Re: thousands of CachedPlan entry per backend
On Thu, Jun 1, 2023 at 4:51 AM James Pang (chaolpan) wrote: > 2) from this line, we saw total 42 blocks ,215 chunks > CacheMemoryContext: 8737352 total in 42 blocks; 1021944 free (215 chunks); > 7715408 used, > > But from sum of it’s child level entrys, total sum(child lines) > block ,trunks show much more than “CacheMemoryContext, is expected to see > that? > Yes, that is expected. The parent context reports only its own direct memory usage and blocks. It does not include the sum of memory usage of its child contexts. Cheers, Jeff >
RE: thousands of CachedPlan entry per backend
these lines about "SPI Plan" are these PL/PGSQL functions related through SPI_prepare plan entry, right? SPI Plan: 1024 total in 1 blocks; 600 free (0 chunks); 424 used CachedPlan: 2048 total in 2 blocks; 304 free (1 chunks); 1744 used: xxx CachedPlanSource: 2048 total in 2 blocks; 200 free (0 chunks); 1848 used: xxx CachedPlanQuery: 2048 total in 2 blocks; 704 free (0 chunks); 1344 used From: Pavel Stehule Sent: Friday, June 2, 2023 12:57 PM To: James Pang (chaolpan) Cc: Laurenz Albe ; pgsql-performance@lists.postgresql.org Subject: Re: thousands of CachedPlan entry per backend pá 2. 6. 2023 v 3:45 odesílatel James Pang (chaolpan) mailto:chaol...@cisco.com>> napsal: these lines about "SPI Plan" are these PL/PGSQL functions related SPI_prepare plan entry, right? Possible to set a GUC to max(cached plan) per backend ? There is no limit for size of system cache. You can use pgbouncer that implicitly refresh session after 1 hour (and this limit can be reduced) Regards Pavel SPI Plan: 1024 total in 1 blocks; 600 free (0 chunks); 424 used CachedPlan: 2048 total in 2 blocks; 304 free (1 chunks); 1744 used: xxx CachedPlanSource: 2048 total in 2 blocks; 200 free (0 chunks); 1848 used: xxx CachedPlanQuery: 2048 total in 2 blocks; 704 free (0 chunks); 1344 used Thanks, James -Original Message- From: Laurenz Albe mailto:laurenz.a...@cybertec.at>> Sent: Thursday, June 1, 2023 8:48 PM To: James Pang (chaolpan) mailto:chaol...@cisco.com>>; Pavel Stehule mailto:pavel.steh...@gmail.com>> Cc: pgsql-performance@lists.postgresql.org<mailto:pgsql-performance@lists.postgresql.org> Subject: Re: thousands of CachedPlan entry per backend On Thu, 2023-06-01 at 08:50 +, James Pang (chaolpan) wrote: > we found thousands of cached plan , since JDBC driver only allow max > 256 cached prepared statements, how backend cache so many sql plans. > If we have one function, when application call that function will make > backend to cache every SQL statement plan in that function too? and for > table triggers, have similar caching behavior ? Yes, as long as the functions are written in PL/pgSQL. It only affects static SQL, that is, nothing that is run with EXECUTE. Yours, Laurenz Albe
Re: thousands of CachedPlan entry per backend
pá 2. 6. 2023 v 3:45 odesílatel James Pang (chaolpan) napsal: >these lines about "SPI Plan" are these PL/PGSQL functions related > SPI_prepare plan entry, right? Possible to set a GUC to max(cached plan) > per backend ? > There is no limit for size of system cache. You can use pgbouncer that implicitly refresh session after 1 hour (and this limit can be reduced) Regards Pavel > > SPI Plan: 1024 total in 1 blocks; 600 free (0 chunks); 424 used > CachedPlan: 2048 total in 2 blocks; 304 free (1 chunks); 1744 used: > xxx > CachedPlanSource: 2048 total in 2 blocks; 200 free (0 chunks); 1848 > used: xxx > CachedPlanQuery: 2048 total in 2 blocks; 704 free (0 chunks); 1344 > used > > Thanks, > > James > > -Original Message- > From: Laurenz Albe > Sent: Thursday, June 1, 2023 8:48 PM > To: James Pang (chaolpan) ; Pavel Stehule < > pavel.steh...@gmail.com> > Cc: pgsql-performance@lists.postgresql.org > Subject: Re: thousands of CachedPlan entry per backend > > On Thu, 2023-06-01 at 08:50 +, James Pang (chaolpan) wrote: > > we found thousands of cached plan , since JDBC driver only allow max > > 256 cached prepared statements, how backend cache so many sql plans. > > If we have one function, when application call that function will make > > backend to cache every SQL statement plan in that function too? and > for table triggers, have similar caching behavior ? > > Yes, as long as the functions are written in PL/pgSQL. > It only affects static SQL, that is, nothing that is run with EXECUTE. > > Yours, > Laurenz Albe >
RE: thousands of CachedPlan entry per backend
these lines about "SPI Plan" are these PL/PGSQL functions related SPI_prepare plan entry, right? Possible to set a GUC to max(cached plan) per backend ? SPI Plan: 1024 total in 1 blocks; 600 free (0 chunks); 424 used CachedPlan: 2048 total in 2 blocks; 304 free (1 chunks); 1744 used: xxx CachedPlanSource: 2048 total in 2 blocks; 200 free (0 chunks); 1848 used: xxx CachedPlanQuery: 2048 total in 2 blocks; 704 free (0 chunks); 1344 used Thanks, James -Original Message- From: Laurenz Albe Sent: Thursday, June 1, 2023 8:48 PM To: James Pang (chaolpan) ; Pavel Stehule Cc: pgsql-performance@lists.postgresql.org Subject: Re: thousands of CachedPlan entry per backend On Thu, 2023-06-01 at 08:50 +, James Pang (chaolpan) wrote: > we found thousands of cached plan , since JDBC driver only allow max > 256 cached prepared statements, how backend cache so many sql plans. > If we have one function, when application call that function will make > backend to cache every SQL statement plan in that function too? and for > table triggers, have similar caching behavior ? Yes, as long as the functions are written in PL/pgSQL. It only affects static SQL, that is, nothing that is run with EXECUTE. Yours, Laurenz Albe
Re: thousands of CachedPlan entry per backend
On Thu, 2023-06-01 at 08:50 +, James Pang (chaolpan) wrote: > we found thousands of cached plan , since JDBC driver only allow max 256 > cached > prepared statements, how backend cache so many sql plans. If we have one > function, > when application call that function will make backend to cache every SQL > statement > plan in that function too? and for table triggers, have similar caching > behavior ? Yes, as long as the functions are written in PL/pgSQL. It only affects static SQL, that is, nothing that is run with EXECUTE. Yours, Laurenz Albe
RE: thousands of CachedPlan entry per backend
Yes, too many cached metadata and we are thinking of a workaround to disconnect the sessions timely. In addition, based on the dumped memory context, I have questions 1) we found thousands of cached plan , since JDBC driver only allow max 256 cached prepared statements, how backend cache so many sql plans. If we have one function, when application call that function will make backend to cache every SQL statement plan in that function too? and for table triggers, have similar caching behavior ? 2) from this line, we saw total 42 blocks ,215 chunks CacheMemoryContext: 8737352 total in 42 blocks; 1021944 free (215 chunks); 7715408 used, But from sum of it’s child level entrys, total sum(child lines) block ,trunks show much more than “CacheMemoryContext, is expected to see that? Thanks, James From: Pavel Stehule Sent: Thursday, June 1, 2023 3:19 PM To: Laurenz Albe Cc: James Pang (chaolpan) ; pgsql-performance@lists.postgresql.org Subject: Re: thousands of CachedPlan entry per backend Hi čt 1. 6. 2023 v 8:53 odesílatel Laurenz Albe mailto:laurenz.a...@cybertec.at>> napsal: On Thu, 2023-06-01 at 03:36 +, James Pang (chaolpan) wrote: > PG V14.8-1 , client using Postgresql JDBC driver we found 40MB process memory > per > backend, from Operating system and memorycontext dump “Grand total:”, both > mached. > But from details, we found almost of entry belong to “CacheMemoryContext”, > from this line CacheMemoryContext: 8737352 total in 42 blocks; 1021944 free > (215 chunks); 7715408 used, > but there are thousands of lines of it’s child, the sum of blocks much more > than “8737352” total in 42 blocks > > Our application use Postgresql JDBC driver with default > parameters(maxprepared statement 256), > there are many triggers, functions in this database, and a few functions run > sql by an extension > pg_background. We have thousands of connections and have big concern why > have thousands of entrys > of cached SQL ? that will consume huge memory , anyway to limit the cached > plan entry to save memory > consumption? Or it looks like an abnormal behavior or bug to see so many > cached plan lines. If you have thousands of connections, that's your problem. You need effective connection pooling. Then 40MB per backend won't be a problem at all. Having thousands of connections will cause other, worse, problems for you. See for example https://www.cybertec-postgresql.com/en/tuning-max_connections-in-postgresql/ If you want to use functions, but don't want to benefit from plan caching, you can set the configuration parameter "plan_cache_mode" to "force_custom_plan". The problem with too big of cached metadata can be forced by too long sessions too. In this case it is good to throw a session (connect) after 1hour or maybe less. Regards Pavel Yours, Laurenz Albe
Re: thousands of CachedPlan entry per backend
Hi čt 1. 6. 2023 v 8:53 odesílatel Laurenz Albe napsal: > On Thu, 2023-06-01 at 03:36 +, James Pang (chaolpan) wrote: > > PG V14.8-1 , client using Postgresql JDBC driver we found 40MB process > memory per > > backend, from Operating system and memorycontext dump “Grand total:”, > both mached. > > But from details, we found almost of entry belong to > “CacheMemoryContext”, > > from this line CacheMemoryContext: 8737352 total in 42 blocks; 1021944 > free (215 chunks); 7715408 used, > > but there are thousands of lines of it’s child, the sum of blocks much > more than “8737352” total in 42 blocks > > > > Our application use Postgresql JDBC driver with default > parameters(maxprepared statement 256), > > there are many triggers, functions in this database, and a few functions > run sql by an extension > > pg_background. We have thousands of connections and have big concern > why have thousands of entrys > > of cached SQL ? that will consume huge memory , anyway to limit the > cached plan entry to save memory > > consumption? Or it looks like an abnormal behavior or bug to see so > many cached plan lines. > > If you have thousands of connections, that's your problem. You need > effective connection pooling. > Then 40MB per backend won't be a problem at all. Having thousands of > connections will cause > other, worse, problems for you. > > See for example > > https://www.cybertec-postgresql.com/en/tuning-max_connections-in-postgresql/ > > If you want to use functions, but don't want to benefit from plan caching, > you can set > the configuration parameter "plan_cache_mode" to "force_custom_plan". > The problem with too big of cached metadata can be forced by too long sessions too. In this case it is good to throw a session (connect) after 1hour or maybe less. Regards Pavel > > Yours, > Laurenz Albe > > >
Re: thousands of CachedPlan entry per backend
On Thu, 2023-06-01 at 03:36 +, James Pang (chaolpan) wrote: > PG V14.8-1 , client using Postgresql JDBC driver we found 40MB process memory > per > backend, from Operating system and memorycontext dump “Grand total:”, both > mached. > But from details, we found almost of entry belong to “CacheMemoryContext”, > from this line CacheMemoryContext: 8737352 total in 42 blocks; 1021944 free > (215 chunks); 7715408 used, > but there are thousands of lines of it’s child, the sum of blocks much more > than “8737352” total in 42 blocks > > Our application use Postgresql JDBC driver with default > parameters(maxprepared statement 256), > there are many triggers, functions in this database, and a few functions run > sql by an extension > pg_background. We have thousands of connections and have big concern why > have thousands of entrys > of cached SQL ? that will consume huge memory , anyway to limit the cached > plan entry to save memory > consumption? Or it looks like an abnormal behavior or bug to see so many > cached plan lines. If you have thousands of connections, that's your problem. You need effective connection pooling. Then 40MB per backend won't be a problem at all. Having thousands of connections will cause other, worse, problems for you. See for example https://www.cybertec-postgresql.com/en/tuning-max_connections-in-postgresql/ If you want to use functions, but don't want to benefit from plan caching, you can set the configuration parameter "plan_cache_mode" to "force_custom_plan". Yours, Laurenz Albe