Peter Geoghegan <pg@heroku.com> writes:
> On Sun, Jan 3, 2016 at 4:32 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> If we fix this, I believe we could also remove the weasel wording that was
>> added to create_policy.sgml in commit 43cd468cf01007f3 about how the
>> system might transiently fail to enforce row security correctly.
> IIUC, then what you say here isn't true, because that issue was about
> a transient failure without the involvement of *any* DDL from start to
> finish. CREATE POLICY accepts subqueries referencing other relations
> in its USING quals. This seems to be idiomatic usage of CREATE POLICY,
> in fact.
> See my original isolation tester test case, where only the setup involves DDL:
> http://www.postgresql.org/message-id/attachment/38467/0001-RLS-isolation-test.patch
Hmm. I agree that this test case's behavior does not depend on CREATE
POLICY's lock mismanagement. I think what is going on here is that the
RLS quals are being checked with an older snapshot than what controls
the output of the UPDATE RETURNING. Even the EPQ recheck that's done
after getting update lock on the "information" row doesn't fix it,
because we *don't* insist on taking an update lock on the "users" table,
so we don't see the new value of that row.
If that diagnosis is correct, you could fix it by changing the RLS
policies' sub-selects to use SELECT FOR UPDATE, though the loss of
concurrency might well be unacceptable.
In any case, the text in create_policy.sgml seems to be a misleading
description of the problem, as it's talking about DDL modifications
which are *not* what's happening here.
regards, tom lane