Function Composition in a Serverless World (OSCON 2018)

Transcript

1. Function Composition in a Serverless World Soam Vasani Platform9 Systems !1 @fissionio
2. First, what’s FaaS? Function-as-a-Service enable developers to deploy parts of an application on an “as needed” basis using short-lived functions. Benefits of FaaS: • Complete abstraction of servers away from the developer • Billing based on consumption and executions, not server instance sizes • Scaling services is simplified !2
3. What is Function Composition? Joining small functions to create bigger, more complex functions. !3
4. Example App !4 Recognize Image Function A Function B Cat Cat Translate Gato
5. !5 How should we combine both functions into one? Recognize + Translate Gato “Composed” Function
6. Many approaches !6 •Manual compilation •Direct function calls or “chaining” •Coordinator pattern •Event-driven •Workflows In increasing order of complexity:
7. Manual Compilation Merge functions on a source code level. • One big function that calls all other individual functions. • One big task from FaaS framework’s point-of-view. !7
8. !8 func recognizeImage(image) { // A: Call API for image tagging. } func translate() { // B: Translate text } func combo() { recognizeImage(…) translate(…) }
9. Pros: • Very simple, no framework needed at all • No serialization overhead Cons: • Function gets bigger; more expensive and may load slowly • Cannot scale independently !9
10. Merged Function Scaling Function A Function B vs. Instance 1 Instance 2 Instance 1 Instance 2 !10 Function B Function B Function A Function A Function B Function A Function A
11. Direct Function Calls (chaining) Form a chain, calling each other over HTTP. • Each task is a separate FaaS function. • Each function knows what comes after it and calls it. !11
12. func recognizeImage(image) { // A: Recognize image // HTTP call to translation function } func translate() { // B: Translate text } !12
13. Pros: Cons: • Each function waits for the next function, wasting $ • Messaging complexity: handling failures, and thinking about fallbacks, retries. • Function updates are error-prone • Cold starts add up • No external components needed • Independent scaling, unlike manual compilation !13
14. Coordinator Pattern Functions that manage the execution of other functions. • One “omniscient” function calls each function (via remote HTTP); manages the execution flow. • Similar to direct functions, except each function is unaware of the other functions. !14
15. !15
16. Pros: • No need to modify the primitive functions • Very flexible; user can manipulate the control flow how they like. (Separation of concerns) Cons: • Overhead of an extra function • Even more cold starts • Coordinator is a long running function (it starts first, and ends last). !16
17. !17 Adding components to support composition • Message Queues • Workflow Systems
18. Event-Driven Function Composition !18 Focus on the data flow instead of the control flow: computation as reactions to events
19. !19
20. !20
21. • Get all the luxury of message queues (e.g. message retries, error handling, persistence). • Loosely coupled functions • Commonly used and well understood architecture. Pros • Web of implicit dependencies. • Difficult to version or upgrade functions. • Supports limited control flow constructs. (e.g. conditional and on-error constructs) Cons !21
22. Workflows Create a “flowchart” of function interactions. !22
23. !23
24. Workflows are everywhere! Business Processes Data Pipelining DevOps Apache Airflow !24
25. FaaS-focused Workflows Azure Logic Apps AWS Step Functions !25
26. Architecture Deploys on Kubernetes and Fission Stores events not state Executes state machines !26
27. • Centralization of composition logic, logging, and visualization • Upgrade a workflow + set of functions “consistently” • Loosely coupled functions • Handles communication complexity (latency, retries, failures, etc.) • Improved performance (better/anticipating scheduling of functions) Pros • More infrastructure complexity • Need to learn workflow-specific language or DSL Cons !27
28. Workflows Demo !28
29. Approaches (recap) !29 •Manual compilation
30. Approaches (recap) !30 •Manual compilation •Direct function calls or “chaining”
31. Approaches (recap) !31 •Manual compilation •Direct function calls or “chaining” •Coordinator pattern
32. Approaches (recap) !32 •Manual compilation •Direct function calls or “chaining” •Coordinator pattern •Event-driven
33. Approaches (recap) !33 •Manual compilation •Direct function calls or “chaining” •Coordinator pattern •Event-driven •Workflows
34. Which approach should you use? !34 • It Depends! • Function size • Function resource usage • Your app’s performance and cost requirements • Tolerance for complexity and operational effort
35. See some examples: https://github.com/fission/faas-composition-patterns !35
36. Thank you! http://fission.io slack.fission.io github.com/fission Twitter: @fissionio !36
37. Additional slides !37
38. • A workflow is just another function! • No magic/privileged functions or constructs • Dynamic tasks: use conditional branches. • (JavaScript) expressions in workflows for simple data selections and transformations • Tools for debugging and inspecting workflows Advanced Concepts !38
39. YAML Workflow RecognizeImage Translate tasks: RecognizeImage: … Translate: … requires: - RecognizeImage This YAML file specifies the sequence of the function logic. !39
40. Pros: • Workflow Framework sees the whole picture • “Consistent” Upgrades (in theory at least…) • Loosely coupled functions, unaware that they are being used in a workflow • Messaging abstracted away: latency, retries, failures, queueing… • Can be reliable/fault-tolerant. • Engine predicts invocations to improve the performance (pre-boot functions). Cons: • Adds additional component = complexity • Workflow-specific language to learn… usually YAML (some sort of learning curve), Usually not a general purpose language • Hard to unit test and debug, depending on maturity of framework !40