Building Better APIs with gRPC - Postmates Android Tech Talk

This talk was a part of the "&droid" event at Postmates HQ.
gRPC is usually used by backend services to communicate with each other.
In the talk we go over our experiences of building gRPC apis for mobile clients.
Speakers:
- Parth Mehta
- Veeren Mandalia

Transcript

2. &droid 
 Postmates Android Tech Talk July 23, 2019
3. About Postmates Postmates is on a mission to help people unlock the best of their cities – and their lives – with an insanely reliable on-demand “anything” network.
4. By the Numbers 4 Major Offices (SFO - BNA - SEA - MEX) 1,300 Employees 3,500 Cities 500K Merchants 140K Couriers Per Week 6.5M Deliveries Per Month
5. Buyer Fleet Seller The Three Apps
6. Speakers James 
 Lee Android Engineering Lead Veeren Mandalia Android Engineering Lead Hai 
 Le Android Engineer Parth 
 Mehta Server Engineering Lead Gaurav 
 Murti Android Engineer Bianca Curutan Mobile Engineer MC
7. Fleet Building Better APIs 
 with gRPC Veeren Mandalia
 Parth Mehta
8. Problem • Improve platform responsiveness and reliability - e.g., Postmates only have a limited time when viewing a delivery offer, every second matters • Reduce engineering time - API definition & implementation
9. • Remote Procedure Call-based • High performance because of multiplexing and compression • Open source • Multiplatform
10. Shared API Definition • Service and models defined in proto definition files • Definition shared between server & Android/iOS clients • Serves as the single source of truth of API definition and documentation
11. Service Definition REST (Kotlin) @GET("v2/courier/events") fun getEvents(@Query("token") token: String, @Query("limit") limit: Int): Observable<EventsResponse>
12. Service Definition gRPC (proto definition file) service FleetService { rpc Events (EventsRequest) returns (stream EventsResponse) {} }
13. Model Definition message EventsResponse { string token = 1; repeated Event events = 2; } message EventsRequest { string token = 1; bool limit = 2; }
14. Tight API Contracts service FleetService { rpc Events (EventsRequest) returns (stream EventsResponse) {} }
15. Tight API Contracts message EventsResponse { string token = 1; repeated Event events = 2; } message EventsRequest { string token = 1; bool limit = 2; }
16. Tight API Contracts message EventsError { oneof error_oneof { InvalidToken invalid_token = 1; UnsupportedLimit unsupported_limit = 2; Error generic = 3; } }
17. gRPC is Multiplatform
18. Generated Code • The protobuf compiler (protoc) then generates language specific framework and application code • Output - Client: Java, Swift - Server: Go, Python
19. Generated Code Client Service Client Model
20. API Updates CI CI message EventsRequest { string token = 1; bool limit = 2; enum event_type = 3; }
21. Creating a Stub val target = "fleet.postmates.com:428" val channel: ManagedChannel = OkHttpChannelBuilder .forTarget(target) .keepAliveTime(15, SECONDS) .keepAliveTimeout(10, SECONDS) .userAgent(“Postmate Android User Agent”) .let { builder -> AndroidChannelBuilder.fromBuilder(builder) .context(context) } .build()
22. Creating a Stub val fleetServiceStub = FleetServiceGrpc.newStub(channel) val fleetServiceBlockingStub = FleetServiceGrpc.newBlockingStub(channel)
23. Make a Request val request = EventsRequest.newBuilder() .setLimit(20) .setToken("token") .build()
24. Make a Request util.getObservableFromRpc<EventsRequest, EventsResponse> { streamObserver -> fleetServiceStub.events(request, streamObserver) }.observeOn(ioScheduler) .subscribe({ eventsResponse -> // Process events, e.g. Show offer to Postmate }, { throwable -> // Handle error })
25. RxJava Utilities fun <Req, Resp> getObservableFromRpc(rpc: (StreamObserver<Resp>) -> Unit): Observable<Resp> { return Observable.fromEmitter({ emitter: Emitter<Resp> -> val responseObserver = object : CancellableStreamObserver<Req, Resp>() { override fun onError(t: Throwable) { super.onError(t) emitter.onError(t) } override fun onNext(value: Resp) = emitter.onNext(value) override fun onCompleted() = emitter.onCompleted() } emitter.setCancellation { responseObserver.cancel() } rpc(responseObserver) }, Emitter.BackpressureMode.LATEST) }
26. RxJava Utilities abstract class CancellableStreamObserver<Req, Resp> : ClientResponseObserver<Req, Resp> { var requestStream: ClientCallStreamObserver<Req>? = null override fun beforeStart(requestStream: ClientCallStreamObserver<Req>?) { this.requestStream = requestStream } @CallSuper override fun onError(t: Throwable) { // Handle error.. } fun cancel() { requestStream?.cancel("Forceful cancellation of stream", StatusException(Status.CANCELLED)) } }
27. Under the Hood Channel Server RPC MSG Server
28. Built on Top of HTTP2
29. Grpc / Http2 Channel Server RPC MSG Stream F F F Connection
30. Performance Multiplexing • Multiple ongoing RPCs over a single connection Compact over the wire • Protobuf over the wire • Compression
31. Bi-Directional Streaming • gRPC supports Bi-Directional streaming • Faster and more efficient data transfer over a persistent connection • Ideal for payloads like events
32. Push and Fetch Server Streaming Server FCM
33. Event Latency on REST vs gRPC 0% 15% 30% 45% 60% 10ms 50ms 100ms 250ms 500ms 750ms 1000ms 1500ms 2000ms SLOW gRPC REST
34. We're Hiring! 
 https://careers.postmates.com