Обсуждение: SELECT ... FOR UPDATE performance costs? alternatives?
All,
I have a stored procedure that I use to manage a queue. I want to pop
an item off the queue to ensure that only one server is processing the
queue item, so inside PGSQL, use SELECT ... FOR UPDATE to lock the row.
Here's how I pop the queue item:
----------------------------------------------------------------------
CREATE OR REPLACE FUNCTION "public"."reserve_next_tcqueue" (in_hostname
varchar, in_status char, in_new_status char) RETURNS bigint AS
$body$
DECLARE
my_reserved_id BIGINT;
BEGIN
/* find and lock a row with the indicated status */
SELECT tcq_id
INTO my_reserved_id
FROM queue q
WHERE q.status = in_status
ORDER BY tcq_id ASC
LIMIT 1
FOR UPDATE;
/* we didn't find anything matching */
IF NOT FOUND THEN
RETURN 0;
END IF;
/* change the status to the new status */
UPDATE queue SET
status = in_new_status,
ts_start = NOW(),
ts_end = NULL,
hostname = COALESCE(in_hostname, hostname)
WHERE tcq_id = my_reserved_id;
/* send back our reserved ID */
RETURN my_reserved_id;
END;
$body$
LANGUAGE 'plpgsql' VOLATILE RETURNS NULL ON NULL INPUT SECURITY INVOKER;
----------------------------------------------------------------------
When my server is under severe load, however, this function begins to
take a long time to execute and I begin to suspect that the FOR UPDATE
lock might be locking the whole table and not just the row.
How do I know if this is the case, how can I optimize this procedure,
and how should I be doing this differently? I'm guessing stacks and
queues would be common patterns handled in the PostgreSQL community.
Thoughts?
-- Dante
On Aug 15, 2007, at 1:09 PM, D. Dante Lorenso wrote: > All, > > I have a stored procedure that I use to manage a queue. I want to > pop an item off the queue to ensure that only one server is > processing the queue item, so inside PGSQL, use SELECT ... FOR > UPDATE to lock the row. Here's how I pop the queue item: > > ---------------------------------------------------------------------- > CREATE OR REPLACE FUNCTION > "public"."reserve_next_tcqueue" (in_hostname varchar, in_status > char, in_new_status char) RETURNS bigint AS > $body$ > DECLARE > my_reserved_id BIGINT; > BEGIN > /* find and lock a row with the indicated status */ > SELECT tcq_id > INTO my_reserved_id > FROM queue q > WHERE q.status = in_status > ORDER BY tcq_id ASC > LIMIT 1 > FOR UPDATE; > > /* we didn't find anything matching */ > IF NOT FOUND THEN > RETURN 0; > END IF; > > /* change the status to the new status */ > UPDATE queue SET > status = in_new_status, > ts_start = NOW(), > ts_end = NULL, > hostname = COALESCE(in_hostname, hostname) > WHERE tcq_id = my_reserved_id; > > /* send back our reserved ID */ > RETURN my_reserved_id; > END; > $body$ > LANGUAGE 'plpgsql' VOLATILE RETURNS NULL ON NULL INPUT SECURITY > INVOKER; > ---------------------------------------------------------------------- > > When my server is under severe load, however, this function begins > to take a long time to execute and I begin to suspect that the FOR > UPDATE lock might be locking the whole table and not just the row. > > How do I know if this is the case, how can I optimize this > procedure, and how should I be doing this differently? I'm > guessing stacks and queues would be common patterns handled in the > PostgreSQL community. > > Thoughts? SELECT ... FOR UPDATE should only be locking the rows returned by your the select statement, in this case the one row. You can check what locks exist on a table (and their type) with the pg_locks system view. Erik Jones Software Developer | Emma® erik@myemma.com 800.595.4401 or 615.292.5888 615.292.0777 (fax) Emma helps organizations everywhere communicate & market in style. Visit us online at http://www.myemma.com
Erik Jones wrote:
> On Aug 15, 2007, at 1:09 PM, D. Dante Lorenso wrote:
>
>> ...to ensure that only one server is processing the queue item, so
>> inside PGSQL, use SELECT ... FOR UPDATE to lock the row...
>>
>> When my server is under severe load, however, this function begins to
>> take a long time to execute and I begin to suspect that the FOR
>> UPDATE lock might be locking the whole table and not just the row.
>> How do I know if this is the case, how can I optimize this procedure,
>> and how should I be doing this differently? ...
>>
>> Thoughts?
>
> SELECT ... FOR UPDATE should only be locking the rows returned by your
> the select statement, in this case the one row. You can check what
> locks exist on a table (and their type) with the pg_locks system view.
>
Is that correct? Documentation section 12.3.1. Table-Level Locks states
'The list below shows the available lock modes ...Remember that all of
these lock modes are table-level locks, even if the name contains the
word "row"'.
I wonder why bother with the SELECT statement at all. Why not just go
straight to the UPDATE statement with something like
UPDATE queue SET
status = in_new_status,
ts_start = NOW(),
ts_end = NULL,
hostname = COALESCE(in_hostname, hostname)
WHERE tcq_id = (SELECT tcq_id FROM queue q WHERE q.status =
in_status ORDER BY tcq_id ASC LIMIT 1);
He may need to trap an exception for the "not found" case, but what's
the big deal with that?
UPDATE statements acquire a ROW EXCLUSIVE on the table, which conflicts,
among other things, with ROW EXCLUSIVE, so it will block other UPDATE
statements initiated by other transactions.
On Aug 15, 2007, at 2:39 PM, btober@ct.metrocast.net wrote: > Erik Jones wrote: >> On Aug 15, 2007, at 1:09 PM, D. Dante Lorenso wrote: >> >>> ...to ensure that only one server is processing the queue item, >>> so inside PGSQL, use SELECT ... FOR UPDATE to lock the row... >>> When my server is under severe load, however, this function >>> begins to take a long time to execute and I begin to suspect that >>> the FOR UPDATE lock might be locking the whole table and not just >>> the row. How do I know if this is the case, how can I optimize >>> this procedure, and how should I be doing this differently? ... >>> >>> Thoughts? >> >> SELECT ... FOR UPDATE should only be locking the rows returned by >> your the select statement, in this case the one row. You can >> check what locks exist on a table (and their type) with the >> pg_locks system view. >> > > Is that correct? Documentation section 12.3.1. Table-Level Locks > states 'The list below shows the available lock modes ...Remember > that all of these lock modes are table-level locks, even if the > name contains the word "row"'. You will notice that SELECT ... FOR UPDATE is not in that list. It's covered in the next section on row level locks. > > I wonder why bother with the SELECT statement at all. Why not just > go straight to the UPDATE statement with something like > > UPDATE queue SET > status = in_new_status, > ts_start = NOW(), > ts_end = NULL, > hostname = COALESCE(in_hostname, hostname) > WHERE tcq_id = (SELECT tcq_id FROM queue q WHERE q.status = > in_status ORDER BY tcq_id ASC LIMIT 1); > > He may need to trap an exception for the "not found" case, but > what's the big deal with that? > > UPDATE statements acquire a ROW EXCLUSIVE on the table, which > conflicts, among other things, with ROW EXCLUSIVE, so it will block > other UPDATE statements initiated by other transactions. That won't work because the update won't lock the row until the select returns. So, if two process execute that at the same time they will both execute the subquery and return the same result, the first will update it and the second will then (redundantly) update it. Erik Jones Software Developer | Emma® erik@myemma.com 800.595.4401 or 615.292.5888 615.292.0777 (fax) Emma helps organizations everywhere communicate & market in style. Visit us online at http://www.myemma.com
Erik Jones wrote: > > On Aug 15, 2007, at 2:39 PM, btober@ct.metrocast.net wrote: > >> Erik Jones wrote: >>> On Aug 15, 2007, at 1:09 PM, D. Dante Lorenso wrote: >>> >>>> ...to ensure that only one server is processing the queue item, so >>>> inside PGSQL, use SELECT ... FOR UPDATE to lock the row... >>>> When my server is under severe load, however, this function begins >>>> to take a long time to execute and I begin to suspect that the FOR >>>> UPDATE lock might be locking the whole table and not just the row. >>>> How do I know if this is the case, how can I optimize this >>>> procedure, and how should I be doing this differently? ... >>>> >>>> Thoughts? >>> >>> SELECT ... FOR UPDATE should only be locking the rows returned by >>> your the select statement, in this case the one row. You can check >>> what locks exist on a table (and their type) with the pg_locks system >>> view. >>> >> >> Is that correct? Documentation section 12.3.1. Table-Level Locks >> states 'The list below shows the available lock modes ...Remember that >> all of these lock modes are table-level locks, even if the name >> contains the word "row"'. > > You will notice that SELECT ... FOR UPDATE is not in that list. It's > covered in the next section on row level locks. >> >> I wonder why bother with the SELECT statement at all. Why not just go >> straight to the UPDATE statement with something like >> >> UPDATE queue SET >> status = in_new_status, >> ts_start = NOW(), >> ts_end = NULL, >> hostname = COALESCE(in_hostname, hostname) >> WHERE tcq_id = (SELECT tcq_id FROM queue q WHERE q.status = >> in_status ORDER BY tcq_id ASC LIMIT 1); >> >> He may need to trap an exception for the "not found" case, but what's >> the big deal with that? >> >> UPDATE statements acquire a ROW EXCLUSIVE on the table, which >> conflicts, among other things, with ROW EXCLUSIVE, so it will block >> other UPDATE statements initiated by other transactions. > > That won't work because the update won't lock the row until the select > returns. So, if two process execute that at the same time they will > both execute the subquery and return the same result, the first will > update it and the second will then (redundantly) update it. It also won't work because I need to change AND read the row. If I only do the update, I don't know what was updated. I still need to return the tcq_id to my application. Maybe the update could look like this: UPDATE queue SET status = in_new_status, ts_start = NOW(), ts_end = NULL, hostname = COALESCE(in_hostname, hostname), WHERE status = in_status; But there I don't have the LIMIT 1, and I also don't know which rows got updated. I supposed there might be some magic to find the OID of the affected rows, but I don't know how what would be done. I need logic like "atomic test and set" or pop 1 item off the queue atomically and tell me what that item was. In my situation, there are a dozen or so machines polling this queue periodically looking for work to do. As more polling is occurring, the locks seem to be taking longer so I was worried table-level locks might be occurring. -- Dante
"D. Dante Lorenso" <dante@lorenso.com> writes:
> ... When my server is under severe load, however, this function begins to
> take a long time to execute and I begin to suspect that the FOR UPDATE
> lock might be locking the whole table and not just the row.
FOR UPDATE only locks the selected row(s). However, the way you've got
this coded, every incoming server will try to select the same row, which
means that whichever one gets there first will block all the others
until it commits (and releases its row lock). Are you careful to commit
immediately after making this update?
regards, tom lane
"D. Dante Lorenso" <dante@lorenso.com> writes: > I need logic like "atomic test and set" or pop 1 item off the queue > atomically and tell me what that item was. > > In my situation, there are a dozen or so machines polling this queue > periodically looking for work to do. As more polling is occurring, > the locks seem to be taking longer so I was worried table-level locks > might be occurring. How quickly after you update the row status are you comitting (and releasing locks)? I have apps that basically do: SELECT id FROM job_table WHERE status = 'New' FOR UPDATE; UPDATE job_table SET status = 'Processing' WHERE id IN (<set of IDs>); COMMIT; -- releases all locks <process each job in the list we got and update its status> This has worked very well for me. -Doug
On Aug 15, 2007, at 9:21 PM, D. Dante Lorenso wrote: > Erik Jones wrote: >> On Aug 15, 2007, at 2:39 PM, btober@ct.metrocast.net wrote: >>> Erik Jones wrote: >>>> On Aug 15, 2007, at 1:09 PM, D. Dante Lorenso wrote: >>>> >>>>> ...to ensure that only one server is processing the queue item, >>>>> so inside PGSQL, use SELECT ... FOR UPDATE to lock the row... >>>>> When my server is under severe load, however, this function >>>>> begins to take a long time to execute and I begin to suspect >>>>> that the FOR UPDATE lock might be locking the whole table and >>>>> not just the row. How do I know if this is the case, how can I >>>>> optimize this procedure, and how should I be doing this >>>>> differently? ... >>>>> >>>>> Thoughts? >>>> >>>> SELECT ... FOR UPDATE should only be locking the rows returned >>>> by your the select statement, in this case the one row. You can >>>> check what locks exist on a table (and their type) with the >>>> pg_locks system view. >>>> >>> >>> Is that correct? Documentation section 12.3.1. Table-Level Locks >>> states 'The list below shows the available lock modes ...Remember >>> that all of these lock modes are table-level locks, even if the >>> name contains the word "row"'. >> You will notice that SELECT ... FOR UPDATE is not in that list. >> It's covered in the next section on row level locks. >>> >>> I wonder why bother with the SELECT statement at all. Why not >>> just go straight to the UPDATE statement with something like >>> >>> UPDATE queue SET >>> status = in_new_status, >>> ts_start = NOW(), >>> ts_end = NULL, >>> hostname = COALESCE(in_hostname, hostname) >>> WHERE tcq_id = (SELECT tcq_id FROM queue q WHERE q.status = >>> in_status ORDER BY tcq_id ASC LIMIT 1); >>> >>> He may need to trap an exception for the "not found" case, but >>> what's the big deal with that? >>> >>> UPDATE statements acquire a ROW EXCLUSIVE on the table, which >>> conflicts, among other things, with ROW EXCLUSIVE, so it will >>> block other UPDATE statements initiated by other transactions. >> That won't work because the update won't lock the row until the >> select returns. So, if two process execute that at the same time >> they will both execute the subquery and return the same result, >> the first will update it and the second will then (redundantly) >> update it. > > It also won't work because I need to change AND read the row. If I > only do the update, I don't know what was updated. I still need to > return the tcq_id to my application. > > Maybe the update could look like this: > > UPDATE queue SET > status = in_new_status, > ts_start = NOW(), > ts_end = NULL, > hostname = COALESCE(in_hostname, hostname), > WHERE status = in_status; > > But there I don't have the LIMIT 1, and I also don't know which > rows got updated. I supposed there might be some magic to find the > OID of the affected rows, but I don't know how what would be done. > > I need logic like "atomic test and set" or pop 1 item off the queue > atomically and tell me what that item was. If you're using 8.2.x there is the RETURNING clause that can be used with UPDATE and INSERT queries. So, you could alter his version to add the FOR UPDATE to the subquery and tack on the RETURNING clause to the UPDATE and you'd have semantically equivalent. > > In my situation, there are a dozen or so machines polling this > queue periodically looking for work to do. As more polling is > occurring, the locks seem to be taking longer so I was worried > table-level locks might be occurring. As I said, watch pg_locks. Also, make sure that this function call isn't happening inside a larger transaction that's not committing right away when the function returns. The lock from the SELECT ... FOR UPDATE isn't released until the enclosing transaction commits or rolls back. Erik Jones Software Developer | Emma® erik@myemma.com 800.595.4401 or 615.292.5888 615.292.0777 (fax) Emma helps organizations everywhere communicate & market in style. Visit us online at http://www.myemma.com
Douglas McNaught wrote:
> "D. Dante Lorenso" <dante@lorenso.com> writes:
> How quickly after you update the row status are you comitting (and
> releasing locks)?
I am calling a stored proc from PHP. Since I do not begin a
transaction, I assume that my call is automatically committed
immediately after invocation.
SELECT reserve_next_tcqueue(?, ?, ?) AS result
> SELECT id FROM job_table WHERE status = 'New' FOR UPDATE;
> UPDATE job_table SET status = 'Processing' WHERE id IN (<set of IDs>);
> COMMIT; -- releases all locks
> <process each job in the list we got and update its status>
> This has worked very well for me.
Yes, this does work well for me also most of the time. It is only when
the database server begins to suffer from severe load (like 3+) that
PostgreSQL begins to log the reserve_next_tcqueue(...) queries as taking
a long time to complete. Here are some examples:
...
Aug 13 16:00:42 shed03 postgres[20264]: [5-2] reserve_next_tcqueue($1,
$2, $3) AS tcq_id]
Aug 13 16:00:53 shed03 postgres[17338]: [5-1] 17338 dbxxx 10.10.20.163
LOG: duration: 3159.208 ms statement: EXECUTE <unnamed> [PREPARE: SELECT
Aug 13 16:00:54 shed03 postgres[20447]: [5-2] reserve_next_tcqueue($1,
$2, $3) AS tcq_id]
Aug 13 16:00:54 shed03 postgres[20470]: [5-1] 20470 dbxxx 10.10.20.51
LOG: duration: 4162.031 ms statement: EXECUTE <unnamed> [PREPARE: SELECT
Aug 13 16:00:54 shed03 postgres[20470]: [5-2] reserve_next_tcqueue($1,
$2, $3) AS tcq_id]
Aug 13 16:00:59 shed03 postgres[20530]: [5-1] 20530 dbxxx 10.10.20.51
LOG: duration: 3672.077 ms statement: EXECUTE <unnamed> [PREPARE: SELECT
...
-- Dante
"D. Dante Lorenso" <dante@lorenso.com> writes:
> Douglas McNaught wrote:
>> How quickly after you update the row status are you comitting (and
>> releasing locks)?
> I am calling a stored proc from PHP. Since I do not begin a
> transaction, I assume that my call is automatically committed
> immediately after invocation.
Have you actually verified that, or are you just assuming it?
I believe that PHP has some sort of autocommit on/off option,
which might possibly be doing things behind your back.
regards, tom lane
Tom Lane wrote: > "D. Dante Lorenso" <dante@lorenso.com> writes: >> Douglas McNaught wrote: >>> How quickly after you update the row status are you comitting (and >>> releasing locks)? > >> I am calling a stored proc from PHP. Since I do not begin a >> transaction, I assume that my call is automatically committed >> immediately after invocation. > > Have you actually verified that, or are you just assuming it? Just assuming. I'm not really sure HOW to verify it, though. > I believe that PHP has some sort of autocommit on/off option, > which might possibly be doing things behind your back. I am using PHP / PDO and all my statements are prepared. Sometimes I will begin a transaction using PDO and do either commit or rollback, but I don't explicitly use transactions if I don't intend to exec more than one statement. Here is what PHP says about auto-commit in PDO: -------- 8< ---------------- 8< ---------------- 8< -------- http://www.php.net/manual/en/ref.pdo.php Unfortunately, not every database supports transactions, so PDO needs to run in what is known as "auto-commit" mode when you first open the connection. Auto-commit mode means that every query that you run has its own implicit transaction, if the database supports it, or no transaction if the database doesn't support transactions. If you need a transaction, you must use the PDO->beginTransaction() method to initiate one. If the underlying driver does not support transactions, a PDOException will be thrown (regardless of your error handling settings: this is always a serious error condition). Once you are in a transaction, you may use PDO->commit() or PDO->rollBack() to finish it, depending on the success of the code you run during the transaction. -------- 8< ---------------- 8< ---------------- 8< -------- So, I feel safe enough with my assumption. I'm not entirely sure about the stored procedure, though. I've recently rewritten the procedure as separate queries, but don't know if that will help until I hit a high peak load again. -- Dante
"D. Dante Lorenso" <dante@lorenso.com> writes:
> Tom Lane wrote:
>> "D. Dante Lorenso" <dante@lorenso.com> writes:
>>> I am calling a stored proc from PHP. Since I do not begin a
>>> transaction, I assume that my call is automatically committed
>>> immediately after invocation.
>>
>> Have you actually verified that, or are you just assuming it?
> Just assuming. I'm not really sure HOW to verify it, though.
Enable query logging on the server and look for BEGIN commands?
regards, tom lane