On Thu, Oct 2, 2014 at 5:37 PM, Peter Geoghegan <pg@heroku.com> wrote:
> Session 3 is an innocent bystander. It goes to query the same table in
> an ordinary, routine way - a SELECT statement. Even though session 2's
> lock is not granted yet, session 3 is not at liberty to skip the queue
> and get its own AccessShareLock. The effect is about the same as if
> session 2 did need to hold an AccessExclusiveLock for ages: read
> queries block for a long time. And yet, in theory session 2's impact
> on production should not be minimal, if we consider something like
> EXPLAIN output.
The explain would show the AccessExclusiveLock, so it would be enough
for a heads-up to kill all idle-in-transaction holding locks on the
target relation (if killable, or just wait).
Granted, it's something that's not easily automatable, whereas a nowait is.
However, rather than nowait, I'd prefer "cancellable" semantics, that
would cancel voluntarily if any other transaction requests a
conflicting lock, like autovacuum does.
