Finagle Overview

Finagle Overview presented to the NY Scala Meetup

Transcript

1. Finagle
2. Hello!
3. Finagle Extensible RPC Implement a feature once => get the feature in every service => get the feature in every protocol
4. Twitter Futures trait Future[+A] { def map[B](fn: A => B): Future[B] def flatMap[B](fn: A => Future[B]): Future[B] def rescue[B](fn: PartialFunction[Throwable, Future [B]]): Future[B] def handle[B](fn: PartialFunction[Throwable, B]): Future [B] }
5. Twitter Futures Interrupts Maintain context Scheduler
6. Finagle Abstractions Service Server Client
7. Service trait Service[-Req, +Rep] extends (Req => Future[Rep])
8. Server trait Server[Req, Rep] { def serve( addr: SocketAddress, underlying: Service[Req, Rep] ): ListeningServer }
9. Client trait Client[Req, Rep] { def newService( dest: Name, label: String ): Service[Req, Rep] }
11. Load Balancing Load balances to the least loaded Internally organized as a heap Implicitly sends more requests to faster servers
12. Liveness
13. Failure Accrual After X failures in a row, backoff for duration Y
14. Rolling Restart Support If we can’t establish a connection to a host, blackhole it and try to reestablish in the background “Fail Fast”
15. Connection Pooling Tunable parameters: Low Watermark High Watermark Idle Timeout Max Waiters
16. Stats Collection We collect stats on every request, including histograms of request latency, success rate etc. See: Metrics, Metrics Everywhere (codahale)
17. Distributed Tracing Hackweek project inspired by Dapper Sample a small number of requests Lets you see which upstreams are pummeling your service during an incident Profiling is good for debugging average latency Tracing is good for debugging long tail latency
19. Pipelining Redis Memcached
20. Service Discovery trait ListeningServer { def announce(addr: String): Future[Announcement] }
21. Service Discovery ZK internally Other schemes Adding finagle-serversets to your classpath enables ZK
22. Next?!
23. finagle 6 apis ClientBuilder => Protocol.new Cargo Culting :( Finagle notoriously hard to configure Zero configuration! Works out of the box.
24. old style val client: Service[HttpRequest, HttpResponse] = ClientBuilder() .codec(Http()) .hosts(address) .hostConnectionLimit(1) .tcpConnectTimeout(50.milliseconds) .failFast(true) .daemon(true) .retries(3) .trace(ZipkinTracer.mk(host, port) .build()
25. finagle 6 style Http.newService(name, label)
26. mux A new protocol, layered on other protocols cf. thrift-mux
27. mux discard leasing multiplexing error encoding ping drain initialization
28. error encoding Means you can handle errors separately by just reading the first byte Easy to imagine dropping errors, or minimizing copies etc
29. init + ping True liveness No more failfast or failure accrual
30. draining Signals that a server will start draining Client knows to stop sending new requests
31. multiplexing Pipelining for everyone! Connection pooling mostly disappears
32. discard No longer care about response Current world: close connection Biggest cause of connection churn New world: send a discard request
33. leasing First target: GC avoidance Useful elsewhere too
34. Finagle Resources twitter.github.io/finagle finaglers@google groups
35. Work on Finagle github.com/twitter/finagle Docs! Github Issues!
36. Thanks! twitter: @mnnakamura github: mosesn email: moses@twitter.com
37. Work for Twitter! https://twitter.com/jobs We’re hiring! t.co/nymeetup
38. Questions? https://twitter.com/jobs/engineering t.co/nymeetup email: moses@twitter.com