Challenges with MongoDB - GAO, Stone. Tech Lead of UMeng.com

MongoDB Shanghai 2012

weibo.com/stonegao

Transcript

1. Challenges with MongoDB MongoDB Beijing 2012 Stone Gao Monday, April 2, 2012
2. About Me Tech Lead at Umeng.com Monday, April 2, 2012
3. MongoDB is Awesome • Document-oriented storage • Full Index Support • Replication & High Availability • Auto-Sharding • Querying • Fast In-Place Updates • Map/Reduce • GridFS Monday, April 2, 2012
4. But... This talk is not Yet Another Talk about it’s Awesomeness but challenges with MongoDB Monday, April 2, 2012
5. Outline 1. Global Write Lock Sucks 2. Auto-Sharding is not that Reliable 3. Schema-less is Over Rated 4. Community Contribution is Quite Low 5. Attitude Matters Monday, April 2, 2012
6. 1. Global Write Lock Sucks http://www.clker.com/cliparts/3/3/5/D/X/b/locked-exclamation-mark-padlock-hi.png Monday, April 2, 2012
7. 1. Global Write Lock Sucks mongod db-n doc2 doc1 collection1 doc2 doc1 collection2 db-1 doc2 doc1 collection1 doc2 doc1 collection2 mysqld db-n doc2 doc1 table1 doc2 doc1 table2 db-1 doc2 doc1 table1 doc2 doc1 table2 VS. DB Process Lock VS. Row Lock single global write lock for the entire server (process) Monday, April 2, 2012
8. 1. Global Write Lock Sucks Intel SSD 320 RAID10 & mongostat Nearly all data in RAM, lock ratio is pretty high and bunch of Queued Writes(qw) 39.5K Rread IOPS / 23K Write IOPS Monday, April 2, 2012
9. 1. Global Write Lock Sucks Intel SSD 320 RAID10 & mongostat Nearly all data in RAM, lock ratio is pretty high and bunch of Queued Writes(qw) 39.5K Rread IOPS / 23K Write IOPS Monday, April 2, 2012
10. 1. Global Write Lock Sucks Intel SSD 320 RAID10 & mongostat Nearly all data in RAM, lock ratio is pretty high and bunch of Queued Writes(qw) 39.5K Rread IOPS / 23K Write IOPS Monday, April 2, 2012
11. 1. Global Write Lock Sucks Intel SSD 320 RAID10 & mongostat Nearly all data in RAM, lock ratio is pretty high and bunch of Queued Writes(qw) 39.5K Rread IOPS / 23K Write IOPS Monday, April 2, 2012
12. Possible Solutions/Workarounds #1 Wait for lock related issues on JIRA •SERVER-1240 : Collection level locking https://jira.mongodb.org/browse/SERVER-1240 Planning Bucket A Vote (154) •SERVER-1241 : Intra collection locking (maybe extent) https://jira.mongodb.org/browse/SERVER-1241 Planning Bucket A Vote (25) •SERVER-2563 : When hitting disk, yield lock - phase 1 https://jira.mongodb.org/browse/SERVER-2563 Fixed in 1.9.1 Vote (25) • any time we actually have to hit disk. so if a memory mapped page is not in ram, then we should yield update by _id, remove, long cursor iteration •SERVER-1169 : Record level locking https://jira.mongodb.org/browse/SERVER-1169 Rejected Vote (1) and more ... Monday, April 2, 2012
13. Possible Solutions/Workarounds #2 One Collection per DB to Reduce Lock Ratio But you can go no further Use Auto-Sharding to the rescue ? Monday, April 2, 2012
14. 2. Auto-Sharding is not that Reliable http://www.autoinsurancecompanies.com/wp-content/uploads/2011/11/reliable.jpg Monday, April 2, 2012
15. Auto-Sharding is not that Reliable Monday, April 2, 2012
16. Problems with Auto-Sharding • MongoDB can’t figure out how many docs in a collection after sharding • Balancer dead lock [Balancer] skipping balancing round during ongoing split or move activity.) [Balancer] dist_lock lock failed because taken by.... [Balancer] Assertion failure cm s/balance.cpp... • Uneven shard load distribution • ... (Note: I did the experiment before 2.0. So some of the issues might be fixed or improved in new versions of MongoDB coz it’s evolving very fast) Monday, April 2, 2012
17. 0) Turn off the balancer (balancing won't understand your locations, but it shouldn't matter b/c you're using hashed shard keys) 1) Shard the empty collection over the shard key { location : 1, hash : 1 } 2) run db.runCommand({ split : "<coll>", middle : { "location":"DEN", "hash": "8000...0" }}) 3) run db.runCommand({ split : "<coll>", middle : { "location":"SC", "hash": "0000...0" }}) 4) move those empty chunks to whatever shards you want - Greg Studer Possible Solutions/Workarounds #1 Manual Chunk Pre-Splitting http://www.mongodb.org/display/DOCS/Splitting+Shard+Chunks https://groups.google.com/d/msg/mongodb-user/tYBFKSMM3cU/TiYtoOiNMgEJ http://blog.zawodny.com/2011/03/06/mongodb-pre-splitting-for-faster-data-loading-and-importing/ Monday, April 2, 2012
18. Possible Solutions/Workarounds #2 https://github.com/twitter/gizzard/raw/master/doc/forwarding_table.png SERVER-2001 : Option to hash shard key https://jira.mongodb.org/browse/SERVER-2001 Unresolved Fix Version/s: 2.1.1 Vote (27) “The lack of hashing based read/write distribution amongst available shards is a huge issue for us now. We're actually considering implementing an app-side layer to do this but that obviously has a number of serious drawbacks.” - Remon van Vliet “Seems like a good idea : we implemented hashed shard key on client-side : operation rate sky rocked ( x3 and less variability). Balancing is moreover quicker and done during our very heavy insertion process : perfect !” - Grégoire Seux Monday, April 2, 2012
19. Possible Solutions/Workarounds #3 Plain-old Application Level Sharding https://github.com/twitter/gizzard/raw/master/doc/forwarding_table.png Monday, April 2, 2012
20. 3. Schema-less is Over Rated http://images.sodahead.com/polls/001635729/1863780_overrated_answer_2_xlarge.jpeg Monday, April 2, 2012
21. Schema-less is Over Rated Schema-Free (schema-less) is not free. It means repeat the schema in every docs (records) ! Monday, April 2, 2012
22. Possible Solutions/Workarounds #1 Use Short Key Names ref : http://christophermaier.name/blog/2011/05/22/MongoDB-key-names {"sequence":"AHAHSPGPGSAVKLPAPHSVGKSALR", "location":{ "chromosome":"19", "strand":"-", "begin":"51067007", "end":"51067085" }} {"s":"AHAHSPGPGSAVKLPAPHSVGKSALR", "l":{ "c":"19", "s":"-", "b":"51067007", "e":"51067085" }} 1.6 billion documents 243 GB 183 GB 60 GB saved! Monday, April 2, 2012
23. Possible Solutions/Workarounds #2 SERVER-863 : Tokenize the field names https://jira.mongodb.org/browse/SERVER-863 planned but not scheduled Vote (66) “Most collections, even if they don’t contain the same structure , they contain similar. So it would make a lot of sense and save a lot of space to tokenize the field names.” “The overall benefit as mentioned by other users is that you reduce the amount of storage/RAM taken up by redundant data in each document (so you can use less resources per request, hence gain more throughput and capacity), while importantly also freeing the developer from having to pick short and hard to read field names as a workaround for a technical limitation.” - Andrew Armstrong Monday, April 2, 2012
24. Possible Solutions/Workarounds #3 SERVER-164 : Option to store data compressed https://jira.mongodb.org/browse/SERVER-164 planned but not scheduled Vote (126) “The way oracle handles this is transparent to the database server at the block engine level. They compress the blocks similar to how SAN store's handle it rather than at a record level. They use zlib type compression and the overhead is less than 5 percent. Due to the IO access reduction in both number of blocks touched, and amount of data transferred, the overall effect is a cumulative speed increase. Should MongoDB do it this way? Maybe? But at the end of the day, the architecture must make Mongo more scalable, as well as increase the ability limit the storage footprint.” - Michael D. Joy Monday, April 2, 2012
25. 4. Community Contribution is Quite Low http://www.thompsoncrg.com/wp-content/themes/zoomtechnic/images/slide/img3.jpg Monday, April 2, 2012
26. Community Contribution is Quite Low https://github.com/mongodb/mongo/graphs/impact https://github.com/mongodb/mongo/contributors Monday, April 2, 2012
27. 5. Attitude Matters Monday, April 2, 2012
28. 5. Attitude Matters MongoDB already has the sweetest API in the NoSQL world. Wish more effort invested in fixing the Hard Problems : locking, sharding, storage engine... http://www.mongodb.org/display/DOCS/SQL+to+Mongo+Mapping+Chart Monday, April 2, 2012
29. We are hiring • Backend Engineer (MongoDB, Hadoop, HBase, Storm, Scala, Java, Ruby, Clojure) • Data Mining Engineer • DevOps Engineer • Front End Engineer We are doing bigdata analytics hr@umeng.com Monday, April 2, 2012
30. Contact • Email : stones.gao@gmail.com gaolei@umeng.com • Twitter: @stonegao Monday, April 2, 2012
31. Thanks Q & A Monday, April 2, 2012