Re: P3 and concurrenty

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
11 messages Options
Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Sven Van Caekenberghe-2
Hi Petter,

[ CC-ing the Pharo Users list ]

P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.

Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.

Is that what you are doing ?

If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:

That will then serialise requests and possibly block one onto the other.

HTH,

Sven

PS: another thing to take care of if closing your sql connections when the session is no longer needed.

PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.

> On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
>
> Hi Sven
>
> We are using Pharo as our backend in a project and we have run into a problem with P3.
>
> The problem seems to be connected to compiled sql statements and concurrency.
>
> We keep getting this error: Bindcomplete message expected
>
> Problem seems to be easy to reproduce:
>
> 1) Compile any sql statement
> 2) Use this statement in a query twice (!) in a teapot endpoint
>
> The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
>
> One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
>
> If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
>
> My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
>
> Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
>
> Best regards
>
> Petter Egesund (I wrote the heysql-package based on P3)


Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Petter
Hi Sven and thanks for answering!

I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.

I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?

It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time,. Creating a pool sounds like the right solution to me now - any meaning about this?

Petter


On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
Hi Petter,

[ CC-ing the Pharo Users list ]

P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.

Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.

Is that what you are doing ?

If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:

That will then serialise requests and possibly block one onto the other.

HTH,

Sven

PS: another thing to take care of if closing your sql connections when the session is no longer needed.

PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.

> On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
>
> Hi Sven
>
> We are using Pharo as our backend in a project and we have run into a problem with P3.
>
> The problem seems to be connected to compiled sql statements and concurrency.
>
> We keep getting this error: Bindcomplete message expected
>
> Problem seems to be easy to reproduce:
>
> 1) Compile any sql statement
> 2) Use this statement in a query twice (!) in a teapot endpoint
>
> The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
>
> One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
>
> If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
>
> My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
>
> Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
>
> Best regards
>
> Petter Egesund (I wrote the heysql-package based on P3)

Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Sven Van Caekenberghe-2
Hi Petter,

> On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
>
> Hi Sven and thanks for answering!
>
> I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.

It is useable, but not concurrently.

> I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?

Yes and no, see further.

> It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.

Yes, follow the pointer that I gave you, it is not hard.

> Creating a pool sounds like the right solution to me now - any meaning about this?

Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...

I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.

Running

  P3ClientTest new runBenchmark1Bench.

on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.

If I modify this slightly to use a new client/connection each time, like this

  [
    (P3Client url: 'psql://sven@localhost')
      query: 'SELECT * FROM benchmark1';
      close
  ] benchFor: 5 seconds

I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.

And note that this is not using prepared statements.

So I would start by opening/closing a connection each time you need it.

HTH,

Sven

> Petter
>
>
> On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> [ CC-ing the Pharo Users list ]
>
> P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
>
> Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
>
> Is that what you are doing ?
>
> If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
>
> That will then serialise requests and possibly block one onto the other.
>
> HTH,
>
> Sven
>
> PS: another thing to take care of if closing your sql connections when the session is no longer needed.
>
> PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
>
> > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> >
> > Hi Sven
> >
> > We are using Pharo as our backend in a project and we have run into a problem with P3.
> >
> > The problem seems to be connected to compiled sql statements and concurrency.
> >
> > We keep getting this error: Bindcomplete message expected
> >
> > Problem seems to be easy to reproduce:
> >
> > 1) Compile any sql statement
> > 2) Use this statement in a query twice (!) in a teapot endpoint
> >
> > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> >
> > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> >
> > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> >
> > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> >
> > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> >
> > Best regards
> >
> > Petter Egesund (I wrote the heysql-package based on P3)
>


Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Petter
Yes, thanks for good feedback.

I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.

Petter

On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
Hi Petter,

> On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
>
> Hi Sven and thanks for answering!
>
> I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.

It is useable, but not concurrently.

> I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?

Yes and no, see further.

> It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.

Yes, follow the pointer that I gave you, it is not hard.

> Creating a pool sounds like the right solution to me now - any meaning about this?

Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...

I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.

Running

  P3ClientTest new runBenchmark1Bench.

on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.

If I modify this slightly to use a new client/connection each time, like this

  [
    (P3Client url: 'psql://sven@localhost')
      query: 'SELECT * FROM benchmark1';
      close
  ] benchFor: 5 seconds

I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.

And note that this is not using prepared statements.

So I would start by opening/closing a connection each time you need it.

HTH,

Sven

> Petter
>
>
> On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> [ CC-ing the Pharo Users list ]
>
> P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
>
> Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
>
> Is that what you are doing ?
>
> If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
>
> That will then serialise requests and possibly block one onto the other.
>
> HTH,
>
> Sven
>
> PS: another thing to take care of if closing your sql connections when the session is no longer needed.
>
> PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
>
> > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> >
> > Hi Sven
> >
> > We are using Pharo as our backend in a project and we have run into a problem with P3.
> >
> > The problem seems to be connected to compiled sql statements and concurrency.
> >
> > We keep getting this error: Bindcomplete message expected
> >
> > Problem seems to be easy to reproduce:
> >
> > 1) Compile any sql statement
> > 2) Use this statement in a query twice (!) in a teapot endpoint
> >
> > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> >
> > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> >
> > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> >
> > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> >
> > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> >
> > Best regards
> >
> > Petter Egesund (I wrote the heysql-package based on P3)
>

Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Sven Van Caekenberghe-2
Hi Petter,

https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.

Here is the class comment:

======

I am P3ConnectionPool.

I offer a pool of shared PSQL connections (P3Client instances) as a resource.

After configuring me with at least a url, you use me by calling #withConnection:

  pool := P3ConnectionPool url: 'psql://sven@localhost'.
  pool withConnection: [ :p3Client |
          p3Client query: 'SELECT table_name FROM information_schema.tables' ].
  pool close.

When a connection is in use, it is not part of the pool.
When a connection is returned/released to the pool, it becomes available for reuse.

The pool's size is the number of open connection ready to be reused.
The pool's capacity is the maximum number of connection that will be pooled.
Excess connections will be closed when they are returned/released to the pool.

New connections are created as needed.
You can set a #configurator to further initialize new connections.
You can use #warmUp or #warmUp: to precreate a number of connections.

When an error occurs, the connection should not be reused and be closed by the caller.

======

There are some unit tests as well.

Let me know if this works for you.

Sven

> On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
>
> Yes, thanks for good feedback.
>
> I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
>
> Petter
>
> On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> >
> > Hi Sven and thanks for answering!
> >
> > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
>
> It is useable, but not concurrently.
>
> > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
>
> Yes and no, see further.
>
> > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
>
> Yes, follow the pointer that I gave you, it is not hard.
>
> > Creating a pool sounds like the right solution to me now - any meaning about this?
>
> Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
>
> I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
>
> Running
>
>   P3ClientTest new runBenchmark1Bench.
>
> on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
>
> If I modify this slightly to use a new client/connection each time, like this
>
>   [
>     (P3Client url: 'psql://sven@localhost')
>       query: 'SELECT * FROM benchmark1';
>       close
>   ] benchFor: 5 seconds
>
> I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
>
> And note that this is not using prepared statements.
>
> So I would start by opening/closing a connection each time you need it.
>
> HTH,
>
> Sven
>
> > Petter
> >
> >
> > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > [ CC-ing the Pharo Users list ]
> >
> > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> >
> > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> >
> > Is that what you are doing ?
> >
> > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> >
> > That will then serialise requests and possibly block one onto the other.
> >
> > HTH,
> >
> > Sven
> >
> > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> >
> > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> >
> > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > >
> > > Hi Sven
> > >
> > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > >
> > > The problem seems to be connected to compiled sql statements and concurrency.
> > >
> > > We keep getting this error: Bindcomplete message expected
> > >
> > > Problem seems to be easy to reproduce:
> > >
> > > 1) Compile any sql statement
> > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > >
> > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > >
> > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > >
> > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > >
> > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > >
> > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > >
> > > Best regards
> > >
> > > Petter Egesund (I wrote the heysql-package based on P3)
> >
>


Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Petter
Hi and thanks for an answer above my expectations to our problem. 

This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)

I will try out and get back as soon as we have tested.

One quick question - are the prepared statements which are connected to the pool thread safe?

Petter

On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
Hi Petter,

https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.

Here is the class comment:

======

I am P3ConnectionPool.

I offer a pool of shared PSQL connections (P3Client instances) as a resource.

After configuring me with at least a url, you use me by calling #withConnection:

  pool := P3ConnectionPool url: 'psql://sven@localhost'.
  pool withConnection: [ :p3Client |
          p3Client query: 'SELECT table_name FROM information_schema.tables' ].
  pool close.

When a connection is in use, it is not part of the pool.
When a connection is returned/released to the pool, it becomes available for reuse.

The pool's size is the number of open connection ready to be reused.
The pool's capacity is the maximum number of connection that will be pooled.
Excess connections will be closed when they are returned/released to the pool.

New connections are created as needed.
You can set a #configurator to further initialize new connections.
You can use #warmUp or #warmUp: to precreate a number of connections.

When an error occurs, the connection should not be reused and be closed by the caller.

======

There are some unit tests as well.

Let me know if this works for you.

Sven

> On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
>
> Yes, thanks for good feedback.
>
> I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
>
> Petter
>
> On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> >
> > Hi Sven and thanks for answering!
> >
> > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
>
> It is useable, but not concurrently.
>
> > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
>
> Yes and no, see further.
>
> > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
>
> Yes, follow the pointer that I gave you, it is not hard.
>
> > Creating a pool sounds like the right solution to me now - any meaning about this?
>
> Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
>
> I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
>
> Running
>
>   P3ClientTest new runBenchmark1Bench.
>
> on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
>
> If I modify this slightly to use a new client/connection each time, like this
>
>   [
>     (P3Client url: 'psql://sven@localhost')
>       query: 'SELECT * FROM benchmark1';
>       close
>   ] benchFor: 5 seconds
>
> I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
>
> And note that this is not using prepared statements.
>
> So I would start by opening/closing a connection each time you need it.
>
> HTH,
>
> Sven
>
> > Petter
> >
> >
> > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > [ CC-ing the Pharo Users list ]
> >
> > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> >
> > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> >
> > Is that what you are doing ?
> >
> > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> >
> > That will then serialise requests and possibly block one onto the other.
> >
> > HTH,
> >
> > Sven
> >
> > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> >
> > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> >
> > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > >
> > > Hi Sven
> > >
> > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > >
> > > The problem seems to be connected to compiled sql statements and concurrency.
> > >
> > > We keep getting this error: Bindcomplete message expected
> > >
> > > Problem seems to be easy to reproduce:
> > >
> > > 1) Compile any sql statement
> > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > >
> > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > >
> > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > >
> > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > >
> > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > >
> > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > >
> > > Best regards
> > >
> > > Petter Egesund (I wrote the heysql-package based on P3)
> >
>

Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Sven Van Caekenberghe-2


> On 10 Feb 2020, at 16:30, Petter Egesund <[hidden email]> wrote:
>
> Hi and thanks for an answer above my expectations to our problem.
>
> This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)
>
> I will try out and get back as soon as we have tested.
>
> One quick question - are the prepared statements which are connected to the pool thread safe?

Prepared statements are scoped to a single connection in PSQL, AFAIK.

What you will probably want to do is use a #configurator: block to set up your prepared statements once, like you do now. Then each thread/process grabs its own connection from the pool, (which will be already configured properly) and uses it, for itself, and finally returns it to the pool.

Interacting with the pool should be fully thread safe. Once you are using a connection, it not longer is, but is also does not have to, since one connection equals one client thread.

> Petter
>
> On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.
>
> Here is the class comment:
>
> ======
>
> I am P3ConnectionPool.
>
> I offer a pool of shared PSQL connections (P3Client instances) as a resource.
>
> After configuring me with at least a url, you use me by calling #withConnection:
>
>   pool := P3ConnectionPool url: 'psql://sven@localhost'.
>   pool withConnection: [ :p3Client |
>           p3Client query: 'SELECT table_name FROM information_schema.tables' ].
>   pool close.
>
> When a connection is in use, it is not part of the pool.
> When a connection is returned/released to the pool, it becomes available for reuse.
>
> The pool's size is the number of open connection ready to be reused.
> The pool's capacity is the maximum number of connection that will be pooled.
> Excess connections will be closed when they are returned/released to the pool.
>
> New connections are created as needed.
> You can set a #configurator to further initialize new connections.
> You can use #warmUp or #warmUp: to precreate a number of connections.
>
> When an error occurs, the connection should not be reused and be closed by the caller.
>
> ======
>
> There are some unit tests as well.
>
> Let me know if this works for you.
>
> Sven
>
> > On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
> >
> > Yes, thanks for good feedback.
> >
> > I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
> >
> > Petter
> >
> > On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> > >
> > > Hi Sven and thanks for answering!
> > >
> > > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
> >
> > It is useable, but not concurrently.
> >
> > > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
> >
> > Yes and no, see further.
> >
> > > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
> >
> > Yes, follow the pointer that I gave you, it is not hard.
> >
> > > Creating a pool sounds like the right solution to me now - any meaning about this?
> >
> > Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
> >
> > I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
> >
> > Running
> >
> >   P3ClientTest new runBenchmark1Bench.
> >
> > on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
> >
> > If I modify this slightly to use a new client/connection each time, like this
> >
> >   [
> >     (P3Client url: 'psql://sven@localhost')
> >       query: 'SELECT * FROM benchmark1';
> >       close
> >   ] benchFor: 5 seconds
> >
> > I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
> >
> > And note that this is not using prepared statements.
> >
> > So I would start by opening/closing a connection each time you need it.
> >
> > HTH,
> >
> > Sven
> >
> > > Petter
> > >
> > >
> > > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > Hi Petter,
> > >
> > > [ CC-ing the Pharo Users list ]
> > >
> > > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> > >
> > > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> > >
> > > Is that what you are doing ?
> > >
> > > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> > >
> > > That will then serialise requests and possibly block one onto the other.
> > >
> > > HTH,
> > >
> > > Sven
> > >
> > > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> > >
> > > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> > >
> > > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > > >
> > > > Hi Sven
> > > >
> > > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > > >
> > > > The problem seems to be connected to compiled sql statements and concurrency.
> > > >
> > > > We keep getting this error: Bindcomplete message expected
> > > >
> > > > Problem seems to be easy to reproduce:
> > > >
> > > > 1) Compile any sql statement
> > > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > > >
> > > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > > >
> > > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > > >
> > > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > > >
> > > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > > >
> > > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > > >
> > > > Best regards
> > > >
> > > > Petter Egesund (I wrote the heysql-package based on P3)
> > >
> >
>


Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Petter
Yes, I see - that sounds sensible.

Could it be an idea to subclass the pool-class and to prepare statements in this method, if there is a kind of init-method in the class? What do you think?

Petter

On Mon, Feb 10, 2020 at 4:36 PM Sven Van Caekenberghe <[hidden email]> wrote:


> On 10 Feb 2020, at 16:30, Petter Egesund <[hidden email]> wrote:
>
> Hi and thanks for an answer above my expectations to our problem.
>
> This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)
>
> I will try out and get back as soon as we have tested.
>
> One quick question - are the prepared statements which are connected to the pool thread safe?

Prepared statements are scoped to a single connection in PSQL, AFAIK.

What you will probably want to do is use a #configurator: block to set up your prepared statements once, like you do now. Then each thread/process grabs its own connection from the pool, (which will be already configured properly) and uses it, for itself, and finally returns it to the pool.

Interacting with the pool should be fully thread safe. Once you are using a connection, it not longer is, but is also does not have to, since one connection equals one client thread.

> Petter
>
> On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
> Hi Petter,
>
> https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.
>
> Here is the class comment:
>
> ======
>
> I am P3ConnectionPool.
>
> I offer a pool of shared PSQL connections (P3Client instances) as a resource.
>
> After configuring me with at least a url, you use me by calling #withConnection:
>
>   pool := P3ConnectionPool url: 'psql://sven@localhost'.
>   pool withConnection: [ :p3Client |
>           p3Client query: 'SELECT table_name FROM information_schema.tables' ].
>   pool close.
>
> When a connection is in use, it is not part of the pool.
> When a connection is returned/released to the pool, it becomes available for reuse.
>
> The pool's size is the number of open connection ready to be reused.
> The pool's capacity is the maximum number of connection that will be pooled.
> Excess connections will be closed when they are returned/released to the pool.
>
> New connections are created as needed.
> You can set a #configurator to further initialize new connections.
> You can use #warmUp or #warmUp: to precreate a number of connections.
>
> When an error occurs, the connection should not be reused and be closed by the caller.
>
> ======
>
> There are some unit tests as well.
>
> Let me know if this works for you.
>
> Sven
>
> > On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
> >
> > Yes, thanks for good feedback.
> >
> > I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
> >
> > Petter
> >
> > On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> > >
> > > Hi Sven and thanks for answering!
> > >
> > > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
> >
> > It is useable, but not concurrently.
> >
> > > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
> >
> > Yes and no, see further.
> >
> > > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
> >
> > Yes, follow the pointer that I gave you, it is not hard.
> >
> > > Creating a pool sounds like the right solution to me now - any meaning about this?
> >
> > Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
> >
> > I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
> >
> > Running
> >
> >   P3ClientTest new runBenchmark1Bench.
> >
> > on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
> >
> > If I modify this slightly to use a new client/connection each time, like this
> >
> >   [
> >     (P3Client url: 'psql://sven@localhost')
> >       query: 'SELECT * FROM benchmark1';
> >       close
> >   ] benchFor: 5 seconds
> >
> > I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
> >
> > And note that this is not using prepared statements.
> >
> > So I would start by opening/closing a connection each time you need it.
> >
> > HTH,
> >
> > Sven
> >
> > > Petter
> > >
> > >
> > > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > Hi Petter,
> > >
> > > [ CC-ing the Pharo Users list ]
> > >
> > > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> > >
> > > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> > >
> > > Is that what you are doing ?
> > >
> > > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> > >
> > > That will then serialise requests and possibly block one onto the other.
> > >
> > > HTH,
> > >
> > > Sven
> > >
> > > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> > >
> > > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> > >
> > > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > > >
> > > > Hi Sven
> > > >
> > > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > > >
> > > > The problem seems to be connected to compiled sql statements and concurrency.
> > > >
> > > > We keep getting this error: Bindcomplete message expected
> > > >
> > > > Problem seems to be easy to reproduce:
> > > >
> > > > 1) Compile any sql statement
> > > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > > >
> > > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > > >
> > > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > > >
> > > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > > >
> > > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > > >
> > > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > > >
> > > > Best regards
> > > >
> > > > Petter Egesund (I wrote the heysql-package based on P3)
> > >
> >
>

Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Sven Van Caekenberghe-2
Hi Petter,

> On 10 Feb 2020, at 16:46, Petter Egesund <[hidden email]> wrote:
>
> Yes, I see - that sounds sensible.
>
> Could it be an idea to subclass the pool-class and to prepare statements in this method, if there is a kind of init-method in the class? What do you think?

That would have indeed been one approach but for this situation I opted for a configurator block.

P3ConnectionPool>>#configurator: oneArgumentBlock
  "Set oneArgumentBlock to be my configurator.
   This is code that will be executed once on each newly created connection (P3Client).
   The default configuration asserts #isWorking on the argument"

The default is

  configurator := [ :p3Client | self assert: p3Client isWorking ]

which basically does an extra validation and ensures the opening of the connection.

Now, a little piece of functionality was still missing, I added it with the following commit:

  https://github.com/svenvc/P3/commit/e3f161d9f9ae34b3b020ad924b3c0f116f68c0b0

Your usage scenario should now be like this: you create a P3ConnectionPool with the proper URL, then you set your configurator to add your prepared statements (you could delegate this to a helper class if you have a lot of them).

  pool configurator: [ :p3Client |
    p3Client prepare: 'SELECT ...' named: 'select1'.
    p3Client prepare: 'INSET ...' named: 'insert1' ]

Later, when using a connection from the pool, you refer to your prepared statements by name.

  pool withConnection: [ :p3Client |
    (p3Client preparedStatementNamed: 'select1')
      execute: { args } ]

Sven

PS: have a look at P3PreparedStatementTest>>#testNamedPreparedStatement for an operation example (minus the connection pooling).

> On Mon, Feb 10, 2020 at 4:36 PM Sven Van Caekenberghe <[hidden email]> wrote:
>
>
> > On 10 Feb 2020, at 16:30, Petter Egesund <[hidden email]> wrote:
> >
> > Hi and thanks for an answer above my expectations to our problem.
> >
> > This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)
> >
> > I will try out and get back as soon as we have tested.
> >
> > One quick question - are the prepared statements which are connected to the pool thread safe?
>
> Prepared statements are scoped to a single connection in PSQL, AFAIK.
>
> What you will probably want to do is use a #configurator: block to set up your prepared statements once, like you do now. Then each thread/process grabs its own connection from the pool, (which will be already configured properly) and uses it, for itself, and finally returns it to the pool.
>
> Interacting with the pool should be fully thread safe. Once you are using a connection, it not longer is, but is also does not have to, since one connection equals one client thread.
>
> > Petter
> >
> > On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.
> >
> > Here is the class comment:
> >
> > ======
> >
> > I am P3ConnectionPool.
> >
> > I offer a pool of shared PSQL connections (P3Client instances) as a resource.
> >
> > After configuring me with at least a url, you use me by calling #withConnection:
> >
> >   pool := P3ConnectionPool url: 'psql://sven@localhost'.
> >   pool withConnection: [ :p3Client |
> >           p3Client query: 'SELECT table_name FROM information_schema.tables' ].
> >   pool close.
> >
> > When a connection is in use, it is not part of the pool.
> > When a connection is returned/released to the pool, it becomes available for reuse.
> >
> > The pool's size is the number of open connection ready to be reused.
> > The pool's capacity is the maximum number of connection that will be pooled.
> > Excess connections will be closed when they are returned/released to the pool.
> >
> > New connections are created as needed.
> > You can set a #configurator to further initialize new connections.
> > You can use #warmUp or #warmUp: to precreate a number of connections.
> >
> > When an error occurs, the connection should not be reused and be closed by the caller.
> >
> > ======
> >
> > There are some unit tests as well.
> >
> > Let me know if this works for you.
> >
> > Sven
> >
> > > On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
> > >
> > > Yes, thanks for good feedback.
> > >
> > > I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
> > >
> > > Petter
> > >
> > > On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > Hi Petter,
> > >
> > > > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> > > >
> > > > Hi Sven and thanks for answering!
> > > >
> > > > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
> > >
> > > It is useable, but not concurrently.
> > >
> > > > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
> > >
> > > Yes and no, see further.
> > >
> > > > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
> > >
> > > Yes, follow the pointer that I gave you, it is not hard.
> > >
> > > > Creating a pool sounds like the right solution to me now - any meaning about this?
> > >
> > > Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
> > >
> > > I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
> > >
> > > Running
> > >
> > >   P3ClientTest new runBenchmark1Bench.
> > >
> > > on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
> > >
> > > If I modify this slightly to use a new client/connection each time, like this
> > >
> > >   [
> > >     (P3Client url: 'psql://sven@localhost')
> > >       query: 'SELECT * FROM benchmark1';
> > >       close
> > >   ] benchFor: 5 seconds
> > >
> > > I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
> > >
> > > And note that this is not using prepared statements.
> > >
> > > So I would start by opening/closing a connection each time you need it.
> > >
> > > HTH,
> > >
> > > Sven
> > >
> > > > Petter
> > > >
> > > >
> > > > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > > Hi Petter,
> > > >
> > > > [ CC-ing the Pharo Users list ]
> > > >
> > > > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> > > >
> > > > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> > > >
> > > > Is that what you are doing ?
> > > >
> > > > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> > > >
> > > > That will then serialise requests and possibly block one onto the other.
> > > >
> > > > HTH,
> > > >
> > > > Sven
> > > >
> > > > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> > > >
> > > > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> > > >
> > > > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > > > >
> > > > > Hi Sven
> > > > >
> > > > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > > > >
> > > > > The problem seems to be connected to compiled sql statements and concurrency.
> > > > >
> > > > > We keep getting this error: Bindcomplete message expected
> > > > >
> > > > > Problem seems to be easy to reproduce:
> > > > >
> > > > > 1) Compile any sql statement
> > > > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > > > >
> > > > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > > > >
> > > > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > > > >
> > > > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > > > >
> > > > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > > > >
> > > > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > > > >
> > > > > Best regards
> > > > >
> > > > > Petter Egesund (I wrote the heysql-package based on P3)
> > > >
> > >
> >
>


Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

Petter
Perfect, this is exactly what I needed. Thank so much!!

Petter

On Mon, Feb 10, 2020 at 8:44 PM Sven Van Caekenberghe <[hidden email]> wrote:
Hi Petter,

> On 10 Feb 2020, at 16:46, Petter Egesund <[hidden email]> wrote:
>
> Yes, I see - that sounds sensible.
>
> Could it be an idea to subclass the pool-class and to prepare statements in this method, if there is a kind of init-method in the class? What do you think?

That would have indeed been one approach but for this situation I opted for a configurator block.

P3ConnectionPool>>#configurator: oneArgumentBlock
  "Set oneArgumentBlock to be my configurator.
   This is code that will be executed once on each newly created connection (P3Client).
   The default configuration asserts #isWorking on the argument"

The default is

  configurator := [ :p3Client | self assert: p3Client isWorking ]

which basically does an extra validation and ensures the opening of the connection.

Now, a little piece of functionality was still missing, I added it with the following commit:

  https://github.com/svenvc/P3/commit/e3f161d9f9ae34b3b020ad924b3c0f116f68c0b0

Your usage scenario should now be like this: you create a P3ConnectionPool with the proper URL, then you set your configurator to add your prepared statements (you could delegate this to a helper class if you have a lot of them).

  pool configurator: [ :p3Client |
    p3Client prepare: 'SELECT ...' named: 'select1'.
    p3Client prepare: 'INSET ...' named: 'insert1' ]

Later, when using a connection from the pool, you refer to your prepared statements by name.

  pool withConnection: [ :p3Client |
    (p3Client preparedStatementNamed: 'select1')
      execute: { args } ]

Sven

PS: have a look at P3PreparedStatementTest>>#testNamedPreparedStatement for an operation example (minus the connection pooling).

> On Mon, Feb 10, 2020 at 4:36 PM Sven Van Caekenberghe <[hidden email]> wrote:
>
>
> > On 10 Feb 2020, at 16:30, Petter Egesund <[hidden email]> wrote:
> >
> > Hi and thanks for an answer above my expectations to our problem.
> >
> > This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)
> >
> > I will try out and get back as soon as we have tested.
> >
> > One quick question - are the prepared statements which are connected to the pool thread safe?
>
> Prepared statements are scoped to a single connection in PSQL, AFAIK.
>
> What you will probably want to do is use a #configurator: block to set up your prepared statements once, like you do now. Then each thread/process grabs its own connection from the pool, (which will be already configured properly) and uses it, for itself, and finally returns it to the pool.
>
> Interacting with the pool should be fully thread safe. Once you are using a connection, it not longer is, but is also does not have to, since one connection equals one client thread.
>
> > Petter
> >
> > On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.
> >
> > Here is the class comment:
> >
> > ======
> >
> > I am P3ConnectionPool.
> >
> > I offer a pool of shared PSQL connections (P3Client instances) as a resource.
> >
> > After configuring me with at least a url, you use me by calling #withConnection:
> >
> >   pool := P3ConnectionPool url: 'psql://sven@localhost'.
> >   pool withConnection: [ :p3Client |
> >           p3Client query: 'SELECT table_name FROM information_schema.tables' ].
> >   pool close.
> >
> > When a connection is in use, it is not part of the pool.
> > When a connection is returned/released to the pool, it becomes available for reuse.
> >
> > The pool's size is the number of open connection ready to be reused.
> > The pool's capacity is the maximum number of connection that will be pooled.
> > Excess connections will be closed when they are returned/released to the pool.
> >
> > New connections are created as needed.
> > You can set a #configurator to further initialize new connections.
> > You can use #warmUp or #warmUp: to precreate a number of connections.
> >
> > When an error occurs, the connection should not be reused and be closed by the caller.
> >
> > ======
> >
> > There are some unit tests as well.
> >
> > Let me know if this works for you.
> >
> > Sven
> >
> > > On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
> > >
> > > Yes, thanks for good feedback.
> > >
> > > I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
> > >
> > > Petter
> > >
> > > On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > Hi Petter,
> > >
> > > > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> > > >
> > > > Hi Sven and thanks for answering!
> > > >
> > > > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
> > >
> > > It is useable, but not concurrently.
> > >
> > > > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
> > >
> > > Yes and no, see further.
> > >
> > > > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
> > >
> > > Yes, follow the pointer that I gave you, it is not hard.
> > >
> > > > Creating a pool sounds like the right solution to me now - any meaning about this?
> > >
> > > Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
> > >
> > > I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
> > >
> > > Running
> > >
> > >   P3ClientTest new runBenchmark1Bench.
> > >
> > > on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
> > >
> > > If I modify this slightly to use a new client/connection each time, like this
> > >
> > >   [
> > >     (P3Client url: 'psql://sven@localhost')
> > >       query: 'SELECT * FROM benchmark1';
> > >       close
> > >   ] benchFor: 5 seconds
> > >
> > > I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
> > >
> > > And note that this is not using prepared statements.
> > >
> > > So I would start by opening/closing a connection each time you need it.
> > >
> > > HTH,
> > >
> > > Sven
> > >
> > > > Petter
> > > >
> > > >
> > > > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > > Hi Petter,
> > > >
> > > > [ CC-ing the Pharo Users list ]
> > > >
> > > > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> > > >
> > > > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> > > >
> > > > Is that what you are doing ?
> > > >
> > > > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> > > >
> > > > That will then serialise requests and possibly block one onto the other.
> > > >
> > > > HTH,
> > > >
> > > > Sven
> > > >
> > > > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> > > >
> > > > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> > > >
> > > > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > > > >
> > > > > Hi Sven
> > > > >
> > > > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > > > >
> > > > > The problem seems to be connected to compiled sql statements and concurrency.
> > > > >
> > > > > We keep getting this error: Bindcomplete message expected
> > > > >
> > > > > Problem seems to be easy to reproduce:
> > > > >
> > > > > 1) Compile any sql statement
> > > > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > > > >
> > > > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > > > >
> > > > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > > > >
> > > > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > > > >
> > > > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > > > >
> > > > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > > > >
> > > > > Best regards
> > > > >
> > > > > Petter Egesund (I wrote the heysql-package based on P3)
> > > >
> > >
> >
>

Reply | Threaded
Open this post in threaded view
|

Re: P3 and concurrenty

ducasse
Amazing experience!
This is the power of Pharo at work.
Thanks a lot Sven.

S

On 10 Feb 2020, at 21:24, Petter Egesund <[hidden email]> wrote:

Perfect, this is exactly what I needed. Thank so much!!

Petter

On Mon, Feb 10, 2020 at 8:44 PM Sven Van Caekenberghe <[hidden email]> wrote:
Hi Petter,

> On 10 Feb 2020, at 16:46, Petter Egesund <[hidden email]> wrote:
>
> Yes, I see - that sounds sensible.
>
> Could it be an idea to subclass the pool-class and to prepare statements in this method, if there is a kind of init-method in the class? What do you think?

That would have indeed been one approach but for this situation I opted for a configurator block.

P3ConnectionPool>>#configurator: oneArgumentBlock
  "Set oneArgumentBlock to be my configurator.
   This is code that will be executed once on each newly created connection (P3Client).
   The default configuration asserts #isWorking on the argument"

The default is

  configurator := [ :p3Client | self assert: p3Client isWorking ]

which basically does an extra validation and ensures the opening of the connection.

Now, a little piece of functionality was still missing, I added it with the following commit:

  https://github.com/svenvc/P3/commit/e3f161d9f9ae34b3b020ad924b3c0f116f68c0b0

Your usage scenario should now be like this: you create a P3ConnectionPool with the proper URL, then you set your configurator to add your prepared statements (you could delegate this to a helper class if you have a lot of them).

  pool configurator: [ :p3Client |
    p3Client prepare: 'SELECT ...' named: 'select1'.
    p3Client prepare: 'INSET ...' named: 'insert1' ]

Later, when using a connection from the pool, you refer to your prepared statements by name.

  pool withConnection: [ :p3Client |
    (p3Client preparedStatementNamed: 'select1')
      execute: { args } ]

Sven

PS: have a look at P3PreparedStatementTest>>#testNamedPreparedStatement for an operation example (minus the connection pooling).

> On Mon, Feb 10, 2020 at 4:36 PM Sven Van Caekenberghe <[hidden email]> wrote:
>
>
> > On 10 Feb 2020, at 16:30, Petter Egesund <[hidden email]> wrote:
> >
> > Hi and thanks for an answer above my expectations to our problem.
> >
> > This is really awesome, I was just about to start coding a pool myself, but getting it directly from the author of the base library is off course way better :)
> >
> > I will try out and get back as soon as we have tested.
> >
> > One quick question - are the prepared statements which are connected to the pool thread safe?
>
> Prepared statements are scoped to a single connection in PSQL, AFAIK.
>
> What you will probably want to do is use a #configurator: block to set up your prepared statements once, like you do now. Then each thread/process grabs its own connection from the pool, (which will be already configured properly) and uses it, for itself, and finally returns it to the pool.
>
> Interacting with the pool should be fully thread safe. Once you are using a connection, it not longer is, but is also does not have to, since one connection equals one client thread.
>
> > Petter
> >
> > On Mon, Feb 10, 2020 at 12:13 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > Hi Petter,
> >
> > https://github.com/svenvc/P3/commit/a6b409d0d92cb92bf9b44452908bb9033523b863 adds a connection pool.
> >
> > Here is the class comment:
> >
> > ======
> >
> > I am P3ConnectionPool.
> >
> > I offer a pool of shared PSQL connections (P3Client instances) as a resource.
> >
> > After configuring me with at least a url, you use me by calling #withConnection:
> >
> >   pool := P3ConnectionPool url: '<a href="psql://sven@localhost" class="">psql://sven@localhost'.
> >   pool withConnection: [ :p3Client |
> >           p3Client query: 'SELECT table_name FROM information_schema.tables' ].
> >   pool close.
> >
> > When a connection is in use, it is not part of the pool.
> > When a connection is returned/released to the pool, it becomes available for reuse.
> >
> > The pool's size is the number of open connection ready to be reused.
> > The pool's capacity is the maximum number of connection that will be pooled.
> > Excess connections will be closed when they are returned/released to the pool.
> >
> > New connections are created as needed.
> > You can set a #configurator to further initialize new connections.
> > You can use #warmUp or #warmUp: to precreate a number of connections.
> >
> > When an error occurs, the connection should not be reused and be closed by the caller.
> >
> > ======
> >
> > There are some unit tests as well.
> >
> > Let me know if this works for you.
> >
> > Sven
> >
> > > On 9 Feb 2020, at 19:04, Petter Egesund <[hidden email]> wrote:
> > >
> > > Yes, thanks for good feedback.
> > >
> > > I will try the pooled way, I think - not primarily because of speed, but due to that our library is built around prepared connections.
> > >
> > > Petter
> > >
> > > On Sun, Feb 9, 2020 at 6:01 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > Hi Petter,
> > >
> > > > On 9 Feb 2020, at 17:27, Petter Egesund <[hidden email]> wrote:
> > > >
> > > > Hi Sven and thanks for answering!
> > > >
> > > > I use Teapot with one common sql-connecton, not one for each user session. At startup I create several sql statements and these does not seem to be usable from different Teapot request at the same time.
> > >
> > > It is useable, but not concurrently.
> > >
> > > > I could create one connection pr. session and then close the database connection when the user leaves, but then I also would need to create all the prepared sql-statements for each session, which does not sound right to me?
> > >
> > > Yes and no, see further.
> > >
> > > > It seems I have will have to look into the other solution, and see if I can use some mutex stuff to avoid several tasks acessessing the same resources at the same time.
> > >
> > > Yes, follow the pointer that I gave you, it is not hard.
> > >
> > > > Creating a pool sounds like the right solution to me now - any meaning about this?
> > >
> > > Yes, you could create a connection pool. But that is harder than it sounds: what is the minimum size, the maximum size, what do you do when you go over it, how do you make sure that a resource (connection) is clean when returning it to the pool (given authentication, possible errors), ...
> > >
> > > I don't know what you are doing, but I think you focus too much on performance issues. I would first try to get the code correct and worry about performance later on.
> > >
> > > Running
> > >
> > >   P3ClientTest new runBenchmark1Bench.
> > >
> > > on my machine gives me a BenchmarkResult(67 iterations in 5 seconds 57 milliseconds. 13.249 per second). This is a query that returns 10.000 records with 5 columns. It is reusing the same client/connection for all iterations.
> > >
> > > If I modify this slightly to use a new client/connection each time, like this
> > >
> > >   [
> > >     (P3Client url: '<a href="psql://sven@localhost" class="">psql://sven@localhost')
> > >       query: 'SELECT * FROM benchmark1';
> > >       close
> > >   ] benchFor: 5 seconds
> > >
> > > I get a BenchmarkResult(65 iterations in 5 seconds 34 milliseconds. 12.912 per second) which almost as fast. Of course, for smaller queries, the connect/disconnect overhead will be more significant.
> > >
> > > And note that this is not using prepared statements.
> > >
> > > So I would start by opening/closing a connection each time you need it.
> > >
> > > HTH,
> > >
> > > Sven
> > >
> > > > Petter
> > > >
> > > >
> > > > On Sun, Feb 9, 2020 at 4:52 PM Sven Van Caekenberghe <[hidden email]> wrote:
> > > > Hi Petter,
> > > >
> > > > [ CC-ing the Pharo Users list ]
> > > >
> > > > P3Client is not built/designed to be used by multiple processes concurrently. Each database connection is represented by an instance of P3Client and holds some state both at the client as well as at the server side.
> > > >
> > > > Typically, in a multi user server application, each connection should have its own P3Client / psql connection. For example, in Seaside, a custom WASession subclass gives each session/user its own p3 connection/client.
> > > >
> > > > Is that what you are doing ?
> > > >
> > > > If not, you could wrap your db accessing code so that mutual exclusion is provided. For example, you can have a look at AbstractCache #beThreadSafe and #critical:
> > > >
> > > > That will then serialise requests and possibly block one onto the other.
> > > >
> > > > HTH,
> > > >
> > > > Sven
> > > >
> > > > PS: another thing to take care of if closing your sql connections when the session is no longer needed.
> > > >
> > > > PS: Zinc HTTP does also provide a session mechanism (ZnServerSession[Manager]) but these work with cookies and typically won't help with a REST access pattern.
> > > >
> > > > > On 9 Feb 2020, at 14:21, Petter Egesund <[hidden email]> wrote:
> > > > >
> > > > > Hi Sven
> > > > >
> > > > > We are using Pharo as our backend in a project and we have run into a problem with P3.
> > > > >
> > > > > The problem seems to be connected to compiled sql statements and concurrency.
> > > > >
> > > > > We keep getting this error: Bindcomplete message expected
> > > > >
> > > > > Problem seems to be easy to reproduce:
> > > > >
> > > > > 1) Compile any sql statement
> > > > > 2) Use this statement in a query twice (!) in a teapot endpoint
> > > > >
> > > > > The run some concurrent queries, like "curl http://localhost:8080/endpoint & curl http://localhost:8080/endpoint.." (add several curls after here).
> > > > >
> > > > > One could also use ex. siege (https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html) for easy reproducing.
> > > > >
> > > > > If we chain the curls after each other, like "curl http://localhost:8080/endpoint && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html && https://manpages.ubuntu.com/manpages/trusty/man1/siege.1.html.." it seems to work fine, so doing the request sequentially seem to work fine.
> > > > >
> > > > > My conclusion is that this must be connected to how teapot handles concurrency in companion with the compiled statements?
> > > > >
> > > > > Any clues on this one? We are on Pharo 8.0 with latest version of P3, PG 9.x)
> > > > >
> > > > > Best regards
> > > > >
> > > > > Petter Egesund (I wrote the heysql-package based on P3)
> > > >
> > >
> >
>