Thread: Cursors and Transactions, why?
Why must a cursor be defined in an open transaction? Obviously there's a good reason, but I can't figure it out. On a high level, what would be involved in allowing a cursor to outlive the transaction that created it? Cursors seem as if they have some nice performance benefits (esp. if you're not using all rows found), but their usefulness drops considerably since you must leave a transaction open. eric
On Apr 5, 2004, at 6:44 PM, Joe Conway wrote: > Eric Ridge wrote: >> Why must a cursor be defined in an open transaction? Obviously >> there's a good reason, but I can't figure it out. On a high level, >> what would be involved in allowing a cursor to outlive the >> transaction that created it? > > Historically I think it was because the memory was released at the end > of the current transaction (i.e. allocations were made in > TopTransactionContext). But as of 7.4, cursors *can* outlive > transactions: > http://www.postgresql.org/docs/current/interactive/sql-declare.html > > WITH HOLD > WITHOUT HOLD holy cow! This is fantastic. I had no idea. <short pause> ooh, and I see FETCH, in 7.4, supports absolute positioning. Must upgrade. thanks! eric
Eric Ridge wrote: > Why must a cursor be defined in an open transaction? Obviously there's > a good reason, but I can't figure it out. On a high level, what would > be involved in allowing a cursor to outlive the transaction that created > it? Historically I think it was because the memory was released at the end of the current transaction (i.e. allocations were made in TopTransactionContext). But as of 7.4, cursors *can* outlive transactions: http://www.postgresql.org/docs/current/interactive/sql-declare.html WITH HOLD WITHOUT HOLD WITH HOLD specifies that the cursor may continue to be used after the transaction that created it successfully commits. WITHOUT HOLD specifies that the cursor cannot be used outside of the transaction that created it. If neither WITHOUT HOLD nor WITH HOLD is specified, WITHOUT HOLD is the default. HTH, Joe
Eric Ridge wrote: > Why must a cursor be defined in an open transaction? Obviously there's > a good reason, but I can't figure it out. On a high level, what would > be involved in allowing a cursor to outlive the transaction that > created it? Because the transaction is what protects the rows that build the result set from being removed by vacuum. In PostgreSQL, a cursor is a running query executor just sitting in the middle of its operation. If the underlying query is for example a simple sequential scan, then the result set is not materialized but every future fetch operation will read directly from the base table. This would obviously get screwed up if vacuum would think nobody needs those rows any more. > > Cursors seem as if they have some nice performance benefits (esp. if > you're not using all rows found), but their usefulness drops > considerably since you must leave a transaction open. And now you know why they are so good if you don't use all rows. This benefit I think goes away if you use Joe Conway's suggestion of WITH HOLD. Jan -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #================================================== JanWieck@Yahoo.com #
On 4/6/04 10:54 AM, "Jan Wieck" <JanWieck@yahoo.com> wrote: >> Cursors seem as if they have some nice performance benefits (esp. if >> you're not using all rows found), but their usefulness drops >> considerably since you must leave a transaction open. > > And now you know why they are so good if you don't use all rows. This > benefit I think goes away if you use Joe Conway's suggestion of WITH HOLD. I tried using WITH HOLD in the following case (using an ecpg C program): foreach row in table A update table B with value from table A commit once every 10,000 updates forend I created a cursor on table A. Without WITH HOLD, obviously I got an error on the next TABLE A fetch because the COMMIT closed the cursor. I added 'WITH HOLD' to the cursor. On the first COMMIT, the application hung. I assume the COMMIT would have completed after some period of time, but I didn't wait that long. There are 20 million rows in table A and 60 million in table B (one to many relationship). Is this hang on COMMIT when using WITH HOLD to be expected? Is there a way around it? I don't think it's reasonable put the entire 60 million updates in a single transaction. The kludge solution I implemented was to write out all the data I needed from table A to a file, then read that file and update table B. Wes
<wespvp@syntegra.com> writes:
> Is this hang on COMMIT when using WITH HOLD to be expected?
Yes. WITH HOLD is not magic, it just makes a materialized copy of the
SELECT result. If you're talking about a multi-million-row result,
it's gonna take awhile.
> The kludge solution I implemented was to write out all the data I
> needed from table A to a file, then read that file and update table B.
In theory at least, that should not be any faster than a WITH HOLD
cursor, since you're effectively replicating the same functionality
outside the database ...
regards, tom lane
On 4/6/04 3:55 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote: >> The kludge solution I implemented was to write out all the data I >> needed from table A to a file, then read that file and update table B. > > In theory at least, that should not be any faster than a WITH HOLD > cursor, since you're effectively replicating the same functionality > outside the database ... Except for the "out of memory" thing... Are you saying that once the first COMMIT completed, all COMMIT's after that would function at normal speed - only the first one has to save the result set? Wes
On Apr 6, 2004, at 11:54 AM, Jan Wieck wrote: > Eric Ridge wrote: > >> Why must a cursor be defined in an open transaction? Obviously >> there's a good reason, but I can't figure it out. On a high level, >> what would be involved in allowing a cursor to outlive the >> transaction that created it? > > Because the transaction is what protects the rows that build the > result set from being removed by vacuum. In PostgreSQL, a cursor is a > running query executor just sitting in the middle of its operation. That's a good thing to know. > If the underlying query is for example a simple sequential scan, then > the result set is not materialized but every future fetch operation > will read directly from the base table. This would obviously get > screwed up if vacuum would think nobody needs those rows any more. Is vacuum the only thing that would muck with the rows? >> Cursors seem as if they have some nice performance benefits (esp. if >> you're not using all rows found), but their usefulness drops >> considerably since you must leave a transaction open. > > And now you know why they are so good if you don't use all rows. This > benefit I think goes away if you use Joe Conway's suggestion of WITH > HOLD. Okay, so WITH HOLD is actually materializing the entire resultset (sequential scan or otherwise)? If that's true, you're right, some of the benefits do go away. I need to setup a 7.4 test server and play with this some, and figure out if the benefits are really what I want them to be. I do appreciate the insight into how cursors work... it helps a lot! eric
<wespvp@syntegra.com> writes:
> On 4/6/04 3:55 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:
>> In theory at least, that should not be any faster than a WITH HOLD
>> cursor, since you're effectively replicating the same functionality
>> outside the database ...
> Except for the "out of memory" thing...
What "out of memory thing"? The tuplestore code is perfectly capable of
spilling to disk --- in fact the usual performance gripe against it has
to do with spilling too soon, because sort_mem is set too small.
regards, tom lane
Eric Ridge wrote: > On Apr 6, 2004, at 11:54 AM, Jan Wieck wrote: >> And now you know why they are so good if you don't use all rows. This >> benefit I think goes away if you use Joe Conway's suggestion of WITH >> HOLD. > > Okay, so WITH HOLD is actually materializing the entire resultset > (sequential scan or otherwise)? If that's true, you're right, some of > the benefits do go away. Keep in mind that the tuplestore stays in memory as long as it fits within sort_mem kilobytes. And you can do: set sort_mem to <some_large_number>; prior to COMMIT, and then set sort_mem to default; after COMMIT, as long as you can afford the memory use. A bit ugly, but it might come in handy ;-) Joe
Eric Ridge wrote: > On Apr 6, 2004, at 11:54 AM, Jan Wieck wrote: >> If the underlying query is for example a simple sequential scan, then >> the result set is not materialized but every future fetch operation >> will read directly from the base table. This would obviously get >> screwed up if vacuum would think nobody needs those rows any more. > > Is vacuum the only thing that would muck with the rows? Vacuum is the only thing that cares for the dustmites, yes. > I need to setup a 7.4 test server and play with this some, and figure > out if the benefits are really what I want them to be. I do appreciate > the insight into how cursors work... it helps a lot! Experience and knowledge can only be replaced by more experience and more knowledge. Jan -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #================================================== JanWieck@Yahoo.com #
On Apr 7, 2004, at 7:51 AM, Jan Wieck wrote: > Eric Ridge wrote: >> On Apr 6, 2004, at 11:54 AM, Jan Wieck wrote: >>> If the underlying query is for example a simple sequential scan, >>> then the result set is not materialized but every future fetch >>> operation will read directly from the base table. This would >>> obviously get screwed up if vacuum would think nobody needs those >>> rows any more. >> Is vacuum the only thing that would muck with the rows? > > Vacuum is the only thing that cares for the dustmites, yes. And WITH HOLD is strong enough to defend against a vacuum, I hope... >> I need to setup a 7.4 test server and play with this some, and figure >> out if the benefits are really what I want them to be. I do >> appreciate the insight into how cursors work... it helps a lot! > > Experience and knowledge can only be replaced by more experience and > more knowledge. Very wise words. My real problem is that the JDBC drivers (and I assume this is true for all client interfaces) buffer the results of a SELECT in memory, because the backend pushes out all the tuples as the response. I'm not dealing with a large number of rows (only a few thousand), but they've very wide, and many contain fields with multi-megabyte data. In some situations, when I've got a lot of open ResultSets, the JVM throws OutOfMemory errors. One half-baked thought was to hack the JDBC drivers to have 'em gzip large resultsets in memory. Wouldn't completely solve the problem, but would probably help quite a bit. But the better solution is to use cursors. We're not in a position to upgrade to 7.4 just yet, so we'll just deal with the OutOfMemory errors until we can. eric
On Apr 7, 2004, at 12:43 AM, Joe Conway wrote: > Eric Ridge wrote: >> On Apr 6, 2004, at 11:54 AM, Jan Wieck wrote: >>> And now you know why they are so good if you don't use all rows. >>> This benefit I think goes away if you use Joe Conway's suggestion of >>> WITH HOLD. >> Okay, so WITH HOLD is actually materializing the entire resultset >> (sequential scan or otherwise)? If that's true, you're right, some >> of the benefits do go away. > > Keep in mind that the tuplestore stays in memory as long as it fits > within sort_mem kilobytes. And you can do: More good information. Thanks! Is the tuplestore basically just an array of ItemPointer-s? In mean, it's not a copy of each entire row, is it? eric
On 4/6/04 11:09 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote: > What "out of memory thing"? The tuplestore code is perfectly capable of > spilling to disk --- in fact the usual performance gripe against it has > to do with spilling too soon, because sort_mem is set too small. I tried doing a mass update of all rows with a single SQL statement in psql and after it ran for many hours, I got 'out of memory'. I didn't try that using C and WITH HOLD. I assumed it ran out of swap space, but was sleeping at the time. Wes
<wespvp@syntegra.com> writes:
> On 4/6/04 11:09 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:
>> What "out of memory thing"? The tuplestore code is perfectly capable of
>> spilling to disk --- in fact the usual performance gripe against it has
>> to do with spilling too soon, because sort_mem is set too small.
> I tried doing a mass update of all rows with a single SQL statement in psql
> and after it ran for many hours, I got 'out of memory'.
It's unlikely that that had anything to do with tuplestores. My guess
is that you had some AFTER triggers on the updated table (eg, for
foreign key checks) and that what we ran out of memory for was the list
of deferred trigger events. Last I checked, there was still not
provision to spill that to disk when it gets big :-(
regards, tom lane
On Wed, 7 Apr 2004, Eric Ridge wrote: > > My real problem is that the JDBC drivers (and I assume this is true for > all client interfaces) buffer the results of a SELECT in memory, > because the backend pushes out all the tuples as the response. I'm not > dealing with a large number of rows (only a few thousand), but they've > very wide, and many contain fields with multi-megabyte data. In some > situations, when I've got a lot of open ResultSets, the JVM throws > OutOfMemory errors. > The 7.4 jdbc driver has the ability to use cursors behind the scenes on queries. This is done by calling Statement.setFetchSize(n) to retrieve n rows at a time. There are a number of other restrictions: you must be in a transaction and the ResultSet type must be FORWARD_ONLY. You can use the 7.4 jdbc driver against a 7.3 server as well so this may provide some relief. Kris Jurka
Eric Ridge wrote: > Is the tuplestore basically just an array of ItemPointer-s? In mean, > it's not a copy of each entire row, is it? Yup, it is copied: src/backend/utils/sort/tuplestore.c:tuplestore_puttuple() 8<-------------------------------------------------------- /* * Accept one tuple and append it to the tuplestore. * * Note that the input tuple is always copied; the caller need not save * it. 8<-------------------------------------------------------- Joe
On Apr 7, 2004, at 2:01 PM, Kris Jurka wrote: > The 7.4 jdbc driver has the ability to use cursors behind the scenes on > queries. This is done by calling Statement.setFetchSize(n) to > retrieve n > rows at a time. There are a number of other restrictions: you must be > in > a transaction and the ResultSet type must be FORWARD_ONLY. You can use > the 7.4 jdbc driver against a 7.3 server as well so this may provide > some > relief. I remember reading about this long ago. It's not really an option in my little world because of the open transaction bit. My solution is eventually going to be to upgrade to 7.4... heh, probably about the time 7.5 is released. *sigh* eric
On Apr 7, 2004, at 7:48 PM, Joe Conway wrote: > Eric Ridge wrote: >> Is the tuplestore basically just an array of ItemPointer-s? In mean, >> it's not a copy of each entire row, is it? > > Yup, it is copied: wow. I should go back and read the archives to see if this was discussed already, but I can't help but wonder if there's a way to only copy pointers to the tuples. I realize VACUUM could screw it up, but maybe something could be invented (or re-used) to help guard against that. eric
Eric Ridge <ebr@tcdi.com> writes:
> wow. I should go back and read the archives to see if this was
> discussed already, but I can't help but wonder if there's a way to only
> copy pointers to the tuples. I realize VACUUM could screw it up, but
> maybe something could be invented (or re-used) to help guard against
> that.
Still looking for that free lunch, eh?
If you want to leave the tuples on-disk, then you hold a transaction
open to ensure that VACUUM won't whisk them out from under you. That's
what the normal non-HOLD cursor case will do.
If you don't want to leave the tuples on-disk, you have to copy them
someplace. You can do that with a HOLD cursor.
AFAICS any other solution will simply reinvent one or the other of these
techniques.
regards, tom lane