From June 13 ~15:10UTC to ~15:50 UTC the public WDQS endpoint in eqiad was overloaded by a bot to the point where it was not serving user queries. There is no reason to think that this bot was malicious. To mitigate this, the python-requests user agent is temporarily banned from accessing WDQS, consistent with our user agent policy.
The WDQS public endpoint in eqiad was unavailable from ~15:25 to ~15:45 UTC.
The python-requests user agent is still being banned, we are waiting to implement a more gentle solution before removing this ban.
The internal WDQS endpoint was not impacted.
Problem was detected by the Icinga LVS probe.
All times in UTC.
- 15:10: load starts to increase on the public wdqs eqiad cluster
- 15:31: Icinga LVS alert for wdqs.svc.eqiad.wmnet
- identifying and throttling bots is a hard problem
- we need to take more drastic action to protect the stability of the service (aggressively throttle generic user agents)
What went well?
- problem was detected automatically in a timely manner
- good collaboration and clear communication between
What went poorly?
- while we do have logic to throttle abusive bots, this throttling was not sufficient to protect the service
- we are still banning python-requests as a user agent, which affects a number of bots
Where did we get lucky?
- This happened during SRE offsite, when most SRE are in the same timezone. Luckily this wasn't when all of them were sleeping!
- We know that our throttling of bot is far from perfect (this is a hard problem). Some idea are already being discussed on phab:T219477
- ban logs are too verbose on disk (mitigated by https://gerrit.wikimedia.org/r/c/operations/puppet/+/516837)
- our user agent policy discourages the use of generic user agents, we should start to be more agressive in throttling generic UA or ban requests with no UA (https://gerrit.wikimedia.org/r/c/wikidata/query/rdf/+/516834)
- our own clients should follow our policy, for example our icinga checks should have a meaningful UA (https://gerrit.wikimedia.org/r/c/operations/puppet/+/517032)
- remove the ban on the python-requests UA once we have a more gently throttling solution in place