Re: BUG #12990: Missing pg_multixact/members files (appears to have wrapped, then truncated)

On Tue, Apr 21, 2015 at 5:12 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> On Tue, Apr 21, 2015 at 12:34 AM, Alvaro Herrera <alvherre(at)2ndquadrant(dot)com>
> wrote:
>>
>> Alvaro Herrera wrote:
>>
>> > The fix is to raise an ERROR when generating a new multixact, if we
>> > detect that doing so would get close to the oldest multixact that the
>> > system knows about. If that happens, the solution is to vacuum so that
>> > the "oldest" point is advanced a bit more and you have room to generate
>> > more multixacts. In production, you would typically adjust the
>> > multixact freeze parameters so that "oldest multixact" is advanced more
>> > aggressively and you don't hit the ERROR.
>>
>> Here's a patch. I have tested locally and it closes the issue for me.
>> If those affected can confirm that it stops the file removal from
>> happening, I'd appreciate it.
>>
>
> 1. Do you think it makes sense to give warning in SetMultiXactIdLimit()
> if we have already reached offsetWarnLimit as we give for multiWarnLimit?

Amit and I discussed this offline. Yes, we could include a warning
message here, for consistency with the warnings you get about xid
wraparound. Concretely I think it means that you would also get
warnings about being being near the member space limit from vacuums,
rather than just from attempts to allocate new multixact IDs. The
test to detect an impending members-would-wrap ERROR would be similar
to what we do in GetNewMultiXactId, so something like:

MultiXactOffsetWouldWrap(offsetStopLimit,
nextOffset,
MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT *
OFFSET_WARN_SEGMENTS)

I'm not sure whether it's worth writing an extra patch for this
though, because if you're in this situation, your logs are already
overflowing with warnings from the regular backends that are
generating multixacts. Thoughts anyone?

--
Thomas Munro
http://www.enterprisedb.com