Обсуждение: processing large amount of rows with plpgsql
hello everybody, we are trying to move the data from table1 into table2 using a plpgsql stored procedure which is performing simple a dataconversion there are about 50 million rows the tables are relatively simple, less than a dozen columns, most are integer, a couple are char(32) and one is varchar holdingURLs what happens is that when we execute the stored procedure, the execution eats up the 17 GB free space on the server and theserver crashes if we apply a limit of 1 million rows, the execution completes successfully in about a minute we understand, by now, that plpgsql functions are executed in their own transaction, which commits when the function comesto an end the question is - are plpgsql functions suitable for such massive operations on data and can this work without so much disk space is beingeaten for something which should be simply "read-change-write, read-change-write, read-change-write, …"? i. e. any wayto force commit inside, or so? or should we rather implement this operation in some external scripting language (speed is not that important, this is aone time conversion) thank you, geert
On Wed, Aug 8, 2012 at 2:41 PM, Geert Mak <pobox@verysmall.org> wrote: > hello everybody, > > we are trying to move the data from table1 into table2 using a plpgsql stored procedure which is performing simple a dataconversion > > there are about 50 million rows > > the tables are relatively simple, less than a dozen columns, most are integer, a couple are char(32) and one is varcharholding URLs > > what happens is that when we execute the stored procedure, the execution eats up the 17 GB free space on the server andthe server crashes > > if we apply a limit of 1 million rows, the execution completes successfully in about a minute > > we understand, by now, that plpgsql functions are executed in their own transaction, which commits when the function comesto an end > > the question is - > > are plpgsql functions suitable for such massive operations on data and can this work without so much disk space is beingeaten for something which should be simply "read-change-write, read-change-write, read-change-write, …"? i. e. any wayto force commit inside, or so? > > or should we rather implement this operation in some external scripting language (speed is not that important, this isa one time conversion) What is the general structure of the procedure? In particular, how are you browsing and updating the rows? There is (almost) no way to force commit inside a function -- there has been some discussion about stored procedure and/or autonomous transaction feature in terms of getting there. I say 'almost' because you can emulate some aspects of autonomous transactions with dblink, but that may not be a very good fit for your particular case. merlin
On 08.08.2012, at 22:04, Merlin Moncure wrote: > What is the general structure of the procedure? In particular, how > are you browsing and updating the rows? Here it is - BEGIN for statistics_row in SELECT * FROM statistics ORDER BY time ASC LOOP ... ... here some very minimal transformation is done ... and the row is written into the second table ... END LOOP; RETURN 1; END; > There is (almost) no way to > force commit inside a function -- So what you are saying is that this behavior is normal and we should either equip ourselves with enough disk space (whichI am trying now, it is a cloud server, which I am resizing to gain more disk space and see what will happen) or doit with an external (scripting) language? > there has been some discussion about > stored procedure and/or autonomous transaction feature in terms of > getting there. > > I say 'almost' because you can emulate some aspects of autonomous > transactions with dblink, but that may not be a very good fit for your > particular case. I met already dblink mention in this context somewhere... Though if plpgsql performs well with more disk space, I'll leaveit for now. It is a one time operation this one. Thank you, Geert
> > There is (almost) no way to > > force commit inside a function -- > > So what you are saying is that this behavior is normal and we should > either equip ourselves with enough disk space (which I am trying now, > it is a cloud server, which I am resizing to gain more disk space and > see what will happen) or do it with an external (scripting) language? > Hello, a relative simple way to workaround your performance/resource problem is to slice the update. e.g.: create function myupdate(slice int) ... for statistics_row in SELECT * FROM statistics WHERE id % 16 = slice or: WHERE hashtext(id::text) % 16 = slice ... and then call your function with the values 1..15 (when using 16 slices) Use a power of 2 for the number of slices. It may be faster to use many slices and this allows to do the job in parallel on a few threads. HTH, Marc Mamin
oops, > and then call your function with the values 1..15 (when using 16 it should of course be 0..15 Marc Mamin > -----Original Message----- > From: pgsql-general-owner@postgresql.org [mailto:pgsql-general- > owner@postgresql.org] On Behalf Of Marc Mamin > Sent: Donnerstag, 9. August 2012 09:12 > To: Geert Mak; pgsql-general@postgresql.org > Subject: Re: [GENERAL] processing large amount of rows with plpgsql > > > > There is (almost) no way to > > > force commit inside a function -- > > > > So what you are saying is that this behavior is normal and we should > > either equip ourselves with enough disk space (which I am trying now, > > it is a cloud server, which I am resizing to gain more disk space and > > see what will happen) or do it with an external (scripting) language? > > > > Hello, > > a relative simple way to workaround your performance/resource problem > is > to slice the update. > > e.g.: > > create function myupdate(slice int) ... > > for statistics_row in > SELECT * FROM statistics > WHERE id % 16 = slice > or: > WHERE hashtext(id::text) % 16 = slice > ... > > and then call your function with the values 1..15 (when using 16 > slices) > > Use a power of 2 for the number of slices. > > It may be faster to use many slices and > this allows to do the job in parallel on a few threads. > > HTH, > > Marc Mamin > > -- > Sent via pgsql-general mailing list (pgsql-general@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-general