Towards Stateful FaaS on Streaming Dataflow

Transcript

1. Berlin • 08 October 2019 Towards Stateful FaaS Flink Forward Marios Fragkoulis & Adil Akhter
2. Outline
3. • Motivation • Rho - FaaS beyond Stateless • Research Direction • Conclusion 3
4. Motivation
5. Web application DB Credit for icons: h9ps://www.fla?con.com 5
6. Web application DB Operations Configura?on Programming model Credit for icons: https://www.flaticon.com 6 Debugging Failure handling Transaction management Fault tolerance Scaling Monitoring Deployment
7. Web application DB IaaS PaaS SaaS Serverless FaaS Credit for icons: https://www.flaticon.com 7 Operations Configura?on Programming model Debugging Failure handling Transaction management Fault tolerance Scaling Monitoring Deployment
8. < > MS Azure Functions Google Cloud func?ons Credit for logos: https://aws.amazon.com/lambda/ https://cloud.google.com/functions/ FaaS 8
9. FaaS Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Cloud storage Credit for icons: https://www.flaticon.com 9 Managed infrastructure Function-based programming model ✅
10. FaaS Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Fn Cloud storage Stateless functions Stateful functions Fn-to-fn calls Coordination ❌ Credit for icons: h9ps://www.fla?con.com 10
11. If not FaaS then what? 11
12. 12 Services Architecture (1): Easiest Implem. Order Business Logic Stock Business Logic Payments Business Logic DB RPC Call RPC Response RPC Call RPC Response RPC Call RPC Response REST Call REST Call ▪ Perform an order iff there is stock available and the payment is cleared. ▪ Services are stateless ▪ Database does the heavy- liVing ▪ High latency, costly coordina?on calls
13. 13 ▪ Make state of each service local to the business logic. ▪ Services now are stateful ▪ Low-latency access to local state ▪ Service calls s?ll expensive ▪ Not obvious how to scale this out. Services Architecture (2): Embedded State/DB Order DB Business Logic Stock DB Business Logic Payments DB Business Logic REST Call REST Call
14. 14 ▪ Each message exchange/change to the state goes through an event-log. ▪ Services are asynchronous/reac?ve. ▪ If we lose state, we replay the log and rebuild it. ▪ Time-travel debugging, audits, etc. are trivial. Services Architecture (3): Event Sourcing Order DB Business Logic Stock DB Business Logic Payments DB Business Logic REST Call REST Call event-log event-log
15. 15 Services Architecture (4): Scalable Deployment RPC Calls Subscribe for Responses event-log event-log Order 1 Business Logic Order 2 Business Logic Order 3 Business Logic Stock 1 Business Logic Stock 2 Business Logic Payment 1 Business Logic event-log DB DB DB DB DB DB
16. 16 ▪ Millions of events per second on a couple of machines ▪ Consistent snapshots of state, with exactly-once guarantees ▪ We can scale-in/out operators to handle varying workloads Does it ring a bell? OP1 OP4 OP2 OP5 Input Message Queues OP3 Output Message Queue Apache Flink
17. Stateful FaaS on Streaming Dataflow 17 Time-travel debugging using checkpoints and message broker Guaranteed message delivery and exactly-once processing Each operator executes a group of functions that share the same state Operator-local state partitioned on key input for scalability and fault-tolerance
18. 18 Dataflow graphs could serve as a scalable backend for microservices and cloud applica?ons based on stateful func?ons. Dataflow graphs as a backend for Stateful Services
19. Recap 19
20. 20
21. 21
22. 22
23. Built using largely ad-hoc, time- consuming, low-level programming.
25. 26
26. 27
27. ρ (rho) Function-as-a-Service beyond Stateless
28. Programming Model
29. Domain Codomain ⇒ λ I O ::
30. 1. Pure Function (Fn) without State 2. Fn with State 3. Fn that supports Orchestration Fn ≡ λ
31. 32 def pingFn (p: Ping): Pong = Pong(id, "Pong!") Stateless λ
32. 34 def pingFn (p: Ping, ctx: ExecutionContext[Pong]): λ[Pong] = ctx.returnWith(Pong(id, "Pong!"))
33. Stateful λ 35 ‑ def sPingFn (p: Ping, ctx: ExecutionContext[Pong], state: PingCounter): λ[Pong] case class PingCounter( name: String, i: Long, allRequestString: Seq[String]) extends ManagedState def pingFn(p: Ping, ctx: ExecutionContext[Pong]): λ[Pong]
34. 36 def sPingFn(p: Ping, ctx: ExecutionContext[Pong], state: PingCounter): λ[Pong] = { val message = ”Pong" for { p ← ctx.persist(Pong(p.shardId, s"$message at ${java.time.Instant.now()}")) } yield p }
35. FnNamespace Every Fn (λ) is part of FnNamespace (can be considered as BoundedContext). 37 FnNamespace λ1 λ2 λn class PingServiceFn extends FnNamespace { override def descriptor: FnNamespaceDescriptor = named("PingService") .withQualifiedPaths( register("ping", pingFn _), register(”sPing", sPingFn _), register(”pingStat", pingStatFn _)) // rest of the implementation }
36. State Management 38 FnNamespace λ1 λ2 λn
37. 39 class PingServiceFn extends FnNamespace { type State = PingCounter def initialState: State = PingCounter("PingCounter", 0, Seq.empty) // rest of the implementation } FnNamespace λ1 λ2 λn
38. 40 FnNamespace λ1 λ2 λn
39. 41 def sPingFn(p: Ping, ctx: ExecutionContext[Pong], state: PingCounter): λ[Pong] = { val message = ”Pong" for { p ← ctx.persist(Pong(p.shardId, s"$message at ${java.time.Instant.now()}")) } yield p } Recall sPingFn
40. class PingServiceFn extends FnNamespace { // … def onPersist: EventHandler[State] = { case (Pong(_, s), PingCounter(n, i, allRequests)) ⇒ PingCounter(n, i + 1, s +: allRequests) case (_, state) ⇒ state // Do nothing with the State } // rest of the implementation } 42 FnNamespace λ1 λ2 λn type EventHandler[S] = PartialFunction[(FnResponse, S), S]
41. Putting it all together 43
42. Execution Semantics
43. 45 We strongly believe that streaming dataflows can have a central place in service-oriented architectures, taking over the execution of acid transactions, ensuring message delivery and processing, in order to perform scalable execution of services.
44. 46 Compiler
45. 47 FnNamespace1 FnNamespace2 FnNamespacen ReqQ1 ResQ1 ReqQ2 ResQ2 ResQn ReqQn
46. 48 CLI Gateway FnNamspace 1 FnNamspace 2 FnNamspace N Events
47. Recall PingServiceFn 49
48. 50 1 2 3
49. Orchestrator λ The provided programming abstraction supports calling another Fn (λ) from the same FnNamespace or from different FnNamespace available in the system. 51
50. 52 OrderService PaymentSerivce StockSerivce reserveBalance prepareOrder
51. 53 for { p ← ctx.callFn[PaymentFnRequest, PaymentFnResult]("PaymentFn.reserveCredit", pr) _ ← ctx.persist(p) // updating the state s ← ctx.callFn[PrepareStockRequest, StockFnResponse]("StockFn.prepareOrder", ps) _ ← ctx.persist(s) // updating the state } yield orderCreationResponse(r, p, s) OrderService PaymentSerivce StockSerivce reserveBalance prepareOrder
52. 54
53. 55 for { p ← ctx.callFn[PaymentFnRequest, PaymentFnResult]("PaymentFn.reserveCredit", pr) _ ← ctx.persist(p) // updating the state s ← ctx.callFn[PrepareStockRequest, StockFnResponse]("StockFn.prepareOrder", ps) _ ← ctx.persist(s) // updating the state } yield orderCreationResponse(r, p, s) OrderService PaymentSerivce StockSerivce reserveBalance prepareOrder
54. 56 FnNamespace1 FnNamespace2 FnNamespacen ReqQ1 ResQ1 ReqQ2 ResQ2 ResQn ReqQn Recall
55. 57 Rho in Action
56. 58 OrderService PaymentSerivce StockSerivce reserveBalance prepareOrder
57. $ rho deploy -n OrderFn --parallelism 2 --skip_jar $ rho deploy -n PaymentFn --parallelism 2 --skip_jar $ rho deploy -n StockFn --parallelism 2 --skip_jar 59
58. 60
59. 61
60. 62
61. 63
62. Performance
63. 65
65. 67 What’s Next?
66. 68
68. Thanks
69. 71 Online introductory course on stream processing starts on January 15th, covering all fundamental stream processing concepts (?me, order, windows, joins, etc.). We are using Apache Flink for assignments & include invited talks from industry & academia. Enrollment is open. tudelft.nl/taming-big-data-streams Shameless plug by Asterios Katsifodimos @kasterios
70. Questions ?
71. 73 Adil Akhter Tech Lead at ING Amsterdam, The Netherlands http://coyoneda.xyz adilakhter
72. References 1. “Stateful FuncCons as a Service in AcCon”, Adil Akhter, Marios Fragkoulis, Asterios Katsifodimos. In the Proceedings of the 45th Interna?onal Conference on Very Large Data Bases (VLDB) 2019. 2. OperaConal Stream Processing: Towards Scalable and Consistent Event-Driven ApplicaCons: Asterios Katsifodimos, Marios Fragkoulis. In the Proceedings of the 22nd Interna?onal Conference on Extending Database Technology (EDBT) 2019. 3. "Benchmarking Distributed Stream Data Processing Systems": Jeyhun Karimov, Tilmann Rabl, Asterios Katsifodimos, Roman Samarev, Henri Heiskanen, Volker Markl. In the Proceedings of the Interna?onal Conference on Data Engineering (ICDE) 2018. 4. “Efficient Window AggregaCon with General Stream Slicing”: Jonas Traub, Philipp M. Grulich, Alejandro Rodriguez Cuellar, Sebas?an Breß, Asterios Katsifodimos, Tilmann Rabl and Volker Markl. In the Proceedings of the 22nd Interna?onal Conference on Extending Database Technology (EDBT) 2019. 74
73. 75