Re: Transaction settings: nowait

From: Yasser Idris <luckyasser(at)gmail(dot)com>
To: pgsql-general(at)postgresql(dot)org
Subject: Re: Transaction settings: nowait
Date: 2009-05-29 03:52:56
Message-ID: 23773848.post@talk.nabble.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-general


Dude, all what u wrote make sense. Only your missing one thing, consider the
following scenario that u already reply to:

>> For. example: I forget to make commit, or rollback on exception then all
>> resources I used (updated) is locked.

>Yes - that's an application bug.

Even if it's application bug, resources shouldn't be locked indefinitely.
What if there are other clients that access this db? Forget abt this, what
if the same problem arises because of a network failure during a transaction
that issued some locks?? What's the solution then, should the resources be
locked forever till someone finds out and kills the process manually?
Although the odds of this scenario might look unlikely to happen, the
results are extremely inconvenient. That's why those features (lock timeout,
transaction time out) are there in every other dbms. It's not abt that u
don't want ur client to wait much before aborting a statement, it's more of
telling ur stupid server to abort a transaction or release a lock in case of
non-graceful abortion (without commit, rollback, or release lock) from the
client side whether it was a programmers fault or any other reason

Craig Ringer wrote:
>
> durumdara wrote:
>
>> If set wait and timeout, the Firebird is waiting for the locked resource
>> (record) for X seconds before it show deadlock error.
>>
>> But when you set no wait, the deadlock error immediately shown by the
>> server.
>
> Waiting on a lock doesn't mean there's a deadlock. A deadlock only
> occurs if two transactions are each waiting on resources the other
> transaction holds.
>
> PostgreSQL *DOES* let you control how long it waits before it checks to
> see if transactions waiting on locks might be deadlocked with another
> waiting transaction. See:
> http://www.postgresql.org/docs/current/static/runtime-config-locks.html
>
> However, it sounds like what you want is the ability to tell PostgreSQL
> that you don't want your queries to ever wait on a lock at all, even if
> it's just that you're waiting for another transaction to finish work.
>
> If that's what you mean: I'm not aware of any way in PostgreSQL to set
> lock wait timeouts at a transaction-wide level, or to set Pg to report
> an error if it would have to wait for a lock.
>
> Others seem to have wanted it enough to write patches for it, but I
> don't think they got merged. See, for example:
> http://archives.postgresql.org/pgsql-hackers/2004-06/msg00935.php
>
> (Please read the discussion on that thread to get an idea of some of the
> problems with the approach you appear to be requesting - if I'm reading
> your post correctly.).
>
> Anyway, it *IS* possible, as others have noted, to use
> SELECT ... FOR { UPDATE | SHARE } NOWAIT
> and/or
> LOCK TABLE ... NOWAIT
> before issuing work. However, this generally means you need to know what
> tuples your queries will touch, including tuples accessed by triggers
> and the like, before you issue your query.
>
> It's also really clumsy, since you often WANT queries to wait on locks
> at least briefly, otherwise you'll get intermittent errors from queries
> that're operating normally just because another query that happened to
> run concurrently briefly locked something the first query wanted.
>
> I must say that I personally can't really see the use of a
> transaction-wide lock wait timeout. Sure, most applications have to be
> prepared to re-issue queries anyway due to transient errors of one sort
> or another, and it's not unreasonable to want to be able to detect the
> difference between a query blocking on a lock, and a query that's just
> taking a long time to do work ... but I don't think that aborting the
> transaction/statement with an error is the right way to do that.
>
> First: In PostgreSQL an error from a transaction results in rollback of
> the whole transaction. This means that if you wanted to re-try after a
> lock wait timeout, you'd need to re-issue the whole transaction, or
> you'd need to use savepoints before each statement to give you
> statement-level retry. Both are clumsy and inefficient. It could also
> easily result in situations where the same group of transactions keep on
> re-trying and fighting over the same locks over and over; you'd waste a
> lot of CPU and I/O repeating work pointlessly, when if you just let the
> transaction wait on the lock everything would go just fine.
>
> In other words, it'd be ugly. I'm also not really sure how much
> difference it makes in practice WHY a statement is taking a long time,
> only that it is. Who cares whether you're waiting on a lock held by
> another transaction, or whether another transaction is using so much
> memory and disk I/O bandwidth that your query is taking forever?
>
> If you really did care about lock waits specifically, it might almost be
> nicer to be able to have the server send asynchronous NOTICE-level
> messages informing the application that the query is blocked on a lock -
> eg "NOTICE: lock_wait_timeout exceeded waiting for SHARE lock on oid
> 44123". That way the client app would know what was happening, but the
> query wouldn't get interrupted unless the app intentionally issued a
> query cancel request.
>
> Of course, that requires more sophisticated app programming and database
> interface driver use than just issuing an error on lock wait (you have
> to process messages, and you have to be prepared to do a query cancel
> request via another connection, probably from another thread), but it's
> a lot cleaner and nicer.
>
> Personally in these situations I just make sure my apps are
> multi-threaded with all database work done in a worker thread and
> controlled by posting events back and forth. That way my UI can keep the
> user informed and keep on updating while the query runs. The UI can also
> make decisions about what to do with excessively long-running queries.
> In the app I'm presently working on, for example, I pop up an infinite
> progress indicator (busy-bar) after a short delay, and I permit the user
> to cancel a long-running transaction if they don't want to keep waiting
> (after all, I have to be able to handle transaction aborts for other
> reasons anyway, so why not?).
>
> In theory, if I was worried about excessive lock wait times, I could
> even use the `pg_catalog.pg_locks' relation to figure out whether it was
> locking issues that was causing the delay. I don't care why queries take
> a while, though, only that they do.
>
> So ... personally, I don't think a lock timeout is a particularly good
> or useful idea. Basically all the use cases are handled (IMO more
> cleanly) by statement_timeout, SELECT ... FOR {SHARE|UPDATE} NOWAIT,
> LOCK TABLE ... NOWAIT, and pg_cancel_query(...) .
>
>> For. example: I forget to make commit, or rollback on exception then all
>> resources I used (updated) is locked.
>
> Yes - that's an application bug.
>
> --
> Craig Ringer
>
> --
> Sent via pgsql-general mailing list (pgsql-general(at)postgresql(dot)org)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-general
>
>

--
View this message in context: http://www.nabble.com/Transaction-settings%3A-nowait-tp23402987p23773848.html
Sent from the PostgreSQL - general mailing list archive at Nabble.com.

In response to

Responses

Browse pgsql-general by date

  From Date Subject
Next Message Scott Marlowe 2009-05-29 05:27:22 Re: Transaction settings: nowait
Previous Message Scott Bailey 2009-05-29 02:37:51 Re: composite type and domain