Mark Dilger <hornschnorter@gmail.com> writes:
> On 11/23/19 8:34 AM, Tom Lane wrote:
>> It suddenly strikes me to worry that we have an XID wraparound hazard
>> for entries in the notify queue.
> Is it worth checking for this condition in autovacuum?
Dunno, maybe. It's a different avenue to consider, at least.
> There shouldn't be too much reason to back-patch any of this, since
> the change in 51004c717 only applies to v13 and onward. Or do you
> see the risk you described as "pretty minimal" as still being large
> enough to outweigh the risk of anything we might back-patch?
There may not be a risk large enough to worry about before 51004c717,
assuming that we discount cases like a single session staying
idle-in-transaction for long enough for the XID counter to wrap
(which'd cause problems for more than just LISTEN/NOTIFY). I haven't
analyzed this carefully enough to be sure. We'd have to consider
that, as well as the complexity of whatever fix we choose for HEAD,
while deciding if we need a back-patch.
regards, tom lane