"Zeugswetter Andreas SB SD" <[EMAIL PROTECTED]> writes:
>> * Is it really a good idea for database-wide ANALYZE to run as a single
>> transaction? Holding all those locks is a recipe for deadlocks, even
>> if they're as inoffensive as AccessShareLocks normally are.
> Wasn't one idea behind that change also to not make the planner create a plan
> from mixed old and new statistics ?
I don't recall that that was part of the discussion. IIRC all we were
after was to let someone invoke ANALYZE from inside a BEGIN/COMMIT block.
A possible compromise is to hack ANALYZE so that if it is invoked when
*not* within a BEGIN block, it runs a separate transaction for each
table. This seems pretty crufty but might satisfy all the requirements.
> I guess that could later be accomplished with
> "begin work; analyze; commit work;" (with subtransactions) though.
AFAICS, locks taken by a (committed) subtransaction can't be released
until top-level commit anyhow. Otherwise they fail to perform one of
the essential functions of locking in an MVCC environment: to delay
another process until the changes you've made are visible to him.
regards, tom lane
---------------------------(end of broadcast)---------------------------
TIP 5: Have you checked our extensive FAQ?
http://www.postgresql.org/docs/faqs/FAQ.html
