• Like

A Case Study of NoSQL Adoption: What Drove Wordnik Non-Relational?

Uploaded on


More in: Technology
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Be the first to comment
No Downloads


Total Views
On Slideshare
From Embeds
Number of Embeds



Embeds 0

No embeds

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

    No notes for slide


  • 1. Why Wordnik went Non-Relational Tony Tam @fehguy
  • 2. What this Talk is About• 5 Key reasons why Wordnik migrated into a Non-Relational database• Process for selection, migration• Optimizations and tips from living survivors of the battle field
  • 3. Why Should You Care?• MongoDB user for 2 years• Lessons learned, analysis, benefits from process• We migrated from MySQL to MongoDB with no downtime• We have interesting/challenging data needs, likely relevant to you
  • 4. More on Wordnik• World’s fastest updating English dictionary • Based on input of text up to 8k words/second • Word Graph as basis to our analysis • Synchronous & asynchronous processing• 10’s of Billions of documents in NR storage• 20M daily REST API calls, billions served • Powered by Swagger OSS API framework Powered API swagger.wordnik.com
  • 5. Architectural History• 2008: Wordnik was born as a LAMP AWS EC2 stack• 2009: Introduced public REST API, powered wordnik.com, partner APIs• 2009: drank NoSQL cool-aid• 2010: Scala• 2011: Micro SOA
  • 6. Non-relational by Necessity• Moved to NR because of ―4S‖ • Speed • Stability • Scaling • Simplicity• But… • MySQL can go a LONG way • Takes right team, right reasons (+ patience) • NR offerings simply too compelling to focus on scaling MySQL
  • 7. Wordnik’s 5 Whys for NoSQL
  • 8. Why #1: Speed bumps with MySQL• Inserting data fast (50k recs/second) caused MySQL mayhem • Maintaining indexes largely to blame • Operations for consistency unnecessary but "cannot be turned off‖• Devised twisted schemes to avoid client blocking • Aka the ―master/slave tango‖
  • 9. Why #2: Retrieval Complexity• Objects typically mapped to tables • Object Hierarchy always => inner + outer joins• Lots of static data, so why join? • “Noun” is not getting renamed in my code’s lifetime! • Logic like this is probably in application logic• Since storage is cheap • I’ll choose speed
  • 10. Why #2: Retrieval Complexity One definition = 10+ joins 50 requests per second!
  • 11. Why #2: Retrieval Complexity• Embed objects in rows ―sort of works‖ • Filtering gets really nasty • Native XML in MySQL? • If a full table-scan is OK…• OK, then cache it! • Layers of caching introduced layers of complexity • Stale data/corruption • Object versionitis • Cache stampedes
  • 12. Why #3: Object Modeling• Object models being compromised for sake of persistence • This is backwards! • Extra abstraction for the wrong reason• OK, then performance suffers • In-application joins across objects • ―Who ran the fetch all query against production?!‖ –any sysadmin• ―My zillionth ORM layer that only I understand‖ (and can maintain)
  • 13. Why #4: Scaling• Needed "cloud friendly storage" • Easy up, easy down! • Startup: Sync your data, and announce to clients when ready for business • Shutdown: Announce your departure and leave• Adding MySQL instances was a dance • Snapshot + bin files mysql> change master to MASTER_HOST=db1, MASTER_USER=xxx, MASTER_ PASSWORD=xxx, MASTER_LOG_FILE=master- relay.000431, MASTER_LOG_POS=1035435402;
  • 14. Why #4: Scaling• What about those VMs? • So convenient! But… they kind of suck • Can the database succeed on a VM?• VM Performance: • Memory, CPU or I/O—Pick only one • Can your database really reduce CPU or disk I/O with lots of RAM?
  • 15. Why #5: Big Picture• BI tools use relational constraints for discovery • Is this the right reason for them? • Can we work around this? • Let’s have a BI tool revolution, too!• True service architecture makes relational constraints impractical/impossible• Distributed sharding makes relational constraints impractical/impossible
  • 16. Why #5: Big Picture• Is your app smarter than your database? • The logic line is probably blurry!• What does count(*) really mean when you add 5k records/sec? • Maybe eventual consistency is not so bad…• 2PC? Do some reading and decide!http://eaipatterns.com/docs/IEEE_Software_Design_2PC.pdf
  • 17. Ok, I’m in!• I thought deciding was easy!? • Many quickly maturing products • Divergent features tackle different needs• Wordnik spent 8 weeks researching and testing NoSQL solutions • This is a long time! (for a startup) • Wrote ODM classes and migrated our data• Surprise! There were surprises • Be prepared to compromise
  • 18. Choice Made, Now What?• We went with MongoDB *** • Fastest to implement • Most reliable • Best community• Why? • Why #1: Fast loading/retrieval • Why #2: Fast ODM (50 tps => 1000 tps!) • Why #3: Document Models === Object models • Why #4: MMF => Kernel-managed memory + RS • Why #5: It’s 2011, is there no progress?
  • 19. More on Why MongoDB• Testing, testing, testing • Used our migration tools to load test • Read from MySQL, write to MongoDB • We loaded 5+ billion documents, many times over• In the end, one server could… • Insert 100k records/sec sustained • Read 250k records/sec sustained • Support concurrent loading/reading
  • 20. Migration & Testing• Iterated ODM mapping multiple times • Some issues • Type Safety cur.next.get("iWasAnIntOnce").asInstanceOf[Long] • Dates as Strings obj.put("a_date", "2011-12-31") != obj.put("a_date", new Date("2011-12-31")) • Storage Size obj.put("very_long_field_name", true) >> obj.put("vsfn", true)
  • 21. Migration & Testing• Expect data model iterations • Wordnik migrated table to Mongo collection "as-is‖ • Easier to migrate, test • _id field used same MySQL PK • Auto Increment? • Used MySQL to ―check-out‖ sequences • One row per mongo collection • Run out of sequences => get more • Need exclusive locks here!
  • 22. Migration & Testing• Sequence generator in-process SequenceGenerator.checkout("doc_metadata,100")• Sequence generator as web service • Centralized UID management
  • 23. Migration & Testing• Expect data access pattern iterations • So much more flexibility! • Reach into objects > db.dictionary_entry.find({"hdr.sr":"cmu"}) • Access to a whole object tree at query time • Overwrite a whole object at once… when desired • Not always! This clobbers the whole record > db.foo.save({_id:18727353,foo:"bar"}) • Update a single field: > db.foo.update({_id:18727353},{$set:{foo:"bar"}})
  • 24. Flip the Switch• Migrate production with zero downtime • We temporarily halted loading data • Added a switch to flip between MySQL/MongoDB • Instrument, monitor, flip it, analyze, flip back• Profiling your code is key • What is slow? • Build this in your app from day 1
  • 25. Flip the Switch
  • 26. Flip the Switch• Storage selected at runtime val h = shouldUseMongoDb match { case true => new MongoDbSentenceDAO case _ => new MySQLDbSentenceDAO } h.find(...)• Hot-swappable storage via configuration • It worked!
  • 27. Then What?• Watch our deployment, many iterations to mapping layer • Settled on in-house, type-safe mapper https://github.com/fehguy/mongodb-benchmark-tools• Some gotchas (of course) • Locking issues on long-running updates (more in a minute)• We want more of this! • Migrated shared files to Mongo GridFS • Easy-IT
  • 28. Performance + Optimization• Loading data is fast! • Fixed collection padding, similarly-sized records • Tail of collection is always in memory • Append faster than MySQL in every case tested• But... random access started getting slow • Indexes in RAM? Yes • Data in RAM? No, > 2TB per server • Limited by disk I/O /seek performance • EC2 + EBS for storage?
  • 29. Performance + Optimization• Moved to physical data center • DAS & 72GB RAM => great uncached performance• Good move? Depends on use case • If ―access anything anytime‖, not many options • You want to support this?
  • 30. Performance + Optimization• Inserts are fast, how about updates? • Well… update => find object, update it, save • Lock acquired at ―find‖, released after ―save‖ • If hitting disk, lock time could be large• Easy answer, pre-fetch on update • Oh, and NEVER do ―update all records‖ against a large collection
  • 31. Performance + Optimization• Indexes • Cant always keep index in ram. MMF "does its thing" • Right-balanced b-tree keeps necessary index hot • Indexes hit disk => mute your pager 1 7 1 2 5 7
  • 32. More Mongo, Please! • We modeled our word graph in mongo• 50M Nodes• 80M Edges• 80 S edge fetch
  • 33. More Mongo, Please!• Analytics rolled-up from aggregation jobs • Send to Hadoop, load to mongo for fast access
  • 34. What’s next• Liberate our models • stop worrying about how to store them (for the most part)• New features almost always NR• Some MySQL left • Less on each release
  • 35. Questions?• See more about Wordnik APIs http://developer.wordnik.com• Migrating from MySQL to MongoDBhttp://www.slideshare.net/fehguy/migrating-from-mysql-to-mongodb-at-wordnik• Maintaining your MongoDB Installation http://www.slideshare.net/fehguy/mongo-sv-tony-tam• Swagger API Framework http://swagger.wordnik.com• Mapping Benchmark https://github.com/fehguy/mongodb-benchmark-tools• Wordnik OSS Tools https://github.com/wordnik/wordnik-oss