The once and future layer 5: Resilient, Twitter-style microservices

What is required to operate microservices at scale? Beyond containers, schedulers, and frameworks, what is actually required to turn hundreds of services, tens of thousands of machines, and millions of requests per second into a unified, performant application? Oliver Gould explores the evolution of Twitter’s stack from monolith to highly distributed microservices and the surprising glue that held it all together: layer 5 in the OSI model, the oft-overlooked session layer.

Oliver offers an overview of Finagle, the high-scale RPC library developed at Twitter and adopted by Pinterest, SoundCloud, ING Bank, and other companies, tracing Finagle’s evolution from a simple library into something much more: a unified, global mechanism for operability and control over a highly disaggregated application architecture. Oliver explains how this mechanism provides Twitter with higher-level, service-based semantics around scalability, reliability, and fault tolerance and how the control over layer 5 afforded by Finagle allowed Twitter to solve some of the most surprising and difficult problems with its highly distributed architecture—when the software architecture diagram and the org chart intersected.

Oliver concludes by introducing linkerd, an open source proxy form of Finagle, which extends Finagle’s operational model to non-JVM or polyglot microservices, and demonstrates how linkerd can be used to “wrap” multiservice applications, independent of application language(s) or infrastructure, to obtain many of the benefits that Finagle provides for Twitter.

Transcript

1. The Once and Future Layer 5 Resilient, Twitter-style microservices oliver gould
 cto, buoyant Velocity NYC, September 22 2016 from
2. oliver gould • founding cto @ buoyant
 open-source microservice infrastructure • previously, tech lead @ twitter: • observability • traffic • core contributor: finagle • creator: linkerd • likes: dogs • dislikes: being woken up for computers @olix0r
 ver@buoyant.io
3. overview • 2010: Riding the Whale • “Microservices” • The Once and Future Layer 5 • Introducing linkerd • Demotime!
4. 2010 A FAILWHALE ODYSSEY
5. Twitter, 2010 107 users 107 tweets/day 102 engineers 101 services 101 deploys/week 102 hosts 10-1 datacenters 101 user-facing outages/week https://blog.twitter.com/2010/measuring-tweets
8. Events https://blog.twitter.com/2013/new-tweets-per-second-record-and-how
9. Asymmetry Photo by @troy
10. Provisioning
11. microservices everything’s easier with (not)
12. scaling teams growing software
13. flexibility
14. performance correctness monitoring debugging efficiency security operability
 resilience
15. microservices are difficult.
16. Resilience is an imperative: our software runs on the truly dismal computers we call datacenters. Besides being heinously
 complex… they are unreliable and prone to
 operator error. Marius Eriksen @marius
 RPC Redux
17. resilience in microservices software you didn’t write hardware you can’t touch network you can’t configure break in new and surprising ways and your customers shouldn’t notice
18. resilient microservices means resilient communication
19. datacenter [1] physical [2] link [3] network [4] transport kubernetes 
 canal, weave, … aws, azure, digitalocean, gce, … business languages, libraries [7] application rpc [5] session [6] presentation json, protobuf, thrift, … http/2, mux, …
20. layer 5 dispatches requests onto layer 4 connections
21. finagle THE ONCE AND FUTURE LAYER 5
22. github.com/twitter/finagle RPC library (JVM) asynchronous built on Netty scala functional strongly typed first commit: Oct 2010 (@nk + @marius)
23. used by…
24. programming finagle val users = Thrift.newIface[UserSvc](“/s/users”)
 val timelines = Thrift.newIface[TimelineSvc](“/s/timeline”) Http.serve(“:8080”, Service.mk[Request, Response] { req => for { user <- users.get(userReq(req)) timeline <- timelines.get(user) } yield renderHTML(user, timeline) })
25. your server is a function trait Service[Req, Rsp] { def apply(req: Req): Future[Rsp] def close(deadline: Time): Future[Unit] }
26. your server is a function trait ServiceFactory[Req, Rsp] { def apply(conn: ClientConnection): Future[Service[Req, Rsp]] def close(deadline: Time): Future[Unit] }
27. your server is a function trait Filter[InReq, OutRsp, OutReq, InRsp] { def apply(req: InReq, service: Service[OutReq, InRsp]): Future[OutRsp] def andThen[A, B](f: Filter[OutReq, InRsp, A, B]): Filter[OutReq, InRsp, A, B] def andThen[A, B](s: Service[A, B]): Service[OutReq, InRsp] def andThen[A, B](sf: ServiceFactory[A, B]): ServiceFactory[OutReq, InRsp] }
28. your server is a function val service: Service[http.Request, http.Response] = recordHandletime andThen traceRequest andThen logRequest andThen timeouts andThen myService val server: ListeningServer = Http.serve(“:8080”, service) val client: ServiceFactory[http.Request, http.Response] = retries andThen Http.newClient(“127.1:8080”)
29. operating finagle transport security service discovery circuit breaking backpressure deadlines retries tracing monitoring keep-alive multiplexing load balancing per-request routing service-level objectives Observe Session timeout Retries Request draining Load balancer Monitor Observe Trace Failure accrual Request timeout Pool Fail fast Expiration Dispatcher
30. layer 5 naming
31. layer 5 naming applications refer to logical names
 requests are bound to concrete names
 delegations express routing /s/users /#/io.l5d.zk/prod/users/http /s => /#/io.l5d.zk/prod/http
32. per-request routing: staging GET / HTTP/1.1
 Host: mysite.com
 Dtab-local: /s/B => /s/B2
33. per-request routing: debug proxy GET / HTTP/1.1
 Host: mysite.com
 Dtab-local: /s/E => /s/P/s/E
34. “It’s slow”
 is the hardest problem you’ll ever debug. Jeff Hodges @jmhodges
 Notes on Distributed Systems for Young Bloods
35. the more components you deploy, the more problems you have
36. the more components you deploy, the more problems you have
37. the more components you deploy, the more problems you have
38. tracing
39. tracing
40. tracing
41. lb algorithms: • round-robin • fewest connections • queue depth • exponentially-weighted moving average (ewma) • aperture load balancing at layer 5
42. timeouts & retries timelines users web db timeout=400ms retries=3 timeout=400ms retries=2 timeout=200ms retries=3 timelines users web db
43. timeouts & retries timelines users web db timeout=400ms retries=3 timeout=400ms retries=2 timeout=200ms retries=3 timelines users web db 800ms! 600ms!
44. deadlines timelines users web db timeout=400ms deadline=323ms deadline=210ms 77ms elapsed 113ms elapsed
45. retries typical: retries=3
46. retries typical: retries=3 worst-case: 300% more load!!!
47. budgets typical: retries=3 better:
 retryBudget=20% worst-case: 300% more load!!! worst-case: 20% more load
48. cancellation timelines users web db timelines users web db timeout!
49. cancellation timelines users web db timelines users web db timeout!
50. Nacking timelines users web db timelines users web db
51. Nacking timelines users web db timelines users web db nack!
52. Nacking timelines users web db timelines users web db requeue
53. magic ops sprinkles transport security service discovery circuit breaking backpressure deadlines retries tracing metrics keep-alive multiplexing load balancing per-request routing service-level objectives Observe Session timeout Retries Request draining Load balancer Monitor Observe Trace Failure accrual Request timeout Pool Fail fast Expiration Dispatcher
55. So just rewrite everything in Finagle?
56. linkerd
57. github.com/buoyantio/linkerd microservice rpc proxy layer-5 router aka l5d built on finagle & netty pluggable http, thrift, … consul, etcd, k8s, marathon, zk, … …
58. magic operability sprinkles transport security service discovery circuit breaking backpressure deadlines retries tracing metrics keep-alive multiplexing load balancing per-request routing service-level objectives Service B instance linkerd Service C instance linkerd Service A instance linkerd
59. namerd service discovery service delegates logical names to service discovery centralized routing policy pluggable consul, etcd, k8s, zk, …
60. namerd
61. demo: gob’s microservice
62. kubernetes
63. host kubelet pod app: b ip: 10.2.3.4 container container pod app: a ip: 10.1.2.3 container
64. host app: a app: b app: a host app: b app: a app: b service-a
65. web word gen l5d l5d l5d
66. web word gen gen-v2 l5d l5d l5d l5d
67. web word gen gen-v2 l5d l5d l5d l5d namerd
68. github.com/buoyantio/linkerd-examples
69. linkerd roadmap • HTTP/2+gRPC linkerd#174 • Deadline Enforcement (in progress) • Dark Traffic • Improved namerd API • All configurable everything
70. more at linkerd.io slack: slack.linkerd.io email: ver@buoyant.io twitter: • @olix0r • @linkerd thanks!