More Related Content
What's hot
What's hot (20)
Similar to Making Session Stores More Intelligent
Similar to Making Session Stores More Intelligent (20)
Recently uploaded
Recently uploaded (20)
Making Session Stores More Intelligent
- 1. Making Session Stores More Intelligent KYLE J. DAVIS TECHNICAL MARKETING MANAGER REDIS LABS
- 2. What is a session store?
- 3. • An chunk of data that is connected to one “user” of a service –”user” can be a simple visitor –or proper user with an account • Often persisted between client and server by a token in a cookie* –Cookie is given by server, stored by browser –Client sends that cookie back to the server on subsequent requests –Server associates that token with data • Often the most frequently used data by that user –Data that is specific to the user –Data that is required for rendering or common use • Often ephemeral and duplicated A session store is…
- 4. Session Storage Uses Cases Traditional • Username • Preferences • Name • “Stateful” data Intelligent • Traditional + • Notifications • Past behaviour –content surfacing –analytical information –personalization
- 5. In a simple world Internet Server Database
- 6. Good problems Internet Server Database Traffic Grows… Struggles
- 7. Good solution Internet Server Database performance restored Session storage on the server
- 8. More good problems Internet Server Database Session storage on the server Struggling
- 9. Problematic Solutions Internet Server Database Session storage on the server Load balanced Session storage on the server
- 10. Multiple Servers + On-server Sessions? Server DatabaseRobin Server #1 – Hello Robin!
- 11. Multiple Servers + On-server Sessions? Server DatabaseRobin Server #3 – Hello ????
- 12. Better solution Internet Server Database Load balanced Redis Session Storage
- 13. What is Redis?
- 14. Who We Are Open source. The leading in-memory database platform, supporting any high performance operational, analytics or hybrid use case. The open source home and commercial provider of Redis Enterprise technology, platform, products & services. 14
- 15. Redis Top Differentiators Simplicity ExtensibilityPerformance NoSQL Benchmark 1 Redis Data Structures 2 3 Redis Modules 15 Lists Hashes Bitmaps Strings Bit field Streams Hyperloglog Sorted Sets Sets Geospatial Indexes
- 16. Performance: The Most Powerful Database Highest Throughput at Lowest Latency in High Volume of Writes Scenario Least Servers Needed to Deliver 1 Million Writes/Sec Benchmarks performed by Avalon Consulting Group Benchmarks published in the Google blog 16 1 Serversusedtoachieve1Mwrites/sec
- 17. Simplicity: Data Structures - Redis’ Building Blocks Lists [ A → B → C → D → E ] Hashes { A: “foo”, B: “bar”, C: “baz” } Bitmaps 0011010101100111001010 Strings "I'm a Plain Text String!�� Bit field {23334}{112345569}{766538} Key 17 2 ”Retrieve the e-mail address of the user with the highest bid in an auction that started on July 24th at 11:00pm PST” ZREVRANGE 07242015_2300 0 0= Streams {id1=time1.seq1(A:“xyz”, B:“cdf”), d2=time2.seq2(D:“abc”, )} Hyperloglog 00110101 11001110 Sorted Sets { A: 0.1, B: 0.3, C: 100 } Sets { A , B , C , D , E } Geospatial Indexes { A: (51.5, 0.12), B: (32.1, 34.7) }
- 18. • Add-ons that use a Redis API to seamlessly support additional use cases and data structures. • Enjoy Redis’ simplicity, super high performance, infinite scalability and high availability. Extensibility: Modules Extend Redis Infinitely • Any C/C++/Go program can become a Module and run on Redis. • Leverage existing data structures or introduce new ones. • Can be used by anyone; Redis Enterprise Modules are tested and certified by Redis Labs. • Turn Redis into a Multi-Model database 18 3
- 19. Redise Pack Managed Fully managed Redise Pack in private datacenters Redise Pack Downloadable Redise software for any enterprise datacenter or cloud environment Redise Cloud Private Fully managed, server- less scaling Redise service in VPCs within AWS, MS Azure, GCP and IBM Softlayer Redise Cloud Fully managed, server- less Redise service on hosted resources within AWS, MS Azure, GCP, IBM Softlayer, Heroku, CF and OpenShift Redis Labs Products 19 or or or DBaaS Software
- 20. Concepts
- 21. • Probabilistic data structure • Hash -> sample bits -> set bits • Properties: –False negatives – not possible –False positives – possible, but controllable –Bits per item stored –Add or check if exists –Like the Tardis, it’s bigger on the inside than outside • Availability: –Redis Module –On top of bitfields Concept: Bloom Filters (presence)
- 22. • Probabilistic data structure • Hash -> count runs -> store runs • Properties: –Estimates unique items –Bits per item stored – 264 unique items in 12kb / error rate 0.81% –Add, count or merge! –Like the Tardis, it’s bigger on the inside than outside • Availability: –All versions of Redis Concept: HyperLogLog (cardinality) engineering.conversantmedia.com
- 23. • It’s just bits! • Fixed starting point, each point represents a moment in time, flip to represent activity • Properties: –Size relative to length of time (byte round) –Count totals or ranges –BITOP (AND/XOR/OR/NOT) • Availability: –All versions of Redis Concept: Bit counting (time series)
- 25. Process • Group of users get notification “Sale on sweaters” • Insert into central table of notifications • Insert row in table with each user of group with notification and seen flag • Each time it is needed, query notifications table where seen flag is false. Traditional Group Notification Pattern
- 26. Traditional Group Notification Pattern Challenges • Adding/removing means touching a row for each user in group. –Fine for groups of 10 users, what about 1 million? –Also multi-step • Storage is proportional to size of group and notifications • Constant DB hits, not easily cacheable • Setting “read” is DB write
- 27. Process • Add notification to single group based structure or table (easily cacheable) • First n notifications are read by all users in group. • The notifications are checked to see if they are in a session-based Bloom filter or not. • Mark read by adding to Bloom filter in session store. Modern & Intelligent Group Notification Pattern
- 28. Modern & Intelligent Group Notification Pattern Advantages • Adding a notification only writes to a single table, single row. • Model fits use – unread assumed. • Fast. Checking for read / writing read is unrelated to number of items in the filter. Consistent. • ~5-bits per item, but Bloom filter doesn’t always grow. • Gentle scaling
- 29. Visual Notification #1 Notification #2 Notification #3 Notification #4 Notification #5 Notification #6 �� ✗ ✔ ✗ ✔ ✗ Notification #1 Notification #2 Notification #3 Notification #4 Notification #5 Notification #6 ✗
- 30. Fresh Content
- 31. Process • Hand pick and rotate a small number of content/items • Stored in DB table • Served out dumbly to users Traditional Content Surfacing Pattern (Basic) Challenges • May serve content multiple times • Freshness is linked to a manual curatorial process
- 32. Traditional Content Surfacing Pattern (Advanced) Process • Batch process builds content list to surface for each user • List is stored in DB Table • Served out to user • Rotated on a schedule Challenges • Not Real-time • May serve content multiple times • Un-cacheable DB content • Hard to scale
- 33. Process • Middleware adds each content read to a Bloom filter stored in the session • Featured content list is built, can be extensive. • Featured items are checked vs Bloom filter on-the-fly Modern & Intelligent Content Surfacing Pattern Advantages • No DB hits for user • Featured content is cacheable • Will not to show content multiple times if read • Tiny storage requirements even at scale • Freshness can be achieved with zero/low human input • Real-time recording of activity – immediate impact on fresh content
- 34. Visual Content #1 Content #2 Content #3 Content #4 Content #5 Content #6 ✔ ✗ ✔ ✗ ✔ ✗ Content #1 Content #1 ✗ Content #3 Content #5
- 35. Activity Pattern Monitoring & Personalization
- 36. • Monitor the usage behaviour –Content viewed –Activity over time –Combinations of content history and activity • Personalize the content based on the behaviour • Seen as difficult to accomplish –Analytics data • Stored in another service • Anonymized –Complicated graph or ML based solutions • Inferences • Black boxes Activity Pattern Monitoring & Personalization?
- 37. • Record site activity with bit counting • Unique page views in HyperLogLog • Leverage the page visit Bloom filter • Simpler counter for pages consumed • Create criteria based on session stored analytics –New to a page? Bloom filter –New to the site? Unique Page view = 1 (HLL) && Previously Visited = false (Bloom) –Inactive user? Sum the bit count over the last five records, if = 0 then inactive –Been to a cluster of pages (infer interest)? Check cluster of pages vs Bloom filter – combo! Activity Pattern Monitoring & Personalization
- 38. • Why is this suddenly possible? –Probabilistic data structures are small/fast –Bit counting is small/fast –Decoupled from operational database • What about privacy? –Legitimate concern –Non-reversible probabilistic structures –Siloed from rest of database Activity Pattern Monitoring & Personalization
- 39. Questions?
- 40. Thank you!