Event Sourcing, Stream Processing and Serverless (Kafka Summit SF19)

In this talk we’ll look at the relationship between three of the most disruptive software engineering paradigms: event sourcing, stream processing and serverless. We’ll debunk some of the myths around event sourcing. We’ll look at the inevitability of event-driven programming in the serverless space and we’ll see how stream processing links these two concepts together with a single ‘database for events’. As the story unfolds we’ll dive into some use cases, examine the practicalities of each approach, particularly the stateful elements, and finally extrapolate how their future relationship is likely to unfold. Key takeaways include: The different flavors of event sourcing and where their value lies. The difference between stream processing at application- and infrastructure-levels. The relationship between stream processors and serverless functions. The practical limits of storing data in Kafka and stream processors like KSQL.

Transcript

1. Event Sourcing, Stream Processing & Serverless Ben Stopford Office of the CTO, Confluent
2. What we’re going to talk about • Event Sourcing • What it is and how does it relate to Event Streaming? • Stream Processing as a kind of “Database” • What does this mean? • Serverless Functions • How do this relate?
3. Can you do event sourcing with Kafka?
4. Traditional Event Sourcing
5. Popular example: Shopping Cart DB Apps Search Apps Apps Database Table matches what the user sees.
6. 12.42 12.44 12.49 12.50 12.59 Event Sourcing stores events, then derives the ‘current state view’ Apps Apps DERIVE Chronological Reduce Event Timeseries of user activity
7. Traditional Event Sourcing (Store immutable events in a database in time order) Apps Search NoSQL Monitoring Security Apps Apps S T R E A M I N G P L AT F O R M Table of events Persist events Apps Apps
8. Traditional Event Sourcing (Read) Apps Search NoSQL Monitoring Security Apps Apps S T R E A M I N G P L AT F O R M Apps Search Monitoring Apps Apps Chronological Reduce on read (done inside the app) Query by customer Id (+session?) - No schema migration - Similar to ’schema on read’
9. 3 Benefits
10. Evidentiary Accountants don’t use erasers (e.g. audit, ledger, git)
11. Replayability Recover corrupted data after a programmatic bug
12. Analytics Keep the data needed to extract trends and behaviors i.e. non-lossy (e.g. insight, metrics, ML)
13. Traditional Event Sourcing • Use a database (any one will do) • Create a table and insert events as they occur • Query all the events associated with your problem* • Reduce them chronologically to get the current state *Aggregate ID in DDD parlance
14. Traditional Event Sourcing with Kafka • Use a database Kafka • Create a table topic insert events as they occur • Query all the events associated with your problem* • Reduce them chronologically to get the current state *Aggregate ID in DDD parlance
15. Confusion: You can’t query Kafka by say Customer Id* *Aggregate ID in DDD parlance
16. If we can’t query by Customer ID then what do we do?
17. CQRS is a tonic: Cache the projection in a ‘View’ Apps Search Monitoring Apps Apps S T R E A M I N G P L AT F O R M Query by customer Id Apps Search NoSQL Apps Apps DWH Hadoop S T R E A M I N G P L AT F O R M View Events/Command Events are the Storage Model Stream Processor Cache/DB/Ktable etc. Regenerate the view rather than doing schema migration
18. CQRS provides the benefits of event sourcing using a “Materialized View”
19. Even with CQRS, Event Sourcing is Hard CQRS helps, but it’s still quite hard if you’re a CRUD app
20. What’s the problem? Harder: • Eventually Consistent • Multi-model (Complexity ∝ #Schemas in the log) • More moving parts Apps Search NoSQL Monitoring Security Apps Apps S T R E A M I N G P L A T F O R M CRUD System CQRS
21. New York Times Website Source of Truth Every article since 1851 https://www.confluent.io/blog/publishing-apache-kafka-new-york-times/ Normalized assets (images, articles, bylines, tags all separate messages) Denormalized into “Content View”
22. If CRUD makes sense there are other ways: audit tables, CDC, etc. Trigger Evidentiary Replayable N/A to web app Analytics CDC
23. More advanced: Use a Bi-Temporal Database
24. Events make most sense where data has to move
25. This is where CQRS comes into its own!
26. Online Transaction Processing: e.g. a Flight Booking System - Flight price served 10,000 x #bookings - Consistency required only at booking time
27. CQRS with event movement Apps Search Monitoring Apps Apps S T R E A M I N G P L AT F O R M Apps Search NoSQL Apps Apps DWH Hadoop S T R E A M I N G P L AT F O R M View Book Flight Apps Search Apps S T R E A M I N G P L A View Apps Search NoSQL Apps DWH S T R E A M I N G P L A View Get Flights Get Flights Get Flights Global Read Central Write
28. The exact same logic applies to microservices
29. Event Sourcing for Microservices Basket Service Fraud Service Billing Service Email Service Basket Events
30. Event Sourcing for Microservices Basket Service Fraud Service Billing Service Email Service Basket Events Events are the storage model Each microservice creates a view that suits its use case
31. Event Sourcing “with a DB” for monoliths. Event Streaming for Microservices & Scale. (Often via. CQRS)
32. Event Streaming
33. Event Streaming is a more general form of Event Sourcing/CQRS Event Streaming • Events as shared data model • Many microservices • Polyglot persistence • Event-Driven processing Traditional Event Sourcing • Events as a storage model • Single microservice • Single DB • data-at-rest
34. Event Streams is about many event sources (Join, Filter, Transform and Summarize) Fraud Service Orders Service Payment Service Customer Service Event Log Projection created in Kafka Streams API
35. KStreams & KSQL have different positioning • KStreams is a library for Dataflow programming: • App Logic & Stream Processor (including state) are combined. • Apps are stateful. • JVM only. • KSQL is a ‘database’ for event preparation: • App sends SQL to a separate process • Apps are stateless • Connect from any language
36. This difference makes most sense if we we look to the future.
37. Cloud & Serverless
38. Thesis • Serverless provides event-driven infrastructure • KSQL is the corollary: an event-driven database
39. Serverless Functions (FaaS) • Write a function • Upload • Configure a trigger (HTTP, Messaging, Object Store, Database, Timer etc.) Request Respond Event Source
40. FaaS in a Nutshell • Fully managed (Runs in a container pool) • Pay for execution time (not resources used) • Auto-scales with load • 0-1000+ concurrent functions • Stateless • Short lived (limit 5-15 mins) • Weak ordering guarantees • Cold start’s can be (very) slow: 100ms – 45s (AWS 250ms-7s)
41. Where is FaaS useful? • Spikey workloads and ‘occasional’ use cases • Use cases that don’t typically warrant massive parallelism e.g. CI systems. • General purpose programming paradigm?
42. But there are open questions
43. Serverless Developer Ecosystem • Runtime diagnostics • Monitoring • Deploy loop • Testing • IDE integration Currently quite poor
44. Harder than current approaches Easier than current approaches Amazon Google Microsoft Serverless programming will likely become prevalent
45. In the future it seems unlikely we’ll manage our own infrastructure. But where will we manage our data?
47. Event-Streaming approaches this from a different angle
48. FaaS is event-driven But it isn’t streaming
49. Complex, Timing issues, Scaling limits Customers Event Source Orders Event Source Payments Event Source Serverless functions handle only one event source FaaS/μS FaaS/μS FaaS/μS
50. A slightly more complex example: Send email only to platinum customers
51. Payments Event Source Event is received by serverless function FaaS/μS
52. Payments Event Source Block and calls the database to get customer+order FaaS/μS Get customer Get order
53. Payments Event Source Is it a ‘Platinum’ customer? FaaS/μS Get customer Get order Is the customer platinum?
54. Payments Event Source Send email if ‘Platinum’ FaaS/μS Get customer Get order Maybe send email
55. Payments Event Source Increase Load: 100 concurrant functions doing IO. FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS
56. Payments Event Source Only send 2 emails. FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS FaaS/μS
57. Send SQL Process boundary Orders Payments KSQL Customers Table Customers KSQL simplifies: App Logic CREATE STREAM foo AS SELECT * FROM orders, payments, customers LEFT JOIN… WHERE customer.type = ‘PLATINUM’ Order Payment Customer KSQL - Handle timing issues - No “per-event” IO. - Price efficient
58. Functions have no additional data dependencies: Everything is in the event!
59. Queries filter out the events you need (much like you filter rows in a database query)
60. FaaS FaaS FaaS KSQL Customers Table KSQL as a “Database” for Event-Driven Infrastructure FaaS FaaS Stateless, elastic compute Prepare the events we need (Sateful) Orders Payments Customers Autoscale with load
61. FaaS Traditional Application Event-Driven Application Application Database KSQL Stateful Data Layer FaaS FaaS FaaS FaaS FaaS Streaming Stateless Stateless Stateless Compute Layer Massive linear scalability with elasticity
62. Event-Driven vs. Event Streaming Event Driven Event Streaming Multiple Event Sources Use Database + ETL + Code Handles automatically Efficiency Extract data from DB in the FaaS (IO) Only the data you need Logic-driven data requests. Call DB from the FaaS (IO) DB/KStreams KqlDB?
64. Event Streaming Platform
65. Summary • Event Streaming provides the benefits of Event Sourcing to microservices and data pipelines. • Events are the data model. • Projections are the serving model: matching to each specific use case • Serving layer can be regenerated from the log (CQRS) • KSQL provides the same benefits for event-driven programs: e.g. preparing the event streams each FaaS application’s specific needs • In serverless architectures this drives efficiency: a ‘database- equivalent’ for event-driven infrastructure.
66. FaaS FaaS FaaS KSQL Can I Build This? FaaS FaaS AWS Lambda / Azure Functions Connectors (in Preview) Hosted KSQL In Preview Confluent Cloud
67. Things I didn’t tell you • Tools like KSQL provide data provisioning, not state mutation. • Use single writers. Try KSQL DB? • Can KSQL handle large state? • Unintended rebalance can stall processing • Static membership (KIP-345) – name the list of stream processors • Increase the timeout for rebalance after node removal (group.max.session.timeout.ms) • Worst case reload: RocksDB ~GbE speed • Can Kafka be used for long term storage? • Log files are immutable once they roll (unless compacted) • Jun spent a decade working on DB2 • Careful: • Historical reads can stall real-time requests (cached) • ZFS has several page cache optimizations • Tiered storage will help
68. Find out More • Peeking Behind the Curtains of Serverless Platforms, Wang et al. • Cloud Programming Simplified: A Berkeley View on Serverless Compute • Neil Avery’s Journey to Event Driven Part 3. The Affinity Between Events, Streams and Serverless. • Designing Event Driven Systems, Ben Stopford
69. Thank you @benstopford Book: https://www.confluent.io/designing-event-driven-systems