Thread: OFFSET impact on Performance???
Hi to all,
I have the following 2 examples. Now, regarding on the offset if it is small(10) or big(>50000) what is the impact on the performance of the query?? I noticed that if I return more data's(columns) or if I make more joins then the query runs even slower if the OFFSET is bigger. How can I somehow improve the performance on this?
Best regards,
Andy.
explain
analyzeSELECT o.id
FROM report r
INNER JOIN orders o ON o.id=r.id_order AND o.id_status=6
ORDER BY 1 LIMIT 10 OFFSET 10
Limit (cost=44.37..88.75 rows=10 width=4) (actual time=0.160..0.275 rows=10 loops=1)
-> Merge Join (cost=0.00..182150.17 rows=41049 width=4) (actual time=0.041..0.260 rows=20 loops=1)
Merge Cond: ("outer".id_order = "inner".id)
-> Index Scan using report_id_order_idx on report r (cost=0.00..157550.90 rows=42862 width=4) (actual time=0.018..0.075 rows=20 loops=1)
-> Index Scan using orders_pkey on orders o (cost=0.00..24127.04 rows=42501 width=4) (actual time=0.013..0.078 rows=20 loops=1)
Filter: (id_status = 6)
Total runtime: 0.373 ms
-> Merge Join (cost=0.00..182150.17 rows=41049 width=4) (actual time=0.041..0.260 rows=20 loops=1)
Merge Cond: ("outer".id_order = "inner".id)
-> Index Scan using report_id_order_idx on report r (cost=0.00..157550.90 rows=42862 width=4) (actual time=0.018..0.075 rows=20 loops=1)
-> Index Scan using orders_pkey on orders o (cost=0.00..24127.04 rows=42501 width=4) (actual time=0.013..0.078 rows=20 loops=1)
Filter: (id_status = 6)
Total runtime: 0.373 ms
explain
analyzeSELECT o.id
FROM report r
INNER JOIN orders o ON o.id=r.id_order AND o.id_status=6
ORDER BY 1 LIMIT 10 OFFSET 1000000Limit (cost=31216.85..31216.85 rows=1 width=4) (actual time=1168.152..1168.152 rows=0 loops=1)
-> Sort (cost=31114.23..31216.85 rows=41049 width=4) (actual time=1121.769..1152.246 rows=42693 loops=1)
Sort Key: o.id
-> Hash Join (cost=2329.99..27684.03 rows=41049 width=4) (actual time=441.879..925.498 rows=42693 loops=1)
Hash Cond: ("outer".id_order = "inner".id)
-> Seq Scan on report r (cost=0.00..23860.62 rows=42862 width=4) (actual time=38.634..366.035 rows=42864 loops=1)
-> Hash (cost=2077.74..2077.74 rows=42501 width=4) (actual time=140.200..140.200 rows=0 loops=1)
-> Seq Scan on orders o (cost=0.00..2077.74 rows=42501 width=4) (actual time=0.059..96.890 rows=42693 loops=1)
Filter: (id_status = 6)
Total runtime: 1170.586 ms
Andrei Bintintan wrote: > Hi to all, > > I have the following 2 examples. Now, regarding on the offset if it > is small(10) or big(>50000) what is the impact on the performance of > the query?? I noticed that if I return more data's(columns) or if I > make more joins then the query runs even slower if the OFFSET is > bigger. How can I somehow improve the performance on this? There's really only one way to do an offset of 1000 and that's to fetch 1000 rows and then some and discard the first 1000. If you're using this to provide "pages" of results, could you use a cursor? -- Richard Huxton Archonet Ltd
> If you're using this to provide "pages" of results, could you use a > cursor? What do you mean by that? Cursor? Yes I'm using this to provide "pages", but If I jump to the last pages it goes very slow. Andy. ----- Original Message ----- From: "Richard Huxton" <dev@archonet.com> To: "Andrei Bintintan" <klodoma@ar-sd.net> Cc: <pgsql-sql@postgresql.org>; <pgsql-performance@postgresql.org> Sent: Thursday, January 20, 2005 2:10 PM Subject: Re: [SQL] OFFSET impact on Performance??? > Andrei Bintintan wrote: >> Hi to all, >> >> I have the following 2 examples. Now, regarding on the offset if it >> is small(10) or big(>50000) what is the impact on the performance of >> the query?? I noticed that if I return more data's(columns) or if I >> make more joins then the query runs even slower if the OFFSET is >> bigger. How can I somehow improve the performance on this? > > There's really only one way to do an offset of 1000 and that's to fetch > 1000 rows and then some and discard the first 1000. > > If you're using this to provide "pages" of results, could you use a > cursor? > > -- > Richard Huxton > Archonet Ltd >
Andrei Bintintan wrote: >> If you're using this to provide "pages" of results, could you use a >> cursor? > > What do you mean by that? Cursor? > > Yes I'm using this to provide "pages", but If I jump to the last pages > it goes very slow. DECLARE mycursor CURSOR FOR SELECT * FROM ... FETCH FORWARD 10 IN mycursor; CLOSE mycursor; Repeated FETCHes would let you step through your results. That won't work if you have a web-app making repeated connections. If you've got a web-application then you'll probably want to insert the results into a cache table for later use. -- Richard Huxton Archonet Ltd
Alex Turner wrote: > I am also very interesting in this very question.. Is there any way > to declare a persistant cursor that remains open between pg sessions? Not sure how this would work. What do you do with multiple connections? Only one can access the cursor, so which should it be? > This would be better than a temp table because you would not have to > do the initial select and insert into a fresh table and incur those > IO costs, which are often very heavy, and the reason why one would > want to use a cursor. I'm pretty sure two things mean there's less difference than you might expect: 1. Temp tables don't fsync 2. A cursor will spill to disk beyond a certain size -- Richard Huxton Archonet Ltd
"Andrei Bintintan" <klodoma@ar-sd.net> writes: > > If you're using this to provide "pages" of results, could you use a cursor? > What do you mean by that? Cursor? > > Yes I'm using this to provide "pages", but If I jump to the last pages it goes > very slow. The best way to do pages for is not to use offset or cursors but to use an index. This only works if you can enumerate all the sort orders the application might be using and can have an index on each of them. To do this the query would look something like: SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 Then you take note of the last value used on a given page and if the user selects "next" you pass that as the starting point for the next page. This query takes the same amount of time no matter how many records are in the table and no matter what page of the result set the user is on. It should actually be instantaneous even if the user is on the hundredth page of millions of records because it uses an index both for the finding the right point to start and for the ordering. It also has the advantage that it works even if the list of items changes as the user navigates. If you use OFFSET and someone inserts a record in the table then the "next" page will overlap the current page. Worse, if someone deletes a record then "next" will skip a record. The disadvantages of this are a) it's hard (but not impossible) to go backwards. And b) it's impossible to give the user a list of pages and let them skip around willy nilly. (If this is for a web page then specifically don't recommend cursors. It will mean you'll have to have some complex session management system that guarantees the user will always come to the same postgres session and has some garbage collection if the user disappears. And it means the URL is only good for a limited amount of time. If they bookmark it it'll break if they come back the next day.) -- greg
Greg Stark wrote: > "Andrei Bintintan" <klodoma@ar-sd.net> writes: > > >>>If you're using this to provide "pages" of results, could you use a cursor? >> >>What do you mean by that? Cursor? >> >>Yes I'm using this to provide "pages", but If I jump to the last pages it goes >>very slow. > > > The best way to do pages for is not to use offset or cursors but to use an > index. This only works if you can enumerate all the sort orders the > application might be using and can have an index on each of them. > > To do this the query would look something like: > > SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 > > Then you take note of the last value used on a given page and if the user > selects "next" you pass that as the starting point for the next page. Greg's is the most efficient, but you need to make sure you have a suitable key available in the output of your select. Also, since you are repeating the query you could get different results as people insert/delete rows. This might or might not be what you want. A similar solution is to partition by date/alphabet or similar, then page those results. That can reduce your resultset to a manageable size. -- Richard Huxton Archonet Ltd
Now I read all the posts and I have some answers. Yes, I have a web aplication. I HAVE to know exactly how many pages I have and I have to allow the user to jump to a specific page(this is where I used limit and offset). We have this feature and I cannot take it out. >> > SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 Now this solution looks very fast, but I cannot implement it, because I cannot jump from page 1 to page xxxx only to page 2. Because I know with this type where did the page 1 ended. And we have some really complicated where's and about 10 tables are involved in the sql query. About the CURSOR I have to read more about them because this is my first time when I hear about. I don't know if temporary tables are a solution, really I don't think so, there are a lot of users that are working in the same time at the same page. So... still DIGGING for solutions. Andy. ----- Original Message ----- From: "Ragnar Hafstað" <gnari@simnet.is> To: <pgsql-performance@postgresql.org> Cc: "Andrei Bintintan" <klodoma@ar-sd.net>; <pgsql-sql@postgresql.org> Sent: Thursday, January 20, 2005 9:23 PM Subject: Re: [PERFORM] [SQL] OFFSET impact on Performance??? > On Thu, 2005-01-20 at 19:12 +0000, Ragnar Hafstað wrote: >> On Thu, 2005-01-20 at 11:59 -0500, Greg Stark wrote: >> >> > The best way to do pages for is not to use offset or cursors but to use >> > an >> > index. This only works if you can enumerate all the sort orders the >> > application might be using and can have an index on each of them. >> > >> > To do this the query would look something like: >> > >> > SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 >> > >> > Then you take note of the last value used on a given page and if the >> > user >> > selects "next" you pass that as the starting point for the next page. >> >> this will only work unchanged if the index is unique. imagine , for >> example if you have more than 50 rows with the same value of col. >> >> one way to fix this is to use ORDER BY col,oid > > and a slightly more complex WHERE clause as well, of course > > gnari > > >
I am also very interesting in this very question.. Is there any way to declare a persistant cursor that remains open between pg sessions? This would be better than a temp table because you would not have to do the initial select and insert into a fresh table and incur those IO costs, which are often very heavy, and the reason why one would want to use a cursor. Alex Turner NetEconomist On Thu, 20 Jan 2005 15:20:59 +0000, Richard Huxton <dev@archonet.com> wrote: > Andrei Bintintan wrote: > >> If you're using this to provide "pages" of results, could you use a > >> cursor? > > > > What do you mean by that? Cursor? > > > > Yes I'm using this to provide "pages", but If I jump to the last pages > > it goes very slow. > > DECLARE mycursor CURSOR FOR SELECT * FROM ... > FETCH FORWARD 10 IN mycursor; > CLOSE mycursor; > > Repeated FETCHes would let you step through your results. That won't > work if you have a web-app making repeated connections. > > If you've got a web-application then you'll probably want to insert the > results into a cache table for later use. > > -- > Richard Huxton > Archonet Ltd > > ---------------------------(end of broadcast)--------------------------- > TIP 3: if posting/reading through Usenet, please send an appropriate > subscribe-nomail command to majordomo@postgresql.org so that your > message can get through to the mailing list cleanly >
On Thu, 2005-01-20 at 19:12 +0000, Ragnar Hafstað wrote: > On Thu, 2005-01-20 at 11:59 -0500, Greg Stark wrote: > > > The best way to do pages for is not to use offset or cursors but to use an > > index. This only works if you can enumerate all the sort orders the > > application might be using and can have an index on each of them. > > > > To do this the query would look something like: > > > > SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 > > > > Then you take note of the last value used on a given page and if the user > > selects "next" you pass that as the starting point for the next page. > > this will only work unchanged if the index is unique. imagine , for > example if you have more than 50 rows with the same value of col. > > one way to fix this is to use ORDER BY col,oid and a slightly more complex WHERE clause as well, of course gnari
On Thu, 2005-01-20 at 11:59 -0500, Greg Stark wrote: > The best way to do pages for is not to use offset or cursors but to use an > index. This only works if you can enumerate all the sort orders the > application might be using and can have an index on each of them. > > To do this the query would look something like: > > SELECT * FROM tab WHERE col > ? ORDER BY col LIMIT 50 > > Then you take note of the last value used on a given page and if the user > selects "next" you pass that as the starting point for the next page. this will only work unchanged if the index is unique. imagine , for example if you have more than 50 rows with the same value of col. one way to fix this is to use ORDER BY col,oid gnari
Alex Turner <armtuk@gmail.com> writes: > I am also very interesting in this very question.. Is there any way to > declare a persistant cursor that remains open between pg sessions? > This would be better than a temp table because you would not have to > do the initial select and insert into a fresh table and incur those IO > costs, which are often very heavy, and the reason why one would want > to use a cursor. TANSTAAFL. How would such a persistent cursor be implemented if not by building a temporary table somewhere behind the scenes? There could be some advantage if the data were stored in a temporary table marked as not having to be WAL logged. Instead it could be automatically cleared on every database start. -- greg
The problems still stays open. The thing is that I have about 20 - 30 clients that are using that SQL query where the offset and limit are involved. So, I cannot create a temp table, because that means that I'll have to make a temp table for each session... which is a very bad ideea. Cursors somehow the same. In my application the Where conditions can be very different for each user(session) apart. The only solution that I see in the moment is to work at the query, or to write a more complex where function to limit the results output. So no replace for Offset/Limit. Best regards, Andy. ----- Original Message ----- From: "Greg Stark" <gsstark@mit.edu> To: <alex@neteconomist.com> Cc: "Richard Huxton" <dev@archonet.com>; "Andrei Bintintan" <klodoma@ar-sd.net>; <pgsql-sql@postgresql.org>; <pgsql-performance@postgresql.org> Sent: Tuesday, January 25, 2005 8:28 PM Subject: Re: [PERFORM] [SQL] OFFSET impact on Performance??? > > Alex Turner <armtuk@gmail.com> writes: > >> I am also very interesting in this very question.. Is there any way to >> declare a persistant cursor that remains open between pg sessions? >> This would be better than a temp table because you would not have to >> do the initial select and insert into a fresh table and incur those IO >> costs, which are often very heavy, and the reason why one would want >> to use a cursor. > > TANSTAAFL. How would such a persistent cursor be implemented if not by > building a temporary table somewhere behind the scenes? > > There could be some advantage if the data were stored in a temporary table > marked as not having to be WAL logged. Instead it could be automatically > cleared on every database start. > > -- > greg > >
Alex Turner wrote: > As I read the docs, a temp table doesn't solve our problem, as it does > not persist between sessions. With a web page there is no guarentee > that you will receive the same connection between requests, so a temp > table doesn't solve the problem. It looks like you either have to > create a real table (which is undesirable becuase it has to be > physicaly synced, and TTFB will be very poor) or create an application > tier in between the web tier and the database tier to allow data to > persist between requests tied to a unique session id. > > Looks like the solutions to this problem is not RDBMS IMHO. It's less the RDBMS than the web application. You're trying to mix a stateful setup (the application) with a stateless presentation layer (the web). If you're using PHP (which doesn't offer a "real" middle layer) you might want to look at memcached. -- Richard Huxton Archonet Ltd
As I read the docs, a temp table doesn't solve our problem, as it does not persist between sessions. With a web page there is no guarentee that you will receive the same connection between requests, so a temp table doesn't solve the problem. It looks like you either have to create a real table (which is undesirable becuase it has to be physicaly synced, and TTFB will be very poor) or create an application tier in between the web tier and the database tier to allow data to persist between requests tied to a unique session id. Looks like the solutions to this problem is not RDBMS IMHO. Alex Turner NetEconomist On Wed, 26 Jan 2005 12:11:49 +0200, Andrei Bintintan <klodoma@ar-sd.net> wrote: > The problems still stays open. > > The thing is that I have about 20 - 30 clients that are using that SQL query > where the offset and limit are involved. So, I cannot create a temp table, > because that means that I'll have to make a temp table for each session... > which is a very bad ideea. Cursors somehow the same. In my application the > Where conditions can be very different for each user(session) apart. > > The only solution that I see in the moment is to work at the query, or to > write a more complex where function to limit the results output. So no > replace for Offset/Limit. > > Best regards, > Andy. > > > ----- Original Message ----- > From: "Greg Stark" <gsstark@mit.edu> > To: <alex@neteconomist.com> > Cc: "Richard Huxton" <dev@archonet.com>; "Andrei Bintintan" > <klodoma@ar-sd.net>; <pgsql-sql@postgresql.org>; > <pgsql-performance@postgresql.org> > Sent: Tuesday, January 25, 2005 8:28 PM > Subject: Re: [PERFORM] [SQL] OFFSET impact on Performance??? > > > > > > Alex Turner <armtuk@gmail.com> writes: > > > >> I am also very interesting in this very question.. Is there any way to > >> declare a persistant cursor that remains open between pg sessions? > >> This would be better than a temp table because you would not have to > >> do the initial select and insert into a fresh table and incur those IO > >> costs, which are often very heavy, and the reason why one would want > >> to use a cursor. > > > > TANSTAAFL. How would such a persistent cursor be implemented if not by > > building a temporary table somewhere behind the scenes? > > > > There could be some advantage if the data were stored in a temporary table > > marked as not having to be WAL logged. Instead it could be automatically > > cleared on every database start. > > > > -- > > greg > > > > > >