Horton

Re: Old row version in hot chain become visible after a freeze

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Alvaro Herrera <alvherre@2ndquadrant.com>
Cc: "Wong, Yi Wen" <yiwong@amazon.com>, Michael Paquier <michael.paquier@gmail.com>, Peter Geoghegan <pg@bowt.ie>, "Wood, Dan" <hexpert@amazon.com>, "pgsql-bugs@postgresql.org" <pgsql-bugs@postgresql.org>
Date: 2017-09-08T17:04:10Z
Lists: pgsql-bugs 
Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> Wong, Yi Wen wrote:
>> lazy_record_dead_tuple may fail to record the heap for later pruning
>> for lazy_vacuum_heap if there is already a sufficiently large number of dead tuples
>> in the array:

> Hmm, ouch, good catch.

> AFAICS this is a shouldn't-happen condition, since we bail out of the
> loop pessimistically as soon as we would be over the array limit if the
> next page were to be full of dead tuples (i.e., we never give the chance
> for overflow to actually happen).  So unless I misunderstand, this could
> only fail if you give maint_work_mem smaller than necessary for one
> pageful of dead tuples, which should be about 1800 bytes ...

> If we wanted to be very sure about this we could add a test and perhaps
> abort the vacuum, but I'm not sure it's worth the trouble.

I think if we're going to depend on that, we should change the logic from
"don't record tuple if no space" to "throw error on no space".

			regards, tom lane

Commits

  1. Fix freezing of a dead HOT-updated tuple

  2. Fix potential data corruption during freeze

  3. Clarify the contract of partition_rbound_cmp().