Re: [HACKERS] Maximum statistics target
On Thu, Mar 20, 2008 at 11:17:10AM -0500, Decibel! wrote: > On Mar 10, 2008, at 1:26 PM, Peter Eisentraut wrote: > > At some point I think it makes a lot more sense to just have VACUUM gather > stats as it goes, rather than have ANALYZE generate a bunch of random IO. > > BTW, when it comes to the case of the OP, perhaps we can build enough > intelligence for the system to understand when the stats follow some type > of pattern (ie: a geometric distribution), and store the stats differently. > -- > Decibel!, aka Jim C. Nasby, Database Architect [EMAIL PROTECTED] > Give your computer some brain candy! www.distributed.net Team #1828 > +1 for opportunistically gathering stats during other I/O such as vacuums and sequential scans. It would be interesting to have a hook to allow processes to attach to the dataflow from other queries. Cheers, Ken -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
On Mar 10, 2008, at 1:26 PM, Peter Eisentraut wrote: Am Montag, 10. März 2008 schrieb Gregory Stark: It's not possible to believe that you'd not notice O(N^2) behavior for N approaching 80 ;-). Perhaps your join columns were unique keys, and thus didn't have any most-common-values? We could remove the hard limit on statistics target and impose the limit instead on the actual size of the arrays. Ie, allow people to specify larger sample sizes and discard unreasonably large excess data (possibly warning them when that happens). I have run some more useful tests now with more distinct values. The planning times do increase, but this is not the primary worry. If you want to spend 20 seconds of planning to speed up your query by 40 seconds, this could surely be a win in some scenarios, and not a catastrophic loss if not. The practical problems lie with memory usage in ANALYZE, in two ways. First, at some point it will try to construct pg_statistic rows that don't fit into the 1GB limit, as mentioned upthread. You get a funny error message and it aborts. This is fixable with some cosmetics. Second, ANALYZE appears to temporarily leak memory (it probably doesn't bother to free things along the way, as most of the code does), and so some not so large statistics targets (say, 4) can get your system swapping like crazy. A crafty user could probably kill the system that way, perhaps even with the restricted settings we have now. I haven't inspected the code in detail yet, but I imagine a few pfree() calls and/or a counter that checks the current memory usage against maintenance_work_mem could provide additional safety. If we could get ANALYZE under control, then I imagine this would provide a more natural upper bound for the statistics targets, and it would be controllable by the administrator. At some point I think it makes a lot more sense to just have VACUUM gather stats as it goes, rather than have ANALYZE generate a bunch of random IO. BTW, when it comes to the case of the OP, perhaps we can build enough intelligence for the system to understand when the stats follow some type of pattern (ie: a geometric distribution), and store the stats differently. -- Decibel!, aka Jim C. Nasby, Database Architect [EMAIL PROTECTED] Give your computer some brain candy! www.distributed.net Team #1828 smime.p7s Description: S/MIME cryptographic signature
Re: [HACKERS] Maximum statistics target
> We could remove the hard limit on statistics target and > impose the limit > instead on the actual size of the arrays. Ie, allow people to > specify larger > sample sizes and discard unreasonably large excess data > (possibly warning them > when that happens). > > That would remove the screw case the original poster had > where he needed to > scan a large portion of the table to see at least one of > every value even > though there were only 169 distinct values. > > -- > Gregory Stark That was my use case, but I wasn't the OP. Your suggestion would satisfy what I was trying to do. However, a higher stats target wouldn't solve my root problem (how the planner uses the gathered stats), and the statistics gathered at 1000 (and indeed at 200) are quite a good representation of what is in the table. I don't like the idea of changing one limit into two limits. Or are you suggesting changing the algorithm that determines how many, and which pages to analyze, perhaps so that it is adaptive to the results of the analysis as it progresses? That doesn't sound easy. Regards, Stephen Denne. Disclaimer: At the Datamail Group we value team commitment, respect, achievement, customer focus, and courage. This email with any attachments is confidential and may be subject to legal privilege. If it is not intended for you please advise by reply immediately, destroy it and do not copy, disclose or use it in any way. __ This email has been scanned by the DMZGlobal Business Quality Electronic Messaging Suite. Please see http://www.dmzglobal.com/services/bqem.htm for details. __ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
Am Montag, 10. März 2008 schrieb Gregory Stark: > > It's not possible to believe that you'd not notice O(N^2) behavior for N > > approaching 80 ;-). Perhaps your join columns were unique keys, and > > thus didn't have any most-common-values? > > We could remove the hard limit on statistics target and impose the limit > instead on the actual size of the arrays. Ie, allow people to specify > larger sample sizes and discard unreasonably large excess data (possibly > warning them when that happens). I have run some more useful tests now with more distinct values. The planning times do increase, but this is not the primary worry. If you want to spend 20 seconds of planning to speed up your query by 40 seconds, this could surely be a win in some scenarios, and not a catastrophic loss if not. The practical problems lie with memory usage in ANALYZE, in two ways. First, at some point it will try to construct pg_statistic rows that don't fit into the 1GB limit, as mentioned upthread. You get a funny error message and it aborts. This is fixable with some cosmetics. Second, ANALYZE appears to temporarily leak memory (it probably doesn't bother to free things along the way, as most of the code does), and so some not so large statistics targets (say, 4) can get your system swapping like crazy. A crafty user could probably kill the system that way, perhaps even with the restricted settings we have now. I haven't inspected the code in detail yet, but I imagine a few pfree() calls and/or a counter that checks the current memory usage against maintenance_work_mem could provide additional safety. If we could get ANALYZE under control, then I imagine this would provide a more natural upper bound for the statistics targets, and it would be controllable by the administrator. -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
"Tom Lane" <[EMAIL PROTECTED]> writes: > Peter Eisentraut <[EMAIL PROTECTED]> writes: >> Am Freitag, 7. März 2008 schrieb Tom Lane: >>> IIRC, egjoinsel is one of the weak spots, so tests involving planning of >>> joins between two tables with large MCV lists would be a good place to >>> start. > >> I have run tests with joining two and three tables with 10 million rows each, >> and the planning times seem to be virtually unaffected by the statistics >> target, for values between 10 and 80. > > It's not possible to believe that you'd not notice O(N^2) behavior for N > approaching 80 ;-). Perhaps your join columns were unique keys, and > thus didn't have any most-common-values? We could remove the hard limit on statistics target and impose the limit instead on the actual size of the arrays. Ie, allow people to specify larger sample sizes and discard unreasonably large excess data (possibly warning them when that happens). That would remove the screw case the original poster had where he needed to scan a large portion of the table to see at least one of every value even though there were only 169 distinct values. -- Gregory Stark EnterpriseDB http://www.enterprisedb.com Ask me about EnterpriseDB's RemoteDBA services! -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
Peter Eisentraut <[EMAIL PROTECTED]> writes: > Am Freitag, 7. März 2008 schrieb Tom Lane: >> IIRC, egjoinsel is one of the weak spots, so tests involving planning of >> joins between two tables with large MCV lists would be a good place to >> start. > I have run tests with joining two and three tables with 10 million rows each, > and the planning times seem to be virtually unaffected by the statistics > target, for values between 10 and 80. It's not possible to believe that you'd not notice O(N^2) behavior for N approaching 80 ;-). Perhaps your join columns were unique keys, and thus didn't have any most-common-values? regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
On Mon, Mar 10, 2008 at 11:36 AM, Peter Eisentraut <[EMAIL PROTECTED]> wrote: > The time to analyze is also quite constant, just before you run out of > memory. :) The MaxAllocSize is the limiting factor in all this. In my > example, statistics targets larger than about 80 created pg_statistic > rows that would have been larger than 1GB, so they couldn't be stored. >From my experience on real life examples, the time to analyze is far from being constant when you raise the statistics target but it may be related to the schema of our tables. cityvox=# \timing Timing is on. cityvox=# show default_statistics_target ; default_statistics_target --- 10 (1 row) Time: 0.101 ms cityvox=# ANALYZE evenement; ANALYZE Time: 406.069 ms cityvox=# ANALYZE evenement; ANALYZE Time: 412.355 ms cityvox=# set default_statistics_target = 30; SET Time: 0.165 ms cityvox=# ANALYZE evenement; ANALYZE Time: 1419.161 ms cityvox=# ANALYZE evenement; ANALYZE Time: 1381.754 ms cityvox=# set default_statistics_target = 100; SET Time: 1.853 ms cityvox=# ANALYZE evenement; ANALYZE Time: 5211.785 ms cityvox=# ANALYZE evenement; ANALYZE Time: 5178.764 ms That said I totally agree that it's not a good idea to have a strict maximum value if we haven't technical reasons for that. -- Guillaume -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
Le Monday 10 March 2008, Peter Eisentraut a écrit : > Am Freitag, 7. März 2008 schrieb Tom Lane: > > I'm not wedded to the number 1000 in particular --- obviously that's > > just a round number. But it would be good to see some performance tests > > with larger settings before deciding that we don't need a limit. > > Well, I'm not saying we should raise the default statistics target. But > setting an arbitrary limit on the grounds that larger values might slow the > system is like limiting the size of tables because larger tables will cause > slower queries. Users should have the option of finding out the best > balance for themselves. If there are concerns with larger statistics > targets, we should document them. I find nothing about this in the > documentation at the moment. I find 2 things: «Increasing the target causes a proportional increase in the time and space needed to do ANALYZE. » in http://www.postgresql.org/docs/current/static/sql-analyze.html and « ... at the price of consuming more space in pg_statistic and slightly more time to compute the estimates» in http://www.postgresql.org/docs/current/static/planner-stats.html But probably not clear enought about time impact in query plan. > > > IIRC, egjoinsel is one of the weak spots, so tests involving planning of > > joins between two tables with large MCV lists would be a good place to > > start. > > I have run tests with joining two and three tables with 10 million rows > each, and the planning times seem to be virtually unaffected by the > statistics target, for values between 10 and 80. They all look more or > less like this: > > test=# explain select * from test1, test2 where test1.a = test2.b; > QUERY PLAN > --- >-- Hash Join (cost=308311.00..819748.00 rows=1000 width=16) >Hash Cond: (test1.a = test2.b) >-> Seq Scan on test1 (cost=0.00..144248.00 rows=1000 width=8) >-> Hash (cost=144248.00..144248.00 rows=1000 width=8) > -> Seq Scan on test2 (cost=0.00..144248.00 rows=1000 > width=8) (5 rows) > > Time: 132,350 ms > > and with indexes > > test=# explain select * from test1, test2 where test1.a = test2.b; > QUERY PLAN > --- >- Merge Join (cost=210416.65..714072.26 rows=1000 > width=16) >Merge Cond: (test1.a = test2.b) >-> Index Scan using test1_index1 on test1 (cost=0.00..282036.13 > rows=1000 width=8) >-> Index Scan using test2_index1 on test2 (cost=0.00..282036.13 > rows=1000 width=8) > (4 rows) > > Time: 168,455 ms > > The time to analyze is also quite constant, just before you run out of > memory. :) The MaxAllocSize is the limiting factor in all this. In my > example, statistics targets larger than about 80 created pg_statistic > rows that would have been larger than 1GB, so they couldn't be stored. > > I suggest that we get rid of the limit of 1000, adequately document > whatever issues might exist with large values (possibly not many, see > above), and add an error message more user-friendly than "invalid memory > alloc request size" for the cases where the value is too large to be > storable. -- Cédric Villemain Administrateur de Base de Données Cel: +33 (0)6 74 15 56 53 http://dalibo.com - http://dalibo.org signature.asc Description: This is a digitally signed message part.
Re: [HACKERS] Maximum statistics target
Am Freitag, 7. März 2008 schrieb Tom Lane: > I'm not wedded to the number 1000 in particular --- obviously that's > just a round number. But it would be good to see some performance tests > with larger settings before deciding that we don't need a limit. Well, I'm not saying we should raise the default statistics target. But setting an arbitrary limit on the grounds that larger values might slow the system is like limiting the size of tables because larger tables will cause slower queries. Users should have the option of finding out the best balance for themselves. If there are concerns with larger statistics targets, we should document them. I find nothing about this in the documentation at the moment. > IIRC, egjoinsel is one of the weak spots, so tests involving planning of > joins between two tables with large MCV lists would be a good place to > start. I have run tests with joining two and three tables with 10 million rows each, and the planning times seem to be virtually unaffected by the statistics target, for values between 10 and 80. They all look more or less like this: test=# explain select * from test1, test2 where test1.a = test2.b; QUERY PLAN - Hash Join (cost=308311.00..819748.00 rows=1000 width=16) Hash Cond: (test1.a = test2.b) -> Seq Scan on test1 (cost=0.00..144248.00 rows=1000 width=8) -> Hash (cost=144248.00..144248.00 rows=1000 width=8) -> Seq Scan on test2 (cost=0.00..144248.00 rows=1000 width=8) (5 rows) Time: 132,350 ms and with indexes test=# explain select * from test1, test2 where test1.a = test2.b; QUERY PLAN Merge Join (cost=210416.65..714072.26 rows=1000 width=16) Merge Cond: (test1.a = test2.b) -> Index Scan using test1_index1 on test1 (cost=0.00..282036.13 rows=1000 width=8) -> Index Scan using test2_index1 on test2 (cost=0.00..282036.13 rows=1000 width=8) (4 rows) Time: 168,455 ms The time to analyze is also quite constant, just before you run out of memory. :) The MaxAllocSize is the limiting factor in all this. In my example, statistics targets larger than about 80 created pg_statistic rows that would have been larger than 1GB, so they couldn't be stored. I suggest that we get rid of the limit of 1000, adequately document whatever issues might exist with large values (possibly not many, see above), and add an error message more user-friendly than "invalid memory alloc request size" for the cases where the value is too large to be storable. -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
Tom Lane wrote: > Martijn van Oosterhout <[EMAIL PROTECTED]> writes: > > On Fri, Mar 07, 2008 at 07:25:25PM +0100, Peter Eisentraut wrote: > >> What's the problem with setting it to ten million if I > have ten million values > >> in the table and I am prepared to spend the resources to > maintain those > >> statistics? > > > That it'll probably take 10 million seconds to calculate the plans > > using it? I think Tom pointed there are a few places that are O(n^2) > > the number entries... > > I'm not wedded to the number 1000 in particular --- obviously that's > just a round number. But it would be good to see some > performance tests > with larger settings before deciding that we don't need a limit. I recently encountered a situation where I would have liked to be able to try a larger limit (amongst other ideas for improving my situation): I have a field whose distribution of frequencies of values is roughly geometric, rather than flat. Total rows = 36 million relpages=504864 Distinct field values in use = 169 10 values account for 50% of the rows. 41 values account for 90% of the rows. After setting statistics target to 1000 for that field, and analyzing the table, the statistics row for that field had 75 most frequent values and a histogram with 76 entries in it. Estimating 151 values in total. For this situation using a larger statistics target should result in more pages being read, and a more accurate record of statistics. It shouldn't result in significantly more work for the planner. It wouldn't solve my problem though, which is frequent over-estimation of rows when restricting by this field with values not known at plan time. Regards, Stephen Denne. Disclaimer: At the Datamail Group we value team commitment, respect, achievement, customer focus, and courage. This email with any attachments is confidential and may be subject to legal privilege. If it is not intended for you please advise by reply immediately, destroy it and do not copy, disclose or use it in any way. __ This email has been scanned by the DMZGlobal Business Quality Electronic Messaging Suite. Please see http://www.dmzglobal.com/services/bqem.htm for details. __ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
Martijn van Oosterhout <[EMAIL PROTECTED]> writes: > On Fri, Mar 07, 2008 at 07:25:25PM +0100, Peter Eisentraut wrote: >> What's the problem with setting it to ten million if I have ten million >> values >> in the table and I am prepared to spend the resources to maintain those >> statistics? > That it'll probably take 10 million seconds to calculate the plans > using it? I think Tom pointed there are a few places that are O(n^2) > the number entries... I'm not wedded to the number 1000 in particular --- obviously that's just a round number. But it would be good to see some performance tests with larger settings before deciding that we don't need a limit. IIRC, egjoinsel is one of the weak spots, so tests involving planning of joins between two tables with large MCV lists would be a good place to start. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Maximum statistics target
On Fri, Mar 07, 2008 at 07:25:25PM +0100, Peter Eisentraut wrote: > What's the problem with setting it to ten million if I have ten million > values > in the table and I am prepared to spend the resources to maintain those > statistics? That it'll probably take 10 million seconds to calculate the plans using it? I think Tom pointed there are a few places that are O(n^2) the number entries... Have a nice day, -- Martijn van Oosterhout <[EMAIL PROTECTED]> http://svana.org/kleptog/ > Please line up in a tree and maintain the heap invariant while > boarding. Thank you for flying nlogn airlines. signature.asc Description: Digital signature
[HACKERS] Maximum statistics target
Related to the concurrent discussion about selectivity estimations ... What is the reason the statistics target is limited to 1000? I've seen more than one case where increasing the statistics target to 1000 improved results and one would have wanted to increase it further. What's the problem with setting it to ten million if I have ten million values in the table and I am prepared to spend the resources to maintain those statistics? -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers