Jump to content

Incidents/2019-02-13 PAWS

From Wikitech
Tracked in Phabricator
Task T216208

Summary

The PAWS service was offline for several hours due to issues connecting to tools.db.svc.eqiad.wmflabs, related to the database being in the limit of usage ( T215993: tools.db.svc.eqiad.wmflabs hitting it's limit?).

See also the following day's related incident: 20190214-labsdb1005

Timeline

All times are in UTC.

  • 03:38 first icinga alert about PAWS being down.
  • 04:26 tools.db.svc.eqiad.wmflabs issues reported in T215993: tools.db.svc.eqiad.wmflabs hitting it's limit?.
  • 04:33 Bryan acknoledges the database issue.
  • 04:43 icinga recovery for PAWS (flapping?).
  • 05:38 second icinga alert about PAWS being down.
  • 05:43 icinga recovery for PAWS (flapping?).
  • 06:41 third icinga alert about PAWS being down.
  • 06:42 Brooke and Arturo start looking into the PAWS issue. Very late/early for both.
  • 07:27 detect an issue with the database (connection refused), didn't pay enough attention since it appeared transient in the logs at the time.
  • 07:47 we start chasing issues regarding networking/proxying related to kubernetes since some parts of that had stopped working
  • 08:37 Giovanni joins the debugging team.
  • 08:45 Brooke goes to sleep, very late in her timezone. Arturo breakfast.
  • 09:28 Giovanni realizes the database issue could be the root issue. Restarts the database.
  • 09:42 icinga recovery for PAWS.

Conclusions

The PAWS setup is really complex from the user-to-server path, lots of small spots to check & debug, at least 3 or 4 proxies are involved.
Also, these proxies are in different CloudVPS projects (project-proxy, paws, and tools).

We overlooked the issue in tools.db.svc.eqiad.wmflabs, probably due to being very late for Brooke and very early for Arturo.
The database issue itself is not weird (capacity saturation) given increased usage of our platforms.

It took us a lot of time to start debugging things properly because PAWS is poorly documented from the administration/design/architecture/operation point of view. Since nobody in the WMF-WMCS team is really involved in mantaining PAWS, is not known which bits are important, what are the common things to check, etc. We weren't familiar with the setup.

In general, PAWS is in a weird state. It's a service that WMF/WMCS wants to offer to the Wikimedia Movement, but we don't have enough human capacity to fully support it.
The truth is that the project is mostly community-maintained, which is probably not enough for the importance of the service within the movement.
It's weird that we get paged at odd hours to fix a service that we aren't really supporting.

We have a basic documentation page which clearly is not enough:

Actionables

Retrieved from "https://wikitech.wikimedia.org/w/index.php?title=Incidents/2019-02-13_PAWS&oldid=1966293"