I'm seeing these in the log:
2014-07-07 14:05:42 mw1015 commonswiki: Update for doc ids: 15731483; error message was: No enabled connection
I don't imagine we're actually out of connections - probably just hitting some other http error and eating it. We should not eat it.
Version: unspecified
Severity: normal
This should only happen after we've tried getMaxConnectionAttempts() times. We never call getConnection() directly, so it should be handled by our callback.
Getting a full stacktrace out of this would help.
Yeah. I've seen this in a maintenance script when I tried to bulk insert too much data. I _think_ it has something to do with commands being too too too big getting interpreted as a retry-able http error. We actually have logic to resubmit the command as singletons but I think we don't get to use it because we've marked all the connections as busted due to the error. Might try to reproduce with stupid huge page.
Change 156803 had a related patch set uploaded by Manybubbles:
Increase timeout on updates
https://gerrit.wikimedia.org/r/#/c/156798/
and
https://gerrit.wikimedia.org/r/#/c/156800/
as well.
Its not stupid huge pages. Its massive influxes of updates all at once, I believe.
Anyway, these commits should make it more stable as well as give us more information when it fails.