You are browsing a read-only backup copy of Wikitech. The live site can be found at wikitech.wikimedia.org

Incidents/2022-07-20 network interruption: Difference between revisions

From Wikitech-static
Jump to navigation Jump to search
imported>RLazarus
No edit summary
 
imported>LSobanski
Line 67: Line 67:
! rowspan="5" |Process
! rowspan="5" |Process
|Was the incident status section actively updated during the incident?
|Was the incident status section actively updated during the incident?
|
|no
|Only one SRE responded during the incident, so the incident doc was created afterward, to organize the timeline and followup items.
|Only one SRE responded during the incident, so the incident doc was created afterward, to organize the timeline and followup items.
|-
|-
Line 79: Line 79:
|-
|-
|Are the documented action items assigned?
|Are the documented action items assigned?
|
|no
|
|
|-
|-
Line 109: Line 109:
|-
|-
! colspan="2" align="right" |Total score (count of all “yes” answers above)
! colspan="2" align="right" |Total score (count of all “yes” answers above)
|
|7
|
|
|}
|}

Revision as of 11:52, 27 July 2022

document status: in-review

Summary

Incident metadata (see Incident Scorecard)
Incident ID 2022-07-20 network interruption Start 2022-07-20 03:16
Task T313382 End 2022-07-20 03:25 (mostly)
People paged 11 Responder count 3
Coordinators rzl Affected metrics/SLOs No relevant SLOs exist. See impact metrics below.
Impact The network was partitioned for 6 minutes.
  • About 1.2 million requests are missing from the appservers (app and API combined) during 16 minutes, or 26% of expected uncached traffic.
  • About 970,000 requests are missing from Varnish-text during 11 minutes, or 0.2% of total traffic.
  • 5XXs at both levels were negligible.
  • Phabricator was unavailable for 32 minutes.
  • The Kubernetes API was at least partially unavailable for 52 minutes, but during a period where no control operations are normally in progress.

At 03:16, the top-of-rack switch asw2-c-eqiad virtual chassis lost connectivity to FPC5, partitioning the network. This caused a hard down event for all hosts in rack C5 (netbox). It also caused additional instability due to how the virtual chassis works, and because it's incorrectly cabled up.

We received a burst of both paging and non-paging alerts: Icinga reporting hosts down; BGP status; application-level errors; and MariaDB replica alerts. At least one user also reported via IRC that they couldn't access metawiki (almost certainly uncacheable traffic, due to logged-in state).

At 03:22, asw2-c-eqiad:fpc5 came back online. Most systems recovered automatically, but some needed manual attention:

  • We received HAProxy failover alerts on dbproxy1018 through 1021, and those needed to be resolved by reloading haproxy manually, as expected.
  • Phabricator's dbproxy had failed over to a read-only replica (as expected) but Phabricator was unavailable for read-only tasks in read-only mode. When users attempted to view a task, they got an error page saying, Unhandled Exception ("AphrontQuery Exception") #1290: The MariaDB server is running with the --read-only option so it cannot execute this statement This was resolved by reloading haproxy, but Phab was expected to be available for reads.
  • The Kubernetes API server alerted for high latency until kube-apiserver was manually restarted on both hosts.

Actionables

  • T313384 Recable eqiad row C switch fabric, so that in the future a failure like this will only impact servers in rack C5.
  • T313382#8090176 Move critical hosts, like DB masters, away from rack C5 until its top-of-rack switch is trustworthy.
  • Yes Done T313382#8090224 Add LibreNMS alerting (and runbook) for this scenario, which will speed up troubleshooting.
  • T313879 Make read-only Phabricator operations possible when its database is in read-only mode.

Scorecard

Incident Engagement™ ScoreCard
Question Answer

(yes/no)

Notes
People Were the people responding to this incident sufficiently different than the previous five incidents? yes
Were the people who responded prepared enough to respond effectively no The IC was able to respond effectively to the downstream failures (DB, appservers, Phab, k8s, etc) but wasn't able to identify the root cause or troubleshoot in LibreNMS effectively due to lack of familiarity.
Were fewer than five people paged? no
Were pages routed to the correct sub-team(s)? no
Were pages routed to online (business hours) engineers? Answer “no” if engineers were paged after business hours. no
Process Was the incident status section actively updated during the incident? no Only one SRE responded during the incident, so the incident doc was created afterward, to organize the timeline and followup items.
Was the public status page updated? no Not justified given the impact
Is there a phabricator task for the incident? yes
Are the documented action items assigned? no
Is this incident sufficiently different from earlier incidents so as not to be a repeat occurrence? yes
Tooling To the best of your knowledge was the open task queue free of any tasks that would have prevented this incident? Answer “no” if there are

open tasks that would prevent this incident or make mitigation easier if implemented.

yes
Were the people responding able to communicate effectively during the incident with the existing tooling? yes
Did existing monitoring notify the initial responders? yes
Were the engineering tools that were to be used during the incident, available and in service? no Reading Phab tasks for context was impossible due to its being unavailable in RO mode
Were the steps taken to mitigate guided by an existing runbook? yes
Total score (count of all “yes” answers above) 7