Augmenting Mongo DB with Treasure Data
Download as PPTX, PDF1 like1,161 views
The document discusses the integration of MongoDB with Treasure Data, highlighting MongoDB's strengths in format and convenience while acknowledging its weaknesses, particularly in integration capabilities and management of large clusters. It explores various use cases, including those for e-commerce, and suggests operational improvements, emphasizing the challenges of data management and analytics. Ultimately, it compares the trade-offs between using MongoDB and Treasure Data, emphasizing ease of use and scalability versus control over outcomes.
Augmenting Mongo DB with Treasure Data
- 2. Augmenting MongoDB with Treasure Data
- 3. About Me • A recovering software engineer turned digital artist once interested in fractals; • now into data visualization based on large datasets rendered directly to GPU (RGL, various Python GL libraries, etc.) • it’s easier these days to manipulate large dataset with limited effort
- 4. Images courtesy of Edureka, 10gen, MongoDB, clipart panda and aperfectworld.org
- 6. relatively speaking Images courtesy of Edureka and MongoDB
- 7. Strengths: Format, Convenience Schema courtesy of Emily Stolfo, MongoDB
- 8. Discuss
- 9. “not so strengths” of MongoDB • Horowitz was also very honest about where and how MongoDB is lacking in its current offering – most notably in terms of integration capabilities and some areas of high performance. • “In the relational world you’ve got a few big boxes, in the MongoDB world you could have 2,000 commodity servers, so you need really great management tools for that. That’s a huge problem for us.” • “The other big thing is automation, where you can have automation tools that let you manage very large clusters all from a very simple pane of glass.” http://diginomica.com/2014/11/10/mongodb-cto-mongo-works-doesnt/ From an interview with MongoDB CTO Eliot Horowitz
- 10. testimony of John R Jensen, Cengage
- 11. hmmm…moar “not so strengths” ;) of MongoDB • The dreaded “Write Lock” • https://news.ycombinator.com/item?id=1691748 • http://www.sarahmei.com/blog/2013/11/11/why-you-should-never-use-mongodb/ - is the data actually relational or not? • Slideshare “where not to use MongoDB” • Slideshare “Hive vs. Cassandra vs. MongoDB” • http://www.slideshare.net/johnrjenson/mongodb-pros-and-cons • “choosing the right NoSQL database” video: https://www.youtube.com/watch?t=34&v=gJFG04Sy6NY • limits of MongoDB: http://docs.mongodb.org/manual/reference/limits/
- 14. What are the steps to scale?
- 15. The Hadoop Story on MongoDB
- 16. Share some use-cases with us, won’t you?
- 17. Complementing MongoDB • Featurewise? • Elastic Search • Use-Case: Stripe • Opslog feature as queryable log of diffs • Hadoop (good for large scale processing), Hive (Treasure Data) • Use-Case: Wish
- 18. What’s wish.com? • Mobile eCommerce - world’s largest mobile shopping mall • Top 10 app on iOS & Android • Product discovery/personalization
- 19. Your Fun, Personalized Shopping Mall
- 20. Wish & MongoDB • Primary database • 64x mongod • AWS -> bare metal (SSDs ftw!)
- 21. Architecture App server Wish App fluentd fluentd aggregation server fluentd fluentd aggregation server fluentd Hadoop/Hive
- 22. Before fluentd
- 23. fluentd!
- 24. Complementing MongoDB • Operationally? • Managing MongoDB is hard (spin up instances with Mongolab) • MongoDB (monitoring products: Ops Manager and MMS) • What’s your pain? What’s are you missing?
- 25. You know what? Managing ANY data is hard.
- 26. Treasure Data as a Cloud Way of complementing Mongo DB
- 27. 6 Batch Imports Streaming Imports Ingestion Batch Processing (Hive, Presto) Storage (Schema-on-read) Console ODBC JDBC AWS S3 PrestogresCLI Luigi-TD FTP GDocs HTTP Put Python PostgreSQL Redshift SFDC Yahoo BDI Tableau Server MySQL … Ingest Analyze Distribute
- 28. Gree (gaming)
- 29. DATA ACCESS > ADVANCED ANALYTICS • Product analytics: data access is a major issue. • “Machine learning” is still simple and “small” in scale (can be done inside Python) • Future work: productized/operationalized machine learning bluetooth iOS/Android SDK Fluentd Python/Pandas SDK Data Science Team (5 people)
- 30. SCHEMA(LESS) COUNTS • Redshift: lots of co-use cases • Event data is semi- structured → Can be modeled as JSON but schemas change • Treasure Data provides a SQL-accessible, semi- structured data lake. email Source of truth/JSON More intensive data processing Hourly/Daily load Big data mart More interactive data processing Ad hoc queries for new data Ad hoc queries for cached data
- 31. DATA COLLECTION IS HARD • Want to assume all data is on S3 or HDFS, but reality is murkier. • Sensor readings available as email attachments • Provide data collection tools for 90% of the use cases. Have APIs ready for 10%. GH SCADA email Parse & transform Import via REST Import data as JSON Analyze via SQL Query Results Data-informed maintenance
- 32. The Sunk Cost Fallacy
- 33. Some revised scenarios • Revised scenario 1: Using Treasure Data for Ingestion and analytics; exporting results to MongoDB for reporting
- 34. Some revised scenarios • Revised scenario 2: Ingestion data into MongoDB and exporting to Treasure Data
- 35. Treasure Data is good for a some of the same things… • less overhead in setup • less - make that practically no - effort to scale • less overhead/effort to use • but -> less fine-tuned control over outcome
- 36. The Tradeoff Control (sharding and replication) vs. Ease (scaling automatically in the cloud) ?
- 37. MongoDB is already excellent for a lot of things +
- 38. Demo!
Editor's Notes
- #5: Just a quick review… Sharding strategies: Range sharding - (shard key divided by e.g. device id by range) Hash Sharding: MongoDB applies a MD5 hash on the key when the subkey is used Tag-Aware Sharding: allows a subset of shards to be tagged, and assigned to a sub-range of the shard key
- #6: The folks at Edureka did a comparative study of different database types. Breakoff discussion: Let’s talk about what kind of databases we’re using, and for what purposes.
- #7: Question to audience: How do mongo and HBase (Plazma?) fall on the boundary between partition tolerance and consistency? (Might consider leaving this slide out
- #8: BSON looks like JSON and translate nicely to things like python dictionaries. Working the Mongo prompt is easy but requires mastering another API/paradigm.
- #9: What are some other strengths of MongoDB
- #10: Managing MongoDB is hard (spin up instances with Mongolab) MongoDB (monitoring products: Ops Manager and MMS)
- #12: If you can lose 5sec. worth of updates, a MongoDB replication pair is just fine. If you can lose a day's worth of updates (or can easily reconstruct the database contents from other sources), you can try out pretty much anything without bad repercussions. If you can't lose anything, you're pretty much limited to the most conservative databases (the SQL bunch).
- #16: Any places where this process could be problematic? One is a failure before cache is written, during finalize. Another could be a failure during any step of the M-R process.
- #22: NOTE: Need transitions on this slide to control how things appear with my story We start in the app which generates the logs. The app synchronously logs to a fluentd running on the same host. There’s no network latency and the load on each local fluentd is trivial, so we’ve never had problems with these getting slow The local fluentd accepts the logs and buffers them on disk for reliability. Periodically, it flushes those buffers to one of the hosts in our fluentd aggregation tier with at-least-once semantics. These run active/active and can be scaled out linearly. They also buffer on disk and periodically flush into Hadoop. As an added bonus, they also flush into S3 for backup. This tier gives us an easy to monitor & manage conduit for our logs to flow through without imposing extra costs on the app. To recap, the logs from our app are buffered by a fluentd on the same host. That reliably forwards to a tier of aggregation fluentds, which forward to Hadoop and S3.
- #33: The sunk cost fallacy is the idea that your sunk costs (unrecoverable) create barriers to adjusting your future spending. For example, “I’m hungry. Therefore I should eat that egg salad in the fridge (even if it’s gone bad) because I’ve already spent the money on it (rather than going for fresh food.”