SearchResult { required int32 estimated_results = 1; // (1 is the tag number) optional string error_message = 2; repeated group Result = 3 { required float score = 4; required fixed64 docid = 5; optional message<WebResultDetails> = 6; … } };
- 22. Protocol Buffers (cont) • Automatically generated language wrappers • Graceful client and server upgrades – systems ignore tags they don't understand, but pass the information through (no need to upgrade intermediate servers) • Serialization/deserialization – high performance (200+ MB/s encode/decode) – fairly compact (uses variable length encodings) – format used to store data persistently (not just for RPCs) • Also allow service specifications: service Search { rpc DoSearch(SearchRequest) returns (SearchResponse); rpc DoSnippets(SnippetRequest) returns (SnippetResponse); rpc Ping(EmptyMessage) returns (EmptyMessage) { { protocol=udp; }; }; • Open source version: http://code.google.com/p/protobuf/
- 23. Designing Efficient Systems Given a basic problem definition, how do you choose the "best" solution? • Best could be simplest, highest performance, easiest to extend, etc. Important skill: ability to estimate performance of a system design – without actually having to build it!
- 24. Numbers Everyone Should Know L1 cache reference 0.5 ns Branch mispredict 5 ns L2 cache reference 7 ns Mutex lock/unlock 25 ns Main memory reference 100 ns Compress 1K bytes with Zippy 3,000 ns Send 2K bytes over 1 Gbps network 20,000 ns Read 1 MB sequentially from memory 250,000 ns Round trip within same datacenter 500,000 ns Disk seek 10,000,000 ns Read 1 MB sequentially from disk 20,000,000 ns Send packet CA->Netherlands->CA 150,000,000 ns
- 25. Back of the Envelope Calculations How long to generate image results page (30 thumbnails)? Design 1: Read serially, thumbnail 256K images on the fly 30 seeks * 10 ms/seek + 30 * 256K / 30 MB/s = 560 ms
- 26. Back of the Envelope Calculations How long to generate image results page (30 thumbnails)? Design 1: Read serially, thumbnail 256K images on the fly 30 seeks * 10 ms/seek + 30 * 256K / 30 MB/s = 560 ms Design 2: Issue reads in parallel: 10 ms/seek + 256K read / 30 MB/s = 18 ms (Ignores variance, so really more like 30-60 ms, probably)
- 27. Back of the Envelope Calculations How long to generate image results page (30 thumbnails)? Design 1: Read serially, thumbnail 256K images on the fly 30 seeks * 10 ms/seek + 30 * 256K / 30 MB/s = 560 ms Design 2: Issue reads in parallel: 10 ms/seek + 256K read / 30 MB/s = 18 ms (Ignores variance, so really more like 30-60 ms, probably) Lots of variations: – caching (single images? whole sets of thumbnails?) – pre-computing thumbnails –… Back of the envelope helps identify most promising…
- 28. Write Microbenchmarks! • Great to understand performance – Builds intuition for back-of-the-envelope calculations • Reduces cycle time to test performance improvements Benchmark Time(ns) CPU(ns) Iterations BM_VarintLength32/0 2 2 291666666 BM_VarintLength32Old/0 5 5 124660869 BM_VarintLength64/0 8 8 89600000 BM_VarintLength64Old/0 25 24 42164705 BM_VarintEncode32/0 7 7 80000000 BM_VarintEncode64/0 18 16 39822222 BM_VarintEncode64Old/0 24 22 31165217
- 29. Know Your Basic Building Blocks Core language libraries, basic data structures, protocol buffers, GFS, BigTable, indexing systems, MySQL, MapReduce, … Not just their interfaces, but understand their implementations (at least at a high level) If you don’t know what’s going on, you can’t do decent back-of-the-envelope calculations!
- 30. Encoding Your Data • CPUs are fast, memory/bandwidth are precious, ergo… – Variable-length encodings – Compression – Compact in-memory representations • Compression/encoding very important for many systems – inverted index posting list formats – storage systems for persistent data • We have lots of core libraries in this area – Many tradeoffs: space, encoding/decoding speed, etc. E.g.: • Zippy: encode@300 MB/s, decode@600MB/s, 2-4X compression • gzip: encode@25MB/s, decode@200MB/s, 4-6X compression
- 31. Designing & Building Infrastructure Identify common problems, and build software systems to address them in a general way • Important not to try to be all things to all people – Clients might be demanding 8 different things – Doing 6 of them is easy – …handling 7 of them requires real thought – …dealing with all 8 usually results in a worse system • more complex, compromises other clients in trying to satisfy everyone Don't build infrastructure just for its own sake • Identify common needs and address them • Don't imagine unlikely potential needs that aren't really there • Best approach: use your own infrastructure (especially at first!) – (much more rapid feedback about what works, what doesn't)
- 32. Design for Growth Try to anticipate how requirements will evolve keep likely features in mind as you design base system Ensure your design works if scale changes by 10X or 20X but the right solution for X often not optimal for 100X
- 33. Interactive Apps: Design for Low Latency • Aim for low avg. times (happy users!) – 90%ile and 99%ile also very important – Think about how much data you’re shuffling around • e.g. dozens of 1 MB RPCs per user request -> latency will be lousy • Worry about variance! – Redundancy or timeouts can help bring in latency tail • Judicious use of caching can help • Use higher priorities for interactive requests • Parallelism helps!
- 34. Making Applications Robust Against Failures Canary requests Failover to other replicas/datacenters Bad backend detection: stop using for live requests until behavior gets better More aggressive load balancing when imbalance is more severe Make your apps do something reasonable even if not all is right – Better to give users limited functionality than an error page
- 35. Add Sufficient Monitoring/Status/Debugging Hooks All our servers: • Export HTML-based status pages for easy diagnosis • Export a collection of key-value pairs via a standard interface – monitoring systems periodically collect this from running servers • RPC subsystem collects sample of all requests, all error requests, all requests >0.0s, >0.05s, >0.1s, >0.5s, >1s, etc. • Support low-overhead online profiling – cpu profiling – memory profiling – lock contention profiling If your system is slow or misbehaving, can you figure out why?
- 36. MapReduce • A simple programming model that applies to many large-scale computing problems • Hide messy details in MapReduce runtime library: – automatic parallelization – load balancing – network and disk transfer optimizations – handling of machine failures – robustness – improvements to core library benefit all users of library!
- 37. Typical problem solved by MapReduce • Read a lot of data • Map: extract something you care about from each record • Shuffle and Sort • Reduce: aggregate, summarize, filter, or transform • Write the results Outline stays the same, map and reduce change to fit the problem
- 38. Example: Rendering Map Tiles Input Map Shuffle Reduce Output Emit each to all Render tile using Geographic Sort by key overlapping latitude- data for all enclosed Rendered tiles feature list (key= Rect. Id) longitude rectangles features I-5 (0, I-5) (0, I-5) 0 Lake Washington (1, I-5) (0, Lake Wash.) WA-520 (0, Lake Wash.) (0, WA-520) I-90 (1, Lake Wash.) … … (0, WA-520) (1, I-90) … (1, I-5) 1 (1, Lake Wash.) (1, I-90) …
- 39. Parallel MapReduce Input data Map Map Map Map Master Shuffle Shuffle Shuffle Reduce Reduce Reduce Partitioned output
- 40. Parallel MapReduce Input data Map Map Map Map Master Shuffle Shuffle Shuffle Reduce Reduce Reduce Partitioned output For large enough problems, it’s more about disk and network performance than CPU & DRAM
- 41. MapReduce Usage Statistics Over Time Aug, ‘04 Mar, ‘06 Sep, '07 Sep, ’09 Number of jobs 29K 171K 2,217K 3,467K Average completion time (secs) 634 874 395 475 Machine years used 217 2,002 11,081 25,562 Input data read (TB) 3,288 52,254 403,152 544,130 Intermediate data (TB) 758 6,743 34,774 90,120 Output data written (TB) 193 2,970 14,018 57,520 Average worker machines 157 268 394 488
- 42. MapReduce in Practice • Abstract input and output interfaces – lots of MR operations don’t just read/write simple files • B-tree files • memory-mapped key-value stores • complex inverted index file formats • BigTable tables • SQL databases, etc. • ... • Low-level MR interfaces are in terms of byte arrays – Hardly ever use textual formats, though: slow, hard to parse – Most input & output is in encoded Protocol Buffer format • See “MapReduce: A Flexible Data Processing Tool” (to appear in upcoming CACM)
- 43. BigTable: Motivation • Lots of (semi-)structured data at Google – URLs: • Contents, crawl metadata, links, anchors, pagerank, … – Per-user data: • User preference settings, recent queries/search results, … – Geographic locations: • Physical entities (shops, restaurants, etc.), roads, satellite image data, user annotations, … • Scale is large – billions of URLs, many versions/page (~20K/version) – Hundreds of millions of users, thousands of q/sec – 100TB+ of satellite image data
- 44. Basic Data Model • Distributed multi-dimensional sparse map (row, column, timestamp) → cell contents Columns Rows • Rows are ordered lexicographically • Good match for most of our applications
- 45. Basic Data Model • Distributed multi-dimensional sparse map (row, column, timestamp) → cell contents “contents:” Columns Rows “www.cnn.com” “<html>…” • Rows are ordered lexicographically • Good match for most of our applications
- 46. Basic Data Model • Distributed multi-dimensional sparse map (row, column, timestamp) → cell contents “contents:” Columns Rows “www.cnn.com” “<html>…” t17 Timestamps • Rows are ordered lexicographically • Good match for most of our applications
- 47. Basic Data Model • Distributed multi-dimensional sparse map (row, column, timestamp) → cell contents “contents:” Columns Rows “www.cnn.com” t11 “<html>…” t17 Timestamps • Rows are ordered lexicographically • Good match for most of our applications
- 48. Basic Data Model • Distributed multi-dimensional sparse map (row, column, timestamp) → cell contents “contents:” Columns Rows t3 “www.cnn.com” t11 “<html>…” t17 Timestamps • Rows are ordered lexicographically • Good match for most of our applications
- 49. Tablets & Splitting “language:” “contents:” “aaa.com” “cnn.com” EN “<html>…” “cnn.com/sports.html” … “website.com” … “zuppa.com/menu.html”
- 50. Tablets & Splitting “language:” “contents:” “aaa.com” “cnn.com” EN “<html>…” “cnn.com/sports.html” Tablets … “website.com” … “zuppa.com/menu.html”
- 51. Tablets & Splitting “language:” “contents:” “aaa.com” “cnn.com” EN “<html>…” “cnn.com/sports.html” Tablets … “website.com” … “yahoo.com/kids.html” … “yahoo.com/kids.html0” … “zuppa.com/menu.html”
- 52. BigTable System Structure Bigtable Cell Bigtable master Bigtable tablet server Bigtable tablet server … Bigtable tablet server
- 53. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server
- 54. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data
- 55. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service
- 56. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service handles failover, monitoring
- 57. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service handles failover, monitoring holds tablet data, logs
- 58. BigTable System Structure Bigtable Cell Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service holds metadata, handles failover, monitoring holds tablet data, logs handles master-election
- 59. BigTable System Structure Bigtable client Bigtable Cell Bigtable client library Bigtable master performs metadata ops + load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service holds metadata, handles failover, monitoring holds tablet data, logs handles master-election
- 60. BigTable System Structure Bigtable client Bigtable Cell Bigtable client library Bigtable master performs metadata ops + Open() load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service holds metadata, handles failover, monitoring holds tablet data, logs handles master-election
- 61. BigTable System Structure Bigtable client Bigtable Cell Bigtable client library Bigtable master performs metadata ops + Open() read/write load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service holds metadata, handles failover, monitoring holds tablet data, logs handles master-election
- 62. BigTable System Structure Bigtable client Bigtable Cell Bigtable client metadata ops library Bigtable master performs metadata ops + Open() read/write load balancing Bigtable tablet server Bigtable tablet server … Bigtable tablet server serves data serves data serves data Cluster scheduling system GFS Lock service holds metadata, handles failover, monitoring holds tablet data, logs handles master-election
- 63. BigTable Status • Design/initial implementation started beginning of 2004 • Production use or active development for 100+ projects: – Google Print – My Search History – Orkut – Crawling/indexing pipeline – Google Maps/Google Earth – Blogger –… • Currently ~500 BigTable clusters • Largest cluster: – 70+ PB data; sustained: 10M ops/sec; 30+ GB/s I/O
- 64. BigTable: What’s New Since OSDI’06? • Lots of work on scaling • Service clusters, managed by dedicated team • Improved performance isolation – fair-share scheduler within each server, better accounting of memory used per user (caches, etc.) – can partition servers within a cluster for different users or tables • Improved protection against corruption – many small changes – e.g. immediately read results of every compaction, compare with CRC. • Catches ~1 corruption/5.4 PB of data compacted
- 65. BigTable Replication (New Since OSDI’06) • Configured on a per-table basis • Typically used to replicate data to multiple bigtable clusters in different data centers • Eventual consistency model: writes to table in one cluster eventually appear in all configured replicas • Nearly all user-facing production uses of BigTable use replication
- 66. BigTable Coprocessors (New Since OSDI’06) • Arbitrary code that runs run next to each tablet in table – as tablets split and move, coprocessor code automatically splits/moves too • High-level call interface for clients – Unlike RPC, calls addressed to rows or ranges of rows • coprocessor client library resolves to actual locations – Calls across multiple rows automatically split into multiple parallelized RPCs • Very flexible model for building distributed services – automatic scaling, load balancing, request routing for apps
- 67. Example Coprocessor Uses • Scalable metadata management for Colossus (next gen GFS-like file system) • Distributed language model serving for machine translation system • Distributed query processing for full-text indexing support • Regular expression search support for code repository
- 68. Current Work: Spanner • Storage & computation system that spans all our datacenters – single global namespace • Names are independent of location(s) of data • Similarities to Bigtable: tables, families, locality groups, coprocessors, ... • Differences: hierarchical directories instead of rows, fine-grained replication • Fine-grained ACLs, replication configuration at the per-directory level – support mix of strong and weak consistency across datacenters • Strong consistency implemented with Paxos across tablet replicas • Full support for distributed transactions across directories/machines – much more automated operation • system automatically moves and adds replicas of data and computation based on constraints and usage patterns • automated allocation of resources across entire fleet of machines
- 69. Design Goals for Spanner • Future scale: ~106 to 107 machines, ~1013 directories, ~1018 bytes of storage, spread at 100s to 1000s of locations around the world, ~109 client machines – zones of semi-autonomous control – consistency after disconnected operation – users specify high-level desires: “99%ile latency for accessing this data should be <50ms” “Store this data on at least 2 disks in EU, 2 in U.S. & 1 in Asia”
- 70. Adaptivity in World-Wide Systems • Challenge: automatic, dynamic world-wide placement of data & computation to minimize latency and/or cost, given constraints on: – bandwidth – packet loss – power – resource usage – failure modes – ... • Users specify high-level desires: “99%ile latency for accessing this data should be <50ms” “Store this data on at least 2 disks in EU, 2 in U.S. & 1 in Asia”
- 71. Building Applications on top of Weakly Consistent Storage Systems • Many applications need state replicated across a wide area – For reliability and availability • Two main choices: – consistent operations (e.g. use Paxos) • often imposes additional latency for common case – inconsistent operations • better performance/availability, but apps harder to write and reason about in this model • Many apps need to use a mix of both of these: – e.g. Gmail: marking a message as read is asynchronous, sending a message is a heavier-weight consistent operation
- 72. Building Applications on top of Weakly Consistent Storage Systems • Challenge: General model of consistency choices, explained and codified – ideally would have one or more “knobs” controlling performance vs. consistency – “knob” would provide easy-to-understand tradeoffs • Challenge: Easy-to-use abstractions for resolving conflicting updates to multiple versions of a piece of state – Useful for reconciling client state with servers after disconnected operation – Also useful for reconciling replicated state in different data centers after repairing a network partition
- 73. Thanks! Questions...? Further reading: • Ghemawat, Gobioff, & Leung. Google File System, SOSP 2003. • Barroso, Dean, & Hölzle . Web Search for a Planet: The Google Cluster Architecture, IEEE Micro, 2003. • Dean & Ghemawat. MapReduce: Simplified Data Processing on Large Clusters, OSDI 2004. • Chang, Dean, Ghemawat, Hsieh, Wallach, Burrows, Chandra, Fikes, & Gruber. Bigtable: A Distributed Storage System for Structured Data, OSDI 2006. • Burrows. The Chubby Lock Service for Loosely-Coupled Distributed Systems. OSDI 2006. • Pinheiro, Weber, & Barroso. Failure Trends in a Large Disk Drive Population. FAST 2007. • Brants, Popat, Xu, Och, & Dean. Large Language Models in Machine Translation, EMNLP 2007. • Barroso & Hölzle. The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines, Morgan & Claypool Synthesis Series on Computer Architecture, 2009. • Malewicz et al. Pregel: A System for Large-Scale Graph Processing. PODC, 2009. • Schroeder, Pinheiro, & Weber. DRAM Errors in the Wild: A Large-Scale Field Study. SEGMETRICS’09. • Protocol Buffers. http://code.google.com/p/protobuf/ These and many more available at: http://labs.google.com/papers.html