Author: afeldman
Description:
pt-kill is already installed on all db's via the percona-toolkit package.
configure a cron to kill queries running as wikiuser (only, long-running
queries from scripts, jobs, and crons run as wikiadmin) for more than 10 (?)
minutes.
i.e. pt-kill --busy-time 60 --match-user wikiuser --kill
On 2013-08-12 17:33:05, afeldman wrote:
i.e. pt-kill --busy-time 60 --match-user wikiuser --kill
to clarify, `--busy-time 60` is secs. so to do 10 mins it would be `--busy-time 600`
(I just read a few parts of the man page)
afeldman wrote:
On Mon Aug 12 20:50:50 2013, jeremyb wrote:
On 2013-08-12 17:33:05, afeldman wrote:
i.e. pt-kill --busy-time 60 --match-user wikiuser --kill
to clarify, --busy-time 60 is secs. so to do 10 mins it would be `--
busy-time 600`(I just read a few parts of the man page)
Thank you very much for clarifying that 10 minutes is 600 seconds and not 60!
https://bugzilla.wikimedia.org/show_bug.cgi?id=52979
According the MaxSem, the query (presumably running as wikiuser) had ben
running for...
[Time] => 140255
so this is starting to get a bit more urgent...
pt-kill is running manually on s5 slaves today, from tin, with verbose logging
to catch info for bug 52979.
As discussed with binasher on IRC, will make it a general module rather than
just a cron job. Plus it needs to report somewhere (probably to the
query_digests host I gues).
pt-kill proved to be too blunt an instrument. Currently testing
gerrit:operations/software/dbtools/arbiter.pl
The solution in place is [1]. It still isn't perfect but good enough for now and doesn't require an external service that can be blocked by max_connections.