Re: PANIC killing vacuum process

"Kevin Grittner" <Kevin(dot)Grittner(at)wicourts(dot)gov> writes:
> Silvio Brandani <silvio(dot)brandani(at)tech(dot)sdb(dot)it> wrote:
>> we have develop a script to execute the vacuum full on all tables
>> of our very big database , since it is a 24 x 7 available system
>> we have not a timeframe to exec the vacuum full.
>> so we try with this script running the vauum full table by table
>> and if the vacuum generate the waiting status for other
>> connections we kill the vacuum .
>> [ and get ]
>> 2010-11-03 14:25:27 CET [19324]: [6-1] PANIC: cannot abort
>> transaction 75073917, it was already committed

> What version of PostgreSQL is this?

Anything pre-9.0 might do that; it's one of the known problems with the
old VACUUM FULL implementation that made us decide to get rid of it.

However, versions that are less than about a year old do have a hack
that should avoid the PANIC for a query-cancel interrupt ... at the
cost of ignoring the cancel interrupt, so that's not all that helpful
a solution here.

>> Is it possible to softly kill a vacuum process without risk a
>> panic ?????

> Normally, yes. VACUUM FULL is more prone to problems than a normal
> vacuum, especially if you are using an old version. There are very
> few circumstances where VACUUM FULL is the right thing to use.

Indeed. If you think you need to use VACUUM FULL on a routine basis,
rethink that.

regards, tom lane