Обсуждение: PREPARE / EXECUTE
First of all PREPARE/EXECUTE is a wonderful thing to speed up things significantly. I wonder if there is a way to store a parsed/rewritten/planned query in a table so that it can be loaded again. This might be useful when it comes to VERY complex queries (> 10 tables). I many applications the situation is like that: a. The user connects to the database. b. The user sends various different queries to the server (some might be the same) c. The user disconnects. If there was a way to store execution plans in a table the user could load the execution plans of the most time consuming stuff into the backend without parsing and optimizing it every time he authenticates. Does it sound useful to anybody? Is it possible to do it or are there some technical problems? Maybe this is worth thinking about. Hans -- *Cybertec Geschwinde u Schoenig* Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria Tel: +43/1/913 68 09; +43/664/233 90 75 www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at <http://cluster.postgresql.at>, www.cybertec.at <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at>
If you were using them that frequently, couldn't you just keep a persistent connection? If it's not used that often, wouldn't the overhead of preparing the query following a new connection become noise? Greg On Wed, 2002-10-23 at 09:24, Hans-Jürgen Schönig wrote: > First of all PREPARE/EXECUTE is a wonderful thing to speed up things > significantly. > I wonder if there is a way to store a parsed/rewritten/planned query in > a table so that it can be loaded again. > > This might be useful when it comes to VERY complex queries (> 10 tables). > I many applications the situation is like that: > > a. The user connects to the database. > b. The user sends various different queries to the server (some might be > the same) > c. The user disconnects. > > If there was a way to store execution plans in a table the user could > load the execution plans of the most time consuming stuff into the > backend without parsing and optimizing it every time he authenticates. > > Does it sound useful to anybody? Is it possible to do it or are there > some technical problems? > > Maybe this is worth thinking about. > > Hans > > -- > *Cybertec Geschwinde u Schoenig* > Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria > Tel: +43/1/913 68 09; +43/664/233 90 75 > www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at > <http://cluster.postgresql.at>, www.cybertec.at > <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at> > > > ---------------------------(end of broadcast)--------------------------- > TIP 6: Have you searched our list archives? > > http://archives.postgresql.org
Hans-Jürgen Schönig <postgres@cybertec.at> writes:
> I wonder if there is a way to store a parsed/rewritten/planned query in
> a table so that it can be loaded again.
The original version of the PREPARE patch used a shared-across-backends
cache for PREPAREd statements. We rejected that for a number of
reasons, one being the increased difficulty of keeping such a cache up
to date. I think actually storing the plans on disk would have all the
same problems, but worse.
regards, tom lane
The idea is not to have it accross multiple backends and having it in sync with the tables in the database. This is not the point. My problem is that I have seen many performance critical applications sending just a few complex queries to the server. The problem is: If you have many queries where the relation "time planner"/"time executor" is very high (eg. complex joins with just one value as the result). These applications stay the same for a long time (maybe even years) and so there is no need to worry about new tables and so forth - maybe there is not even a need to worry about new data. In these cases we could speed up the database significantly just by avoiding the use of the planner: An example: I have a join across 10 tables + 2 subselects across 4 tables on the machine I use for testing: planner: 12 seconds executor: 1 second The application will stay the same forever. I could be 10 times faster if there was a way to load the execution plan into the backend. There is no way to use a persistent connection (many clients on different machines, dynamic IPs, etc. ...) There is no way to have an "invalid" execution plan because there are no changes (new tables etc.) in the database. Also: If people execute a prepared query and it fails they will know why - queries will fail if people drop a table even if these queries are not prepared. A new feature like the one we are discussing might be used rarely but if people use it they will benefit A LOT. If we had a simple ASCII interface to load the stuff into the planner people could save MANY cycles. When talking about tuning it is nice to gain 10% or even 20% but in many cases it does not solve a problem - if a problem can be reduced by 90% it is a REAL gain. Gaining 10% can be done by tweaking the database a little - gaining 1000% cannot be done so it might be worth thinking about it even it the feature is only used by 20% of those users out there. 20% of all postgres users is most likely more than 15.000 people. Again; it is not supposed to be a every-day solution. It is a solution for applications staying the same for a very long time. Hans Tom Lane wrote: >Hans-Jürgen Schönig <postgres@cybertec.at> writes: > > >>I wonder if there is a way to store a parsed/rewritten/planned query in >>a table so that it can be loaded again. >> >> > >The original version of the PREPARE patch used a shared-across-backends >cache for PREPAREd statements. We rejected that for a number of >reasons, one being the increased difficulty of keeping such a cache up >to date. I think actually storing the plans on disk would have all the >same problems, but worse. > > regards, tom lane > >---------------------------(end of broadcast)--------------------------- >TIP 2: you can get off all lists at once with the unregister command > (send "unregister YourEmailAddressHere" to majordomo@postgresql.org) > > -- *Cybertec Geschwinde u Schoenig* Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria Tel: +43/1/913 68 09; +43/664/233 90 75 www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at <http://cluster.postgresql.at>, www.cybertec.at <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at>
On Wed, Oct 23, 2002 at 18:04:01 +0200, Hans-Jürgen Schönig <postgres@cybertec.at> wrote: > > An example: > I have a join across 10 tables + 2 subselects across 4 tables > on the machine I use for testing: > planner: 12 seconds > executor: 1 second > > The application will stay the same forever. > I could be 10 times faster if there was a way to load the execution plan > into the backend. One option you have is to explicitly give the join order. You can look at explain to see what order the joins are done in and then rewrite the sql to force them to be done in that order. This should keep things simple for the planner.
Bruno Wolff III <bruno@wolff.to> writes: > Hans-J�rgen Sch�nig <postgres@cybertec.at> wrote: >> I have a join across 10 tables + 2 subselects across 4 tables >> on the machine I use for testing: >> planner: 12 seconds >> executor: 1 second > One option you have is to explicitly give the join order. Yes, this is exactly the sort of situation where forcing the join order is a big performance win. See http://www.ca.postgresql.org/users-lounge/docs/7.2/postgres/explicit-joins.html regards, tom lane
This is exactly what we do in case of complex stuff. I know that it can help to reduce the problem for the planner. However: If you have explicit joins across 10 tables the SQL statement is not that readable any more and it is still slower than a prepared execution plan. I guess it is worth thinking about prepared plans somewhere on disk. Is there a way to transform ASCII -> plan? Hans Bruno Wolff III wrote: >On Wed, Oct 23, 2002 at 18:04:01 +0200, > Hans-Jürgen Schönig <postgres@cybertec.at> wrote: > > >>An example: >>I have a join across 10 tables + 2 subselects across 4 tables >>on the machine I use for testing: >> planner: 12 seconds >> executor: 1 second >> >>The application will stay the same forever. >>I could be 10 times faster if there was a way to load the execution plan >>into the backend. >> >> -- *Cybertec Geschwinde u Schoenig* Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria Tel: +43/1/913 68 09; +43/664/233 90 75 www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at <http://cluster.postgresql.at>, www.cybertec.at <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at>
Could you use some form of connection proxy where the proxy is actually keeping persistent connections but your application is making transient connections to the proxy? I believe this would result in the desired performance boost and behavior. Now, the next obvious question...anyone know of any proxy apps available for postgresql? Regards, Greg On Wed, 2002-10-23 at 11:04, Hans-Jürgen Schönig wrote: > The idea is not to have it accross multiple backends and having it in > sync with the tables in the database. This is not the point. > My problem is that I have seen many performance critical applications > sending just a few complex queries to the server. The problem is: If you > have many queries where the relation "time planner"/"time executor" is > very high (eg. complex joins with just one value as the result). > These applications stay the same for a long time (maybe even years) and > so there is no need to worry about new tables and so forth - maybe there > is not even a need to worry about new data. In these cases we could > speed up the database significantly just by avoiding the use of the planner: > > An example: > I have a join across 10 tables + 2 subselects across 4 tables > on the machine I use for testing: > planner: 12 seconds > executor: 1 second > > The application will stay the same forever. > I could be 10 times faster if there was a way to load the execution plan > into the backend. > There is no way to use a persistent connection (many clients on > different machines, dynamic IPs, etc. ...) > There is no way to have an "invalid" execution plan because there are no > changes (new tables etc.) in the database. > > Also: If people execute a prepared query and it fails they will know why > - queries will fail if people drop a table even if these queries are not > prepared. > A new feature like the one we are discussing might be used rarely but if > people use it they will benefit A LOT. > > If we had a simple ASCII interface to load the stuff into the planner > people could save MANY cycles. > When talking about tuning it is nice to gain 10% or even 20% but in many > cases it does not solve a problem - if a problem can be reduced by 90% > it is a REAL gain. > Gaining 10% can be done by tweaking the database a little - gaining > 1000% cannot be done so it might be worth thinking about it even it the > feature is only used by 20% of those users out there. 20% of all > postgres users is most likely more than 15.000 people. > > Again; it is not supposed to be a every-day solution. It is a solution > for applications staying the same for a very long time. > > Hans > > > Tom Lane wrote: > > >Hans-Jürgen Schönig <postgres@cybertec.at> writes: > > > > > >>I wonder if there is a way to store a parsed/rewritten/planned query in > >>a table so that it can be loaded again. > >> > >> > > > >The original version of the PREPARE patch used a shared-across-backends > >cache for PREPAREd statements. We rejected that for a number of > >reasons, one being the increased difficulty of keeping such a cache up > >to date. I think actually storing the plans on disk would have all the > >same problems, but worse. > > > > regards, tom lane > > > >---------------------------(end of broadcast)--------------------------- > >TIP 2: you can get off all lists at once with the unregister command > > (send "unregister YourEmailAddressHere" to majordomo@postgresql.org) > > > > > > > -- > *Cybertec Geschwinde u Schoenig* > Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria > Tel: +43/1/913 68 09; +43/664/233 90 75 > www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at > <http://cluster.postgresql.at>, www.cybertec.at > <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at> > > > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Don't 'kill -9' the postmaster
Greg Copeland wrote: >Could you use some form of connection proxy where the proxy is actually >keeping persistent connections but your application is making transient >connections to the proxy? I believe this would result in the desired >performance boost and behavior. > >Now, the next obvious question...anyone know of any proxy apps available >for postgresql? > >Regards, > > Greg > > > There is one load balancing software available which is based on the ACE library. Just have a look at: http://freshmeat.net/projects/dbbalancer/ I haven't tested it up to now. I am now looking for a workaround - I am sure that there are many workarounds for this issue (explicit joins, persistent connections, etc. ...). I thought it might be useful to have something like a data type (or maybe a binary field) used to store execution plans. People could use this feature as some sort of "server side" function or so ... It can be seend as some sort of optimized function in the backend which can be loaded/executed more efficiently. Maybe others would like to see that feature as well. Hans -- *Cybertec Geschwinde u Schoenig* Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria Tel: +43/1/913 68 09; +43/664/233 90 75 www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at <http://cluster.postgresql.at>, www.cybertec.at <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at>
> The idea is not to have it accross multiple backends and having it in > sync with the tables in the database. This is not the point. > My problem is that I have seen many performance critical applications > sending just a few complex queries to the server. The problem is: If you > have many queries where the relation "time planner"/"time executor" is > very high (eg. complex joins with just one value as the result). The standard approach to such a scenario would imho be to write stored procedures for the complex queries (e.g. plpgsql) and use that from the client. Maybe even eliminate a few ping pongs between client and server. Andreas
"Zeugswetter Andreas SB SD" <ZeugswetterA@spardat.at> writes: > The standard approach to such a scenario would imho be to write > stored procedures for the complex queries (e.g. plpgsql) and use > that from the client. Maybe even eliminate a few ping pongs between > client and server. Since PL/PgSQL cached query plans are flushed when the backend exits, how would this help? Regarding the original suggestion of storing prepared plans on disk, I agree with Tom -- it's basically the same idea as storing plans in shared memory, which we previously considered (and Karel implemented), but ultimately decided to remove. IMHO, the utility of this feature doesn't justify the problems that would come with implementing it (see the archives for the original implementation discussions). Cheers, Neil -- Neil Conway <neilc@samurai.com> || PGP Key ID: DB3C29FC
On Wed, Oct 23, 2002 at 11:02:14AM -0400, Tom Lane wrote: > Hans-Jürgen Schönig <postgres@cybertec.at> writes: > > I wonder if there is a way to store a parsed/rewritten/planned query in > > a table so that it can be loaded again. > > The original version of the PREPARE patch used a shared-across-backends > cache for PREPAREd statements. We rejected that for a number of > reasons, one being the increased difficulty of keeping such a cache up > to date. I think actually storing the plans on disk would have all the > same problems, but worse. Right. There's solution: persisten backend (for example like classicapache). This solve problem with lifetime of all persistentcaches.It's already in TODO. Karel -- Karel Zak <zakkr@zf.jcu.cz>http://home.zf.jcu.cz/~zakkr/C, PostgreSQL, PHP, WWW, http://docs.linux.cz, http://mape.jcu.cz
On Wed, 2002-10-23 at 10:39, Greg Copeland wrote: > If you were using them that frequently, couldn't you just keep a > persistent connection? If it's not used that often, wouldn't the > overhead of preparing the query following a new connection become noise? Especially by the time you add in the dependency tracking (drop table, query must go), and modifications to analyze to clear out the stored list. > On Wed, 2002-10-23 at 09:24, Hans-Jürgen Schönig wrote: > > First of all PREPARE/EXECUTE is a wonderful thing to speed up things > > significantly. > > I wonder if there is a way to store a parsed/rewritten/planned query in > > a table so that it can be loaded again. > > > > This might be useful when it comes to VERY complex queries (> 10 tables). > > I many applications the situation is like that: > > > > a. The user connects to the database. > > b. The user sends various different queries to the server (some might be > > the same) > > c. The user disconnects. > > > > If there was a way to store execution plans in a table the user could > > load the execution plans of the most time consuming stuff into the > > backend without parsing and optimizing it every time he authenticates. > > > > Does it sound useful to anybody? Is it possible to do it or are there > > some technical problems? > > > > Maybe this is worth thinking about. > > > > Hans > > > > -- > > *Cybertec Geschwinde u Schoenig* > > Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria > > Tel: +43/1/913 68 09; +43/664/233 90 75 > > www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at > > <http://cluster.postgresql.at>, www.cybertec.at > > <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at> > > > > > > ---------------------------(end of broadcast)--------------------------- > > TIP 6: Have you searched our list archives? > > > > http://archives.postgresql.org > > > > ---------------------------(end of broadcast)--------------------------- > TIP 5: Have you checked our extensive FAQ? > > http://www.postgresql.org/users-lounge/docs/faq.html > -- Rod Taylor
> > > >The standard approach to such a scenario would imho be to write stored procedures >for the complex queries (e.g. plpgsql) and use that from the client. >Maybe even eliminate a few ping pongs between client and server. > >Andreas > > Does it reduce the time taken by the planner? Are server side SQL functions optimized at runtime or at "create function" time? If the function is optimized at runtime it is not a gain. Hans -- *Cybertec Geschwinde u Schoenig* Ludo-Hartmannplatz 1/14, A-1160 Vienna, Austria Tel: +43/1/913 68 09; +43/664/233 90 75 www.postgresql.at <http://www.postgresql.at>, cluster.postgresql.at <http://cluster.postgresql.at>, www.cybertec.at <http://www.cybertec.at>, kernel.cybertec.at <http://kernel.cybertec.at>