Re: 'tuple concurrently updated' error when granting permissions

Adrian Klaver <adrian(dot)klaver(at)aklaver(dot)com> writes:
> On 02/27/2014 06:54 PM, Jason Daly wrote:
>> When I attempt to execute the non-query "GRANT SELECT ON TABLE
>> <schema>.<table> TO <role>", I occasionally encounter an exception from
>> postgresql, 'tuple concurrently updated'.

> My guess this is covered here:
> http://rhaas.blogspot.com/2013/07/mvcc-catalog-access.html

While MVCC catalog scans are a necessary step towards making this kind of
thing safe, they aren't by any means sufficient. It'd be necessary to add
some kind of locking scheme if you want to avoid "tuple concurrently
updated" errors. This is not really any different from the situation
where two transactions both want to update the same row in a user table:
unless the application takes extra steps to serialize the updates, you're
going to get "tuple concurrently updated" errors.

We do have such locking for DDL on tables/indexes, but the theory in the
past has been that it's not worth the trouble for objects represented by
single catalog rows, such as functions or roles. You can't corrupt the
database with concurrent updates on such a row, you'll just get a "tuple
concurrently updated" error from all but the first-to-arrive update.
So the benefit-to-work ratio hasn't been high enough to motivate anyone
to work on it. Possibly this will change sometime in the future, but
I wouldn't hold my breath waiting.

In the meantime, you could consider using an application-managed advisory
lock if you really need such grants to work transparently.

However, might I suggest that you're doing it wrong? If you have enough
traffic on permission grants for a single table that concurrent updates
are a real hazard, it would likely be much better to invent a group role
that holds a relatively static set of rights, and implement the everyday
permissions changes by granting or revoking membership in the group role.
Not only is this using SQL permissions the way they were meant to be used,
but it should remove the concurrent-updates problem, because role
memberships are represented by distinct rows in pg_auth_members. Also,
our implementation isn't terribly efficient for cases where lots and lots
of roles have separate permissions to a single object --- I think the acl
arrays are just searched linearly, and in any case performance would
certainly go south once they got big enough to require toasting.

regards, tom lane