Incident documentation/20200126-app server latency

From Wikitech
Jump to navigation Jump to search

document status: final

Summary

Note: the corresponding maps outage is not treated here. Its timings will just be reported in the timeline for reference.

At 19:47 UTC, the news of Kobe Bryant's death was announced. This caused a surge in both edits and page views, causing a stampede of requests and a general slowdown of both the application servers and the apis. This caused both high contention in editing/parsing as expected, and can be seen by the number of poolcounter full queues, as well as high cpu usage, very high response times, and general unavailability of the MediaWiki application layer (both appservers and api). This was caused by a series of concauses:

  • High contention in editing/parsing, resulting in a lot of workers blocked on waiting for poolcounter.
  • Saturation of the network link of two memcached servers due to the high edit/reparse activity. This caused about 1/9th of the memcached keys to become unavailable
  • A spike in latencies resulting from the failure of the memcached servers caused a rise in the cache-misses for babel calls from the appservers, flooding the api servers. pretty much as the same mechanism of the incident of February 4 (although the ignition of the incident was different).

Given the sustained higher level of requests, it took more than one hour for the clusters to recover.

Number of full queues on poolcounter during the incident.


Impact

1 hour of almost complete unavailability of the backend servers. During that period, a large number of requests to our sites will have timed out or got an error response.

As can be seen from the graph of all requests, the average request rate was just barely under what the normal rate is, but we can estimate there was a large number of requests that we weren't able to answer to. Almost 90% of the backend requests from ATS failed.

Successful responses from all of our caching frontends during the incident.

Detection

The monitoring correctly and promptly detected the issue, and 3 minutes later we received the first page. Given most SRE were already around for a previous incident, most of us didn't need the page to be online.

2020-01-26 19:52:33     +icinga-wm      PROBLEM - MediaWiki memcached error rate on icinga1001
2020-01-26 19:53:57     +icinga-wm      PROBLEM - High average GET latency for mw requests
2020-01-26 19:55:11     +icinga-wm      PROBLEM - LVS HTTPS IPv4 #page on text-lb.codfw.wikimedia.org

Timeline

All times in UTC.

  • 19:50 Mcrouters on almost every host report failures connecting to mc1019 (mostly) and other servers. OUTAGE BEGINS
  • 19.51 Latency spikes up on the appserver cluster. Cache hit ratio for Babel requests is reduced to about 25% of normal
  • 19.52 Investigation begins
  • 19.53 Latency spikes up on the api cluster too.
  • 20.01 Cause is tentatively identified as a side effect of the news spreading and an editing stampede happening
  • 20.08 The largest part of the memcached errors are recovered, but the latencies are still very high (p75 for the appservers at precisely 10 seconds)
  • 20:13 Babel calls are identified as a possible cause.
  • 20:16 Developers are called in order to try to understand what is going on. Given it's sunday and most people are flying to All-Hands, this takes some time.
  • 20:30 Attention also moves to other pages doing cross-cluster requests like ImagePage::printSharedImageText, and to the fact we're reaching the api for babel via the caching layer
  • 20:35 Various mitigations are proposed - including setting fetchDescription => false in wmf-config/filebackend.php
  • 20:35 Kartotherian reaches 100% cpu utilization, starts intermittently paging
  • 20:41 It is noticed that some varnish-fes are constantly crashing
  • 20:43 An anomalous amount of 302s are noticed being returned from the application servers. Upon investigation, those are shown to be mostly the Apple search interface searching for Kobe Bryant.
  • 21:01 It gets decided to rollout a very aggressive timeout in MediaWiki core's HttpRequest request abstraction.
  • 21:04 Most services recover before the fix is deployed. OUTAGE ENDS
  • up to 24:00 - cpu on the maps cluster is still constantly at 100%.

Conclusions

What went well?

  • Monitoring detected the outage quickly

What went poorly?

  • We weren't able to resolve the issues or identify the origin of the problem ourselves. The issue self-resolved.
  • While we identified a few possible causes of the issue, we didn't prepare any fix before the outage was over.

Where did we get lucky?

  • We were aware of the news that caused the outage; this made the whole event less surprising
  • The outage resolved itself likely when the overall traffic waned.

How many people were involved in the remediation?

  • 8 SREs, 4 developers (2 WMF, 1 WMDE, 1 volunteer)

Actionables