Troubleshooting Kafka's socket server: from incident to resolution

1. Troubleshooting Kafka’s socket server from incident to resolution Joel Koshy LinkedIn

2. The incident Occurred a few days after upgrading to pick up quotas and SSL Multi-port KAFKA-1809 KAFKA-1928 SSL KAFKA-1690 x25 x38 October 13 Various quota patches June 3April 5 August 18

3. The incident Broker (which happened to be controller) failed in our queuing Kafka cluster

4. The incident ● Alerts fire; NOC engages SYSOPS/Kafka-SRE ● Kafka-SRE restarts the broker ● Broker failure does not generally cause prolonged application impact ○ but in this incident…

5. The incident Multiple applications begin to report “issues”: socket timeouts to Kafka cluster Posts search was one such impacted application

6. The incident Two brokers report high request and response queue sizes

7. The incident Two brokers report high request queue size and request latencies

8. The incident ● Other observations ○ High CPU load on those brokers ○ Throughput degrades to ~ half the normal throughput ○ Tons of broken pipe exceptions in server logs ○ Application owners report socket timeouts in their logs

9. Remediation Shifted site traffic to another data center “Kafka outage ⇒ member impact Multi-colo is critical!

10. Remediation ● Controller moves did not help ● Firewall the affected brokers ● The above helped, but cluster fell over again after dropping the rules ● Suspect misbehaving clients on broker failure ○ … but x25 never exhibited this issue sudo iptables -A INPUT -p tcp --dport <broker-port> -s <other-broker> -j ACCEPT sudo iptables -A INPUT -p tcp --dport <broker-port> -j DROP

11. “Good dashboards/alerts Skilled operators Clear communication Audit/log all operations CRM principles apply to operations

12. Remediation Friday night ⇒ roll-back to x25 and debug later … but SREs had to babysit the rollback x38 x38 x38 x38 Rolling downgrade

13. Remediation Friday night ⇒ roll-back to x25 and debug later … but SREs had to babysit the rollback x38 x38 x38 x38 Rolling downgrade Move leaders

14. Remediation Friday night ⇒ roll-back to x25 and debug later … but SREs had to babysit the rollback x38 x38 x38 x38 Rolling downgrade Firewall

15. Remediation Friday night ⇒ roll-back to x25 and debug later … but SREs had to babysit the rollback x38 x38 x38 Rolling downgrade Firewall x25

16. Remediation Friday night ⇒ roll-back to x25 and debug later … but SREs had to babysit the rollback x38 x38 x38 Rolling downgrade x25 Move leaders

17. … oh and BTW be careful when saving a lot of public-access/server logs: ● Can cause GC [Times: user=0.39 sys=0.01, real=8.01 secs] ● Use ionice or rsync --bwlimit low low high

18. The investigation

19. ● Test cluster ○ Tried killing controller ○ Multiple rolling bounces ○ Could not reproduce ● Upgraded the queuing cluster to x38 again ○ Could not reproduce ● So nothing… Attempts at reproducing the issue

20. Understanding queue backups…

21. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory Quota manager

22. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory Quota manager New connections

23. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager and then turn off read interest from that connection (for ordering)

24. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time

25. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time

26. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time + remote-time long-poll requests

27. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time + remote-time long-poll requests Hold if quota violated + quota-time

28. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time + remote-time long-poll requests Hold if quota violated + quota-time Await processor + response-queue-time

29. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time + remote-time long-poll requests Hold if quota violated + quota-time Await processor + response-queue-time Write response + response-send-time

30. API layer Life-cycle of a Kafka request Network layer Acceptor Processor Processor Processor Processor Response queue Response queue Response queue Response queue Request queue API handler API handler API handler Clientconnections Purgatory New connections Read request Quota manager Await handling Total time = queue-time Handle request + local-time + remote-time long-poll requests Hold if quota violated + quota-time Await processor + response-queue-time Write response + response-send-time Turn read interest back on from that connection

31. Investigating high request times ● Total time is useful for monitoring ● but high total time is not necessarily bad

32. Investigating high request times ● Total time is useful for monitoring ● but high total time is not necessarily bad Low

33. Investigating high request times ● Total time is useful for monitoring ● but high total time is not necessarily bad Low High but “normal” (purgatory)

34. Investigating high request times ● First look for high local time ○ then high response send time ■ then high remote (purgatory) time → generally non-issue (but caveats described later) ● High request queue/response queue times are effects, not causes

35. High local times during incident (e.g., fetch)

36. How are fetch requests handled? ● Get physical offsets to be read from local log during response ● If fetch from follower (i.e., replica fetch): ○ If follower was out of ISR and just caught-up then expand ISR (ZooKeeper write) ○ Maybe satisfy eligible delayed produce requests (with acks -1) ● Else (i.e., consumer fetch): ○ Record/update byte-rate of this client ○ Throttle the request on quota violation

37. Could these cause high local times? ● Get physical offsets to be read from local log during response ● If fetch from follower (i.e., replica fetch): ○ If follower was out of ISR and just caught-up then expand ISR (ZooKeeper write) ○ Satisfy eligible delayed produce requests (with acks -1) ● Else (i.e., consumer fetch): ○ Record/update byte-rate of this client ○ Throttle the request on quota violation Not using acks -1 Should be fast Maybe Should be fast Delayed outside API thread

38. ISR churn?

39. ● Low ZooKeeper write latencies ● Churn in this incident: effect of some other root cause ● Long request queues can cause churn ○ ⇒ follower fetches timeout ■ ⇒ fall out of ISR (ISR shrink happens asynchronously in separate thread) ○ Outstanding fetch catches up and ISR expands ISR churn? … unlikely

40. High local times during incident (e.g., fetch) Besides, fetch-consumer (not just follower) has high local time

41. Could these cause high local times? ● Get physical offsets to be read from local log during response ● If fetch from follower (i.e., replica fetch): ○ If follower was out of ISR and just caught-up then expand ISR (ZooKeeper write) ○ Satisfy eligible delayed produce requests (with acks -1) ● Else (i.e., consumer fetch): ○ Record/update byte-rate of this client ○ Throttle the request on quota violation Not using acks -1 Should be fast Should be fast Delayed outside API thread Test this…

42. Maintains byte-rate metrics on a per-client-id basis 2015/10/10 03:20:08.393 [] [] [] [logger] Completed request:Name: FetchRequest; Version: 0; CorrelationId: 0; ClientId: 2c27cc8b_ccb7_42ae_98b6_51ea4b4dccf2; ReplicaId: -1; MaxWait: 0 ms; MinBytes: 0 bytes from connection <clientIP>:<brokerPort>-<localAddr>;totalTime:6589, requestQueueTime:6589,localTime:0,remoteTime:0,responseQueueTime:0,sendTime:0, securityProtocol:PLAINTEXT,principal:ANONYMOUS Quota metrics ??!

43. Quota metrics - a quick benchmark for (clientId ← 0 until N) { timer.time { quotaMetrics.recordAndMaybeThrottle(sensorId, 0, DefaultCallBack) } }

44. Quota metrics - a quick benchmark

45. Quota metrics - a quick benchmark Fixed in KAFKA-2664

46. meanwhile in our queuing cluster… due to climbing client-id counts

47. Rolling bounce of cluster forced the issue to recur on brokers that had high clientid metric counts ○ Used jmxterm to check per-client-id metric counts before experiment ○ Hooked up profiler to verify during incident ■ Generally avoid profiling/heapdumps in production due to interference ○ Did not see in earlier rolling bounce due to only a few client-id metrics at the time

48. MACRO ● Observe week-over-week trends ● Formulate theories ● Test theory (micro-experiments) ● Deploy fix and validate MICRO (live debugging) ● Instrumentation ● Attach profilers ● Take heapdumps ● Trace-level logs, tcpdump, etc. Troubleshooting: macro vs micro

49. MACRO ● Observe week-over-week trends ● Formulate theories ● Test theory (micro-experiments) ● Deploy fix and validate MICRO (live debugging) ● Instrumentation ● Attach profilers ● Take heapdumps ● Trace-level logs, tcpdump, etc. Troubleshooting: macro vs micro Generally more effective Sometimes warranted, but invasive and tedious

50. How to fix high local times ● Optimize the request’s handling. For e.g.,: ○ cached topic metadata as opposed to ZooKeeper reads (see KAFKA-901) ○ and KAFKA-1356 ● Make it asynchronous ○ E.g., we will do this for StopReplica in KAFKA-1911 ● Put it in a purgatory (usually if response depends on some condition); but be aware of the caveats: ○ Higher memory pressure if request purgatory size grows ○ Expired requests are handled in purgatory expiration thread (which is good) ○ but satisfied requests are handled in API thread of satisfying request ⇒ if a request satisfies several delayed requests then local time can increase for the satisfying request

51. as for rogue clients… 2015/10/10 03:20:08.393 [] [] [] [logger] Completed request:Name: FetchRequest; Version: 0; CorrelationId: 0; ClientId: 2c27cc8b_ccb7_42ae_98b6_51ea4b4dccf2; ReplicaId: -1; MaxWait: 0 ms; MinBytes: 0 bytes from connection <clientIP>:<brokerPort>-<localAddr>;totalTime:6589, requestQueueTime:6589,localTime:0,remoteTime:0,responseQueueTime:0,sendTime:0, securityProtocol:PLAINTEXT,principal:ANONYMOUS “Get apps to use wrapper libraries that implement good client behavior, shield from API changes and so on…

52. not done yet! this lesson needs repeating...

53. After deploying the metrics fix to some clusters…

54. After deploying the metrics fix to some clusters… Deployment Persistent URP

55. After deploying the metrics fix to some clusters… Applications also begin to report higher than usual consumer lag

56. Root cause: zero-copy broke for plaintext Upgraded cluster Rolled back With fix (KAFKA-2517)

57. The lesson...

58. ● Request queue size ● Response queue sizes ● Request latencies: ○ Total time ○ Local time ○ Response send time ○ Remote time ● Request handler pool idle ratio Monitor these closely!

59. Continuous validation on trunk

61. Local times ConsumerMetadata OffsetFetch ControlledShutdown Offsets (by time) Fetch Produce LeaderAndIsr StopReplica (for delete=true) TopicMetadata UpdateMetadata OffsetCommit These are (typically 1:N) broker-to-broker requests

62. API layerNetwork layer Broker-to-broker request latencies - less critical Processor Response queue API handler Purgatory Acceptor Read bit off so ties up at most one API handler Request queue

63. API layerNetwork layer Broker-to-broker request latencies - less critical Processor Response queue API handler Purgatory Acceptor Request queueBut if requesting broker times out and retries…

64. API layerNetwork layer Broker-to-broker request latencies - less critical Processor Response queue API handler Purgatory Acceptor Request queueBut if requesting broker times out and retries… Processor Response queue API handler

65. API layerNetwork layer Broker-to-broker request latencies - less critical Processor Response queue API handler Purgatory Acceptor Request queue Processor Response queue API handler Configure socket timeouts >> MAX(latency)

66. ● Broker to controller ○ ControlledShutdown (KAFKA-1342) ● Controller to broker ○ StopReplica[delete = true] should be asynchronous (KAFKA-1911) ○ LeaderAndIsr: batch request - maybe worth optimizing or putting in a purgatory? Haven’t looked closely yet… Broker-to-broker request latency improvements

67. The end