Deadlock between UI and importer

[Jeroen T. Vermeulen]

On 2008-12-23 we observed a deadlock between the importer and the translation UI, leading to an error from the import script and two timeout reports from the app servers. (The affected UI user reported retrying once when the server took too long to respond).

The deadlock occurred at the end of a template import, when the importer was recomputing statistics for the template's translations. The UI user was editing one of these translations at the time. The deadlock and timeouts were all reported with the same timestamp, which was apparently the moment the database detected and broke the deadlock.

Deadlock traceback: https://pastebin.canonical.com/12314/

Recomputing the statistics for all of a template's translations does not take much time, but it covers a lot of data. We can reduce the risk by moving this work into a separate transaction, or even one transaction per translation.

[Данило Шеган (danilo)]

FWIW, it might be interested to track if this behaviour is also the cause of occassional PO file stats skew in our update-stats cronjob.

[Jeroen T. Vermeulen (jtv)]

It's one possible reason; simple concurrency would be another. In theory, concurrent SERIALIZABLE transactions would run in a near-perfect illusion of being the only transaction currently executing. In practice, that's expensive to achieve so postgres approximates it. (It's recently been discovered that this approximation doesn't quite meet the standard).

That means that there may well be some cracks in this case that occasional inconsistencies can sneak through. As long as those are few and small, and the cron job catches them anyway, trying to fix that is probably not worth the effort.

[Jeroen T. Vermeulen (jtv)]

It's an easy fix, so might as well deal with it.

[Jeroen T. Vermeulen (jtv)]

Just happened again: http://launchpadlibrarian.net/22404413/x9AeJEyaFcnj4kzCoDkDtSvzWjE.txt

The fix has landed, but not rolled out yet.

[Jeroen T. Vermeulen (jtv)]

Happening really frequently now. Fix to roll out next week.