Re: re : PHP and persistent connections

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Howdy,

It turns out that the number of max_persistent
is linked to the httpd processes in some
difficult-to-describe way.

It's not that hard to describe. The max_persistent/max_links values are
per Apache process.

It was difficult to describe because I was not recieving consistent
results in experiments due to a number of factors. It makes sense now.

Thus if you have:

pgsql.max_persistent = 2

and

MaxClients 300

You could potentially reach a state where you are maintaining 600
persistant connections to the database (if your PHP scripts actually
pg_pconnect() with two different connect strings).

I think that if you are using persistant connections you may as well set
MaxClients to the same number of database backends you are allowing (or
possibly a bit less if you need other connections to the database).
There's no real point in allowing more Maxclients as they'll just start
hitting connect errors anyway.

Well, see, the thing is, we do webhosting for a number of different
domains from the same server so the number of MaxClients needs to be
high. I think that 300 is obscene, as the server is not powerful enough
to handle 300 apache processes without dumping a large number of them
into swap space, not to mention the processing, but no matter what, we
would have several extra postgres backends just hanging around wasting
ram.
Only a few unique persistent connections would be in use at any given
time as only a few domains use the database.

This has made me realize just how completely braindead our server setup
is. ;-) It seems that we would to bring up a seperate database
server, very soon.

Obviously this isn't the most efficient use of backends because a fair
amount of the time the Apache processes won't be using them at all
(they'll be sitting there doing nothing, or serving images or other
static content).

Just what I was thinking. Connection pooling would avoid that, correct?

If your application is big enough you may benefit from serving static
content (images etc) from a different server, so the Apache processes
with persistant connection to backends are being used more heavily for
database work.

True, but in this case probably moving the database to a different server
would make more sense because most of the backends would be serving
content that is completely unrelated to the database.

Ie if you normally have 100 Apache processes running on your webserver
and you are using one persistant connection per process you will need
100 backends. However if at any one time those 60%
of those processes are serving images then you could have an Apache
server on one machine serving those images and only need 40 Apache
processes and therefore 40 backends on the Apache server that serves the
PHP script. You'll be tuning each machine for a more specific task,
rather than having one machine doing all sorts of different stuff.

I do not know what will happen with PHP when there are more than one
different (i.e. different username, database) persistent connections.
I suppose they would be affected by the max_persistent. (?).

If you want persistant connections to two different database/username
pairs then you need to have max_persistant=2, one for each different
connection string.

If you have one database that is used a lot and one that isn't, you may
wish to set max_persistant to 1 and max_clients to 2. Use pg_pconnect()
for the one accessed a lot and pg_connect() for the other. Set Apache
MaxClients to X and the max number of PG backends to X + Y, where Y
allows for the load required by the short lived pg_connect()s.

As you've probably noticed, balancing all this is a rather manual
process.
Perhaps Apache 2.0 will make way for some connection pooling.

I hope that wasn't too confusing.

Your explanation makes perfect sense. A Zen sort of understanding has
come to me through experimenting with different settings.

How would persistent connections fit into a dual-server setup where one
server is handling all of the webserving and the other simply handles the
database data-serving?
The number of backends on the database server would be independent
of the number of Apache processes on the webserver inasmuch as there
could be 75 Apache processes but only 25 are connected to backends on the
database server, correct?
There would not necessarily be any Apache stuff on the database server?

Oh, and if you are using pg_close() I don't think it works
in any currently released PHP4 versions. See:

This seems to be true. I ran into some fun link errors while
connecting and disconnecting more than once in a script.

Thanks again, and again.

gh

At 12:47 PM 11/24/00, GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Oh, and if you are using pg_close() I don't think it works
in any currently released PHP4 versions. See:

This seems to be true. I ran into some fun link errors while
connecting and disconnecting more than once in a script.

This sounds disturbing!

How then should I go about closing persistent connections? Can I close
them at all?

Would pg_close() work if I used it on non-persistent connections?

Thanks in advance,

Mikah

GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Howdy,

It turns out that the number of max_persistent
is linked to the httpd processes in some
difficult-to-describe way.

It's not that hard to describe. The max_persistent/max_links values are
per Apache process.

It was difficult to describe because I was not recieving consistent
results in experiments due to a number of factors. It makes sense now.

I've copied this email exchange over to my PHP folder.. I see what I can do
do to improve the online documentation. :-)

Well, see, the thing is, we do webhosting for a number of different
domains from the same server so the number of MaxClients needs to be
high. I think that 300 is obscene, as the server is not powerful enough
to handle 300 apache processes without dumping a large number of them
into swap space, not to mention the processing, but no matter what, we
would have several extra postgres backends just hanging around wasting
ram.
Only a few unique persistent connections would be in use at any given
time as only a few domains use the database.

Give them their own apache? You can set up two apache instances on one box,
set up one with lots of backends, set up the other to match the applicable
db usage...
You could make a postgres+apache box for these few clients...

This has made me realize just how completely braindead our server setup
is. ;-) It seems that we would to bring up a seperate database
server, very soon.

Depends on the load. I'm serving 429 domains off of PHP/PostgreSQL,
using non-persistant connections (even though almost every page has
a select or two), and it's working just fine. My biggest selects only
return a few hundred rows, my small inserts/updates are done in PHP,
the big ones (4,000+ rows) are just parsed into files that a Perl/cron job
takes care of them. It also depends, obviously, on how you write your
code for all of this, how good the hardware is, etc.
(PII/500, 512Mb of RAM, RH 6.2 for the above)

Obviously this isn't the most efficient use of backends because a fair
amount of the time the Apache processes won't be using them at all
(they'll be sitting there doing nothing, or serving images or other
static content).

Just what I was thinking. Connection pooling would avoid that, correct?

If your application is big enough you may benefit from serving static
content (images etc) from a different server, so the Apache processes
with persistant connection to backends are being used more heavily for
database work.

True, but in this case probably moving the database to a different server
would make more sense because most of the backends would be serving
content that is completely unrelated to the database.

Well, here's the problem:

1 apache/php/postgres thread = 1 possible persistant postgres connection

if you run up 200 threads on _any_ server instance, that means you need
200 waiting backends, if that server is also doing postgres content
with persistant connections anywhere in that server.

I think the idea being referred to works like this:
In a big, mega-hit app, you put your simple content on a simple server,
so the web pages reference GIF's/frames/whatever stored there, rather
than on a heavy-use box. This means that the clients go to *another*
web server for that non-dynamic content.

How would persistent connections fit into a dual-server setup where one
server is handling all of the webserving and the other simply handles the
database data-serving?

Er... well, if you db load was really heavy, this would make sense,
but your problem is about having all of the webserving in one place.

The number of backends on the database server would be independent
of the number of Apache processes on the webserver inasmuch as there
could be 75 Apache processes but only 25 are connected to backends on the
database server, correct?

All 75 Apache processes might eventually try to serve up the db pages.

So all 75 *could* eventually want persistant connections. You can't control
which process gets which page.

There would not necessarily be any Apache stuff on the database server?

Not if you don't want it, no. Keep in mind that using _non_ persistant
connections on this setup will be even slower, as well.

-Ron

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC machine,
which is currently in MacOS land. Your bopping may vary.

On Fri, Nov 24, 2000 at 02:48:18PM +0800, some SMTP stream spewed forth:

At 12:47 PM 11/24/00, GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Oh, and if you are using pg_close() I don't think it works
in any currently released PHP4 versions. See:

This seems to be true. I ran into some fun link errors while
connecting and disconnecting more than once in a script.

This sounds disturbing!

Maybe it should, I thought it was. Who knows.

How then should I go about closing persistent connections? Can I close
them at all?

You cannot, by design and purpose, close persistent connections.
You could kill the postgres backend, but that is not quite the same. ;-)

Would pg_close() work if I used it on non-persistent connections?

My experience has caused me to believe that no, it will not.
That is not final, as I do not have true proof.

Thanks in advance,

No prob, we are here to benefit each other.

It seems like PHP would open other new connections using pg_connect(),
but would not close them. Has anyone had experiences other than this?

gh

On Fri, Nov 24, 2000 at 12:52:34AM -0700, some SMTP stream spewed forth:

GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Howdy,

It turns out that the number of max_persistent
is linked to the httpd processes in some
difficult-to-describe way.

It's not that hard to describe. The max_persistent/max_links values are
per Apache process.

It was difficult to describe because I was not recieving consistent
results in experiments due to a number of factors. It makes sense now.

I've copied this email exchange over to my PHP folder.. I see what I can do
do to improve the online documentation. :-)

Great. Thanks.

Well, see, the thing is, we do webhosting for a number of different
domains from the same server so the number of MaxClients needs to be
high. I think that 300 is obscene, as the server is not powerful enough
to handle 300 apache processes without dumping a large number of them
into swap space, not to mention the processing, but no matter what, we
would have several extra postgres backends just hanging around wasting
ram.
Only a few unique persistent connections would be in use at any given
time as only a few domains use the database.

Give them their own apache? You can set up two apache instances on one box,
set up one with lots of backends, set up the other to match the applicable
db usage...
You could make a postgres+apache box for these few clients...

Er, I think I missed something.
You mean give them their own Apache instance using a seperate ip?

Is it /possible/ to have a group of httpd processes (Apache) share a
group of Postgres backends without having one backend to one httpd?
That would be connection pooling, correct? Which is not yet possible?

This has made me realize just how completely braindead our server setup
is. ;-) It seems that we would to bring up a seperate database
server, very soon.

Depends on the load. I'm serving 429 domains off of PHP/PostgreSQL,
using non-persistant connections (even though almost every page has
a select or two), and it's working just fine. My biggest selects only
return a few hundred rows, my small inserts/updates are done in PHP,
the big ones (4,000+ rows) are just parsed into files that a Perl/cron job
takes care of them. It also depends, obviously, on how you write your
code for all of this, how good the hardware is, etc.
(PII/500, 512Mb of RAM, RH 6.2 for the above)

That makes sense. The only reason I am so zealous about persistent
connections is that I have seen them be 3 times as fast as regular
connections.

Obviously this isn't the most efficient use of backends because a fair
amount of the time the Apache processes won't be using them at all

My main question now is, how can I avoid this?
I would have to go to non-persistent connections, correct?
I think I further understand things now.

So, persistent connections create a one-to-one ratio of
db-using Apache processes and Postgres backends, no matter what?
The only way to avoid such a one-to-one setup would be to
use non-persistent connections or do connection pooling?

So, even if the database were running on a seperate server,
each apache procees on the main server would require one backend process
on the db server?

I have a couple of other questions that I believe are not ansvered in
the docs anywhere.

Do the "persistent-connected" Postgres backends ever timeout or die?
Is it possible to set something like a timeout for persistent connctions?
(Er, would that be something that someone would want
to do? A Bad Thing?)

What happens when the httpd process that held a persistent connection
dies? Does "its" postgres process drop the connection and wait for
others? When the spare apache processes die, the postgres processes
remain.

Thanks.

gh

We're in quote hell.
Yay.

GH wrote:

On Fri, Nov 24, 2000 at 12:52:34AM -0700, some SMTP stream spewed forth:

GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Howdy,

Give them their own apache? You can set up two apache instances on one box,
set up one with lots of backends, set up the other to match the applicable
db usage...
You could make a postgres+apache box for these few clients...

Er, I think I missed something.
You mean give them their own Apache instance using a seperate ip?

Yes.

Apache one, httpd, serves 14 domains, conf files in /usr/local/apache/conf.
pgsql.max_persistent = 1
MaxClients 8

Apache two, httpd2, serves 327 domains, conf files in /usr/local/apache2/conf.
Max clients 150 (no postgres backends, no PHP)

Is it /possible/ to have a group of httpd processes (Apache) share a
group of Postgres backends without having one backend to one httpd?
That would be connection pooling, correct? Which is not yet possible?

Apache's process management, AFAIK, makes this fairly difficult. As in:
"I've never seen it, and I can't find docs on on, maybe v.2 will
have better support for children sharing common resources".

Depends on the load. I'm serving 429 domains off of PHP/PostgreSQL,
using non-persistant connections (even though almost every page has
a select or two), and it's working just fine. My biggest selects only
return a few hundred rows, my small inserts/updates are done in PHP,
the big ones (4,000+ rows) are just parsed into files that a Perl/cron job
takes care of them. It also depends, obviously, on how you write your
code for all of this, how good the hardware is, etc.
(PII/500, 512Mb of RAM, RH 6.2 for the above)

That makes sense. The only reason I am so zealous about persistent
connections is that I have seen them be 3 times as fast as regular
connections.

Hm.

I havn't. In PHP, one connection for the duration of a single
page (pg_connect()) takes as much time as a new persistant connection
(pg_pconnect()). Since you're often only creating one connection per page,
and running a single transaction on it, the main difference
would be in your connection setup... how did you test this? (I'm just
curious). Is it a usage test (real, live, use) or a bench test (push
to a limit that won't be reached in actual use.) I have one horribly
written app, that does maybe 50 _different_ selects on one page,
and it's still under two seconds per user....

Obviously this isn't the most efficient use of backends because a fair
amount of the time the Apache processes won't be using them at all

My main question now is, how can I avoid this?

Serve the postgres pages from a different server instance, on the same
machine, or a different one.

I would have to go to non-persistent connections, correct?

You could use persistant connections on a different server/instance,
or use non-persistant and loose ~10ms per page, less time than your average
10K GIF takes up on a 56K download.

You see, persistant PHP connections offer *no other value*, at all. None.
(it's a common error for new PHP folks to think that a web server
will somehow track their connections.) All it does is reduce setup time on a
page. No "session", no "tracking", nada. It reduces your connection
time for the page, but not significanly enough for users to know,
or care (IME). In web-page uses, the time is pretty much irrelevant,
because you only need one or two connections per page to get most
of your data out. Persistant connections are an interesting idea,
but they don't offer much. See:
http://www.php.net/manual/features.persistent-connections.php

So, persistent connections create a one-to-one ratio of
db-using Apache processes and Postgres backends, no matter what?

Almost. You can have more persistant connections for each apache
child, but each child may look for one. So it may be 5 apache
to 5 postgres, or 5 apache to 50 postgres, if needed (of course,
if you had that many conections, you may want to re-architect anyways)

The only way to avoid such a one-to-one setup would be to
use non-persistent connections or do connection pooling?

I'm still not following you on the "pooling". Apache doesn't, AFAICT,
offer this in each child. Each child is its own application, it's own
apache+php+postgres. Postgres doesn't care. PHP doesn't care. Apache
cares. If you give each child piece 5 postgres connections, and have
10 children, you need up to 50 backends.

So, even if the database were running on a seperate server,
each apache procees on the main server would require one backend process
on the db server?

Yup. If it was going to pull a postgres+PHP page, it would. You see,
apache doesn't work in a space where one apache process can crash the
whole thing. Each piece is isolated. This means that each piece needs
it's own resources. Compare this to other engines, where a single
crash on one serving instance takes down the _entire_ server, and
it makes sense (if the pool is down, it all goes down, a la IIS).

"It scales, but not that way". :-(

-Ronabop

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC machine,
which is currently in MacOS land. Your bopping may vary.

On Fri, Nov 24, 2000 at 04:52:27AM -0700, some SMTP stream spewed forth:

We're in quote hell.
Yay.

Ah, but now the hell thickens. ;-)

GH wrote:

On Fri, Nov 24, 2000 at 12:52:34AM -0700, some SMTP stream spewed forth:

GH wrote:

On Fri, Nov 24, 2000 at 03:17:59PM +1100, some SMTP stream spewed forth:

Howdy,

Give them their own apache? You can set up two apache instances on one box,
set up one with lots of backends, set up the other to match the applicable
db usage...
You could make a postgres+apache box for these few clients...

Er, I think I missed something.
You mean give them their own Apache instance using a seperate ip?

Yes.

Apache one, httpd, serves 14 domains, conf files in /usr/local/apache/conf.
pgsql.max_persistent = 1
MaxClients 8

Apache two, httpd2, serves 327 domains, conf files in /usr/local/apache2/conf.
Max clients 150 (no postgres backends, no PHP)

I see.

Is it /possible/ to have a group of httpd processes (Apache) share a
group of Postgres backends without having one backend to one httpd?
That would be connection pooling, correct? Which is not yet possible?

Apache's process management, AFAIK, makes this fairly difficult. As in:
"I've never seen it, and I can't find docs on on, maybe v.2 will
have better support for children sharing common resources".

Just checking. I had heard (and expected) that it did not -- for the same
reason.

Depends on the load. I'm serving 429 domains off of PHP/PostgreSQL,
using non-persistant connections (even though almost every page has
a select or two), and it's working just fine. My biggest selects only
return a few hundred rows, my small inserts/updates are done in PHP,
the big ones (4,000+ rows) are just parsed into files that a Perl/cron job
takes care of them. It also depends, obviously, on how you write your
code for all of this, how good the hardware is, etc.
(PII/500, 512Mb of RAM, RH 6.2 for the above)

That makes sense. The only reason I am so zealous about persistent
connections is that I have seen them be 3 times as fast as regular
connections.

Hm.

I havn't. In PHP, one connection for the duration of a single
page (pg_connect()) takes as much time as a new persistant connection
(pg_pconnect()). Since you're often only creating one connection per page,
and running a single transaction on it, the main difference
would be in your connection setup... how did you test this? (I'm just
curious). Is it a usage test (real, live, use) or a bench test (push
to a limit that won't be reached in actual use.) I have one horribly
written app, that does maybe 50 _different_ selects on one page,
and it's still under two seconds per user....

"Test" is a strong word. ;-) I have a timer set on a page.
The overall exec time is less that 1-tenth of a second using persistent
connections, so long as a connection exists. Using regular connections,
the exec time soars (;-)) to a whopping 3-tenths or so.
So, no big fat deal. The exec time is low enough that the effects of the
connections shine, but in general are insignificant.
If the script in discussion did anything worthwhile, I doubt that I would
notice anything even close to 3x.

Obviously this isn't the most efficient use of backends because a fair
amount of the time the Apache processes won't be using them at all

My main question now is, how can I avoid this?

Serve the postgres pages from a different server instance, on the same
machine, or a different one.

I would have to go to non-persistent connections, correct?

You could use persistant connections on a different server/instance,
or use non-persistant and loose ~10ms per page, less time than your average
10K GIF takes up on a 56K download.

You see, persistant PHP connections offer *no other value*, at all. None.
(it's a common error for new PHP folks to think that a web server
will somehow track their connections.) All it does is reduce setup time on a
page. No "session", no "tracking", nada. It reduces your connection
time for the page, but not significanly enough for users to know,
or care (IME). In web-page uses, the time is pretty much irrelevant,
because you only need one or two connections per page to get most
of your data out. Persistant connections are an interesting idea,
but they don't offer much. See:
http://www.php.net/manual/features.persistent-connections.php

I have read it (note: the phrasing seems to be a bit "messy"), but
for some reason I must have missed what it was saying. I "get it" now.

So, persistent connections create a one-to-one ratio of
db-using Apache processes and Postgres backends, no matter what?

Almost. You can have more persistant connections for each apache
child, but each child may look for one. So it may be 5 apache
to 5 postgres, or 5 apache to 50 postgres, if needed (of course,
if you had that many conections, you may want to re-architect anyways)

The only way to avoid such a one-to-one setup would be to
use non-persistent connections or do connection pooling?

I'm still not following you on the "pooling". Apache doesn't, AFAICT,

I almost knew that it did not. But I was trying to re-affirm my grasp of
just what "pooling" would do.

offer this in each child. Each child is its own application, it's own
apache+php+postgres. Postgres doesn't care. PHP doesn't care. Apache
cares. If you give each child piece 5 postgres connections, and have
10 children, you need up to 50 backends.

So, even if the database were running on a seperate server,
each apache procees on the main server would require one backend process
on the db server?

Yup. If it was going to pull a postgres+PHP page, it would. You see,
apache doesn't work in a space where one apache process can crash the
whole thing. Each piece is isolated. This means that each piece needs
it's own resources. Compare this to other engines, where a single
crash on one serving instance takes down the _entire_ server, and
it makes sense (if the pool is down, it all goes down, a la IIS).

"It scales, but not that way". :-(

Got it. Maybe this thread will finally pass away now. ;-)

Thanks again.

gh

GH <grasshacker@over-yonder.net> writes:

Do the "persistent-connected" Postgres backends ever timeout or die?

No. A backend will sit patiently for the client to send it another
query or close the connection.

(Barely on topic: in recent releases, the backend does set TCP
"keepalive" mode on the client socket. On a cross-machine connection,
this causes the kernel to ping every so often on an idle connection, to
make sure that the peer machine is still alive and still believes the
connection is open. However, this does not guard against a client
process that is holding connections open without any intention of using
them again soon --- it only protects against half-open connections left
over after a system crash at the client end. In any case, I believe the
total time delay before declaring the connection lost has to be an hour
or more in a spec-compliant TCP implementation.)

Is it possible to set something like a timeout for persistent connctions?
(Er, would that be something that someone would want
to do? A Bad Thing?)

This has been suggested before, but I don't think any of the core
developers consider it a good idea. Having the backend arbitrarily
disconnect on an active client would be a Bad Thing for sure. Hence,
any workable timeout would have to be quite large (order of an
hour, maybe? not milliseconds anyway). And that means that it's not
an effective solution for the problem. Under load, a webserver that
wastes backend connections will run out of available backends long
before a safe timeout would start to clean up after it.

To my mind, a client app that wants to use persistent connections
has got to implement some form of connection pooling, so that it
recycles idle connections back to a "pool" for allocation to task
threads that want to make a new query. And the threads have to release
connections back to the pool as soon as they're done with a transaction.
Actively releasing an idle connection is essential, rather than
depending on a timeout.

I haven't studied PHP at all, but from this conversation I gather that
it's only halfway there...

regards, tom lane

Note: CC'd to Hackers, as this has wandered into deeper feature issues.

Tom Lane wrote:

GH <grasshacker@over-yonder.net> writes:

Do the "persistent-connected" Postgres backends ever timeout or die?

No. A backend will sit patiently for the client to send it another
query or close the connection.

This does have an unfortunate denial-of-service implication, where
an attack can effectively suck up all available backends, and there's
no throttle, no timeout, no way of automatically dropping these....

However, the more likely possibility is similar to the problem that
we see in PHP's persistant connections.... a normally benign connection
is inactive, and yet it isn't dropped. If you have two of these created
every day, and you only have 16 backends, after 8 days you have a lockout.

On a busy web site or another busy application, you can, of course,
exhaust 64 backends in a matter of minutes.

Is it possible to set something like a timeout for persistent connctions?
(Er, would that be something that someone would want
to do? A Bad Thing?)

This has been suggested before, but I don't think any of the core
developers consider it a good idea. Having the backend arbitrarily
disconnect on an active client would be a Bad Thing for sure.

Right.... but I don't think anybody has suggested disconnecting an *active*
client, just inactive ones.

Hence,
any workable timeout would have to be quite large (order of an
hour, maybe? not milliseconds anyway).

The mySQL disconnect starts at around 24 hours. It prevents a slow
accumulation of unused backends, but does nothing for a rapid
accumulation. It can be cranked down to a few minutes AFAIK.

And that means that it's not
an effective solution for the problem. Under load, a webserver that
wastes backend connections will run out of available backends long
before a safe timeout would start to clean up after it.

Depends on how it's set up... you see, this isn't uncharted territory,
other web/db solutions have already fought with this issue. Much
like the number of backends set up for pgsql must be static, a timeout
may wind up being the same way. The critical thing to realize is
that you are timing out _inactive_ connections, not connections
in general. So provided that a connection provided information
about when it was last used, or usage set a counter somewhere, it
could easily be checked.

To my mind, a client app that wants to use persistent connections
has got to implement some form of connection pooling, so that it
recycles idle connections back to a "pool" for allocation to task
threads that want to make a new query. And the threads have to release
connections back to the pool as soon as they're done with a transaction.
Actively releasing an idle connection is essential, rather than
depending on a timeout.

I haven't studied PHP at all, but from this conversation I gather that
it's only halfway there...

Well...... This is exactly how apache and PHP serve pages. The
problem is that apache children aren't threads, they are separate copies
of the application itself. So a single apache thread will re-use the
same connection, over and over again, and give that conection over to
other connections on that apache thread.. so in your above model, it's
not really one client application in the first place.

It's a dynamic number of client applications, between one and hundreds
or so.

So to turn the feature request the other way 'round:
"I have all sorts of client apps, connecting in different ways, to
my server. Some of the clients are leaving their connections open,
but unused. How can I prevent running out of backends, and boot
the inactive users off?"

-Ronabop

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC machine,
which is currently in MacOS land. Your bopping may vary.

At 05:26 PM 11/25/00 -0700, Ron Chmara wrote:

Note: CC'd to Hackers, as this has wandered into deeper feature issues.

Tom Lane wrote:

GH <grasshacker@over-yonder.net> writes:

Do the "persistent-connected" Postgres backends ever timeout or die?

No. A backend will sit patiently for the client to send it another
query or close the connection.

This does have an unfortunate denial-of-service implication, where
an attack can effectively suck up all available backends, and there's
no throttle, no timeout, no way of automatically dropping these....

However, the more likely possibility is similar to the problem that
we see in PHP's persistant connections.... a normally benign connection
is inactive, and yet it isn't dropped. If you have two of these created
every day, and you only have 16 backends, after 8 days you have a lockout.

On a busy web site or another busy application, you can, of course,
exhaust 64 backends in a matter of minutes.

Ugh...the more I read stuff like this the more I appreciate AOlserver's
built-in database API which protects the application from any such
problems altogether. The particular problem being described simply
can't occur in this environment.

- Don Baccus, Portland OR <dhogaza@pacifier.com>
Nature photos, on-line guides, Pacific Northwest
Rare Bird Alert Service and other goodies at
http://donb.photo.net.

"I have all sorts of client apps, connecting in different ways, to
my server. Some of the clients are leaving their connections open,
but unused. How can I prevent running out of backends, and boot
the inactive users off?"

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??

that middleman could be configured to have 16 persistant connections,every
clients would deal with the middleman instead of going direct to the
database,this would be an advantage where multiple PostgreSQL server are
used...

240 apache process are running on a box and there's 60 PostgreSQL instance
running on the machine or another machine:

240 apache process --> middleman --> 60 PostgreSQL process

now if there's multiple Database server:

240 apache process --> middleman --> 12 PostgreSQL for each server (5
servers in this case)

in this case,the middleman could be a shared library which the clients
link to..

what do you think about that ??

Alain Toussaint

At 10:00 PM 11/25/00 -0800, Mitch Vincent wrote:

I've tried quite a bit to use persistent connections with PHP (for over
a year) and always the scripts that I try to use them with behave crazy...
The last time I tried there were problems all over the place with PHP,
variables getting overwritten, certain functions just totally breaking
(date() to name one) and so on.. I know I'm not being specific but my point
is that I think there are some other outstanding PHP issues that play into
this problem as the behavior that I've seen isn't directly related to
PostgreSQL but only happens when I use persistent connections..

I've heard rumors that PHP isn't thoroughly threadsafe, could this be a
source of your problems?

- Don Baccus, Portland OR <dhogaza@pacifier.com>
Nature photos, on-line guides, Pacific Northwest
Rare Bird Alert Service and other goodies at
http://donb.photo.net.

At 12:07 AM 11/26/00 -0500, Alain Toussaint wrote:

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??

that middleman could be configured to have 16 persistant connections,every
clients would deal with the middleman instead of going direct to the
database,this would be an advantage where multiple PostgreSQL server are
used...

Well, this is sort of what AOLserver does for you without any need for
middlemen.

Again, reading stuff like this makes me think "ugh!"

This stuff is really pretty easy, it's amazing to me that the Apache/db
world talks about such kludges when they're clearly not necessary.

My first experience running a website (donb.photo.net) was with Apache
on Linux on an old P100 system in 1996 when few folks had personal photo
sites with >1000 photos on them getting thousands of hits a day. I have
fond memories of those days, and Apache served me (or more properly webserved
my website) well. This site is largely responsible for my reputation that
lets me freelance nature photography to the national media market pretty
much at will. Thus my fondness.

But ... for database stuff the release of AOLserver as first Free Beer,
and now Free Speech software has caused me to abandon Apache and suggestions
like the above just make me cringe.

It shouldn't be that hard, folks.

- Don Baccus, Portland OR <dhogaza@pacifier.com>
Nature photos, on-line guides, Pacific Northwest
Rare Bird Alert Service and other goodies at
http://donb.photo.net.

I've tried quite a bit to use persistent connections with PHP (for over
a year) and always the scripts that I try to use them with behave crazy...
The last time I tried there were problems all over the place with PHP,
variables getting overwritten, certain functions just totally breaking
(date() to name one) and so on.. I know I'm not being specific but my point
is that I think there are some other outstanding PHP issues that play into
this problem as the behavior that I've seen isn't directly related to
PostgreSQL but only happens when I use persistent connections.. I've been
trying to corner the problem for quite some time, it's an elusive one for
sure.. I spoke with the PHP developers 9 or so months ago about the problems
and they didn't seem to pay any attention to it, the thread on the mailing
list was short with the bug report collecting dust at the bottom of the
to-do list I'm sure (as that was back before PHP 4 was even released and
obviously the problem remains)..

Just my $0.02 worth.

-Mitch

----- Original Message -----
From: "Ron Chmara" <ron@Opus1.COM>
To: "Tom Lane" <tgl@sss.pgh.pa.us>; "PostgreSQL Hackers List"
<pgsql-hackers@postgresql.org>
Cc: "GH" <grasshacker@over-yonder.net>; <pgsql-novice@postgresql.org>
Sent: Saturday, November 25, 2000 4:26 PM
Subject: [HACKERS] Re: [NOVICE] Re: re : PHP and persistent connections

Note: CC'd to Hackers, as this has wandered into deeper feature issues.

Tom Lane wrote:

GH <grasshacker@over-yonder.net> writes:

Do the "persistent-connected" Postgres backends ever timeout or die?

No. A backend will sit patiently for the client to send it another
query or close the connection.

This does have an unfortunate denial-of-service implication, where
an attack can effectively suck up all available backends, and there's
no throttle, no timeout, no way of automatically dropping these....

However, the more likely possibility is similar to the problem that
we see in PHP's persistant connections.... a normally benign connection
is inactive, and yet it isn't dropped. If you have two of these created
every day, and you only have 16 backends, after 8 days you have a lockout.

On a busy web site or another busy application, you can, of course,
exhaust 64 backends in a matter of minutes.

Is it possible to set something like a timeout for persistent

connctions?

(Er, would that be something that someone would want
to do? A Bad Thing?)

This has been suggested before, but I don't think any of the core
developers consider it a good idea. Having the backend arbitrarily
disconnect on an active client would be a Bad Thing for sure.

Right.... but I don't think anybody has suggested disconnecting an

*active*

client, just inactive ones.

Hence,
any workable timeout would have to be quite large (order of an
hour, maybe? not milliseconds anyway).

The mySQL disconnect starts at around 24 hours. It prevents a slow
accumulation of unused backends, but does nothing for a rapid
accumulation. It can be cranked down to a few minutes AFAIK.

And that means that it's not
an effective solution for the problem. Under load, a webserver that
wastes backend connections will run out of available backends long
before a safe timeout would start to clean up after it.

Depends on how it's set up... you see, this isn't uncharted territory,
other web/db solutions have already fought with this issue. Much
like the number of backends set up for pgsql must be static, a timeout
may wind up being the same way. The critical thing to realize is
that you are timing out _inactive_ connections, not connections
in general. So provided that a connection provided information
about when it was last used, or usage set a counter somewhere, it
could easily be checked.

To my mind, a client app that wants to use persistent connections
has got to implement some form of connection pooling, so that it
recycles idle connections back to a "pool" for allocation to task
threads that want to make a new query. And the threads have to release
connections back to the pool as soon as they're done with a transaction.
Actively releasing an idle connection is essential, rather than
depending on a timeout.

I haven't studied PHP at all, but from this conversation I gather that
it's only halfway there...

Well...... This is exactly how apache and PHP serve pages. The
problem is that apache children aren't threads, they are separate copies
of the application itself. So a single apache thread will re-use the
same connection, over and over again, and give that conection over to
other connections on that apache thread.. so in your above model, it's
not really one client application in the first place.

It's a dynamic number of client applications, between one and hundreds
or so.

So to turn the feature request the other way 'round:
"I have all sorts of client apps, connecting in different ways, to
my server. Some of the clients are leaving their connections open,
but unused. How can I prevent running out of backends, and boot
the inactive users off?"

-Ronabop

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC

machine,

Well, this is sort of what AOLserver does for you without any need for
middlemen.

i agree that AolServer is good karma,i've been reading various docs on
Aolserver since Philip Greenspun talked about it on linuxworld and i'm glad
that there's some java support being coded for it (im my opinion,it's the only
advantage that Apache had over AolServer for me).

Again, reading stuff like this makes me think "ugh!"

This stuff is really pretty easy, it's amazing to me that the Apache/db
world talks about such kludges when they're clearly not necessary.

well...i was using Apache as an example due to it DB model but the stuff i
was talking would work quite well in the case of multiple DB server
hosting differents table and you want to maintain location
independance,here's an example:

you have 7 Database server,5 are online and the other 2 are for
maintenance and/or development purpose,for simplicity,we'll name the
server database1.example.net to
database7.example.net,database4.example.net is currently doing a dump and
database6.example.net is loading the dump from database4,then,you
reconfigure the middleman so it redirect all request from database4 to
database6:

vim /etc/middleman.conf

and then a sighup to the middleman so it reread its config file:

killall -HUP middleman

this would update the middleman's shared lib with the new configuration
info (and BTW,i just extended my idea from a single shared lib to a
daemon/shared lib combo).

now i'm off to get the dog out for a walk and then,take a nap,see ya !!

Alain Toussaint

I'm sure that this, if true, could certainly be the source of the problems
I've seen... I can't comment on if PHP is completely threadsafe, I know that
some of the modules (for lack of a better word) aren't, possible the ClibPDF
library I'm using. I'll check into it.

Thanks!

-Mitch

----- Original Message -----
From: "Don Baccus" <dhogaza@pacifier.com>
To: "Mitch Vincent" <mitch@venux.net>; "PostgreSQL Hackers List"
<pgsql-hackers@postgresql.org>
Cc: <pgsql-novice@postgresql.org>
Sent: Saturday, November 25, 2000 9:18 PM
Subject: Re: [HACKERS] Re: [NOVICE] Re: re : PHP and persistent connections

At 10:00 PM 11/25/00 -0800, Mitch Vincent wrote:

I've tried quite a bit to use persistent connections with PHP (for

over

a year) and always the scripts that I try to use them with behave

crazy...

The last time I tried there were problems all over the place with PHP,
variables getting overwritten, certain functions just totally breaking
(date() to name one) and so on.. I know I'm not being specific but my

point

is that I think there are some other outstanding PHP issues that play

into

On Sun, 26 Nov 2000, Alain Toussaint wrote:

"I have all sorts of client apps, connecting in different ways, to
my server. Some of the clients are leaving their connections open,
but unused. How can I prevent running out of backends, and boot
the inactive users off?"

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??

I don't see it solving anything. You just move the connection
management problem from the database to the middleman (in the industry
such a thing would be called a query multiplexor). Multiplexors have
often been used in the past to solve this problem, because the database
could not be extended or protected.

Besides, if you are an n-tier developer, this isn't a problem as your
middle tier not does connection management, but some logic as well. At
the end of the day, PHP/Apache is just not suitable for complex
applications.

Tom

Don Baccus wrote:

At 12:07 AM 11/26/00 -0500, Alain Toussaint wrote:

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??
that middleman could be configured to have 16 persistant connections,every
clients would deal with the middleman instead of going direct to the
database,this would be an advantage where multiple PostgreSQL server are
used...

Well, this is sort of what AOLserver does for you without any need for
middlemen.

What if you have a server farm of 8 AOL servers, and 12 perl clients, and
3 MS Access connections, leaving things open? Is AOLserver parsing the
Perl DBD/DBI, connects, too? So you're using AOLserver as (cough) a
middleman? <g>

Again, reading stuff like this makes me think "ugh!"
This stuff is really pretty easy, it's amazing to me that the Apache/db
world talks about such kludges when they're clearly not necessary.

How does AOL server time out access clients, ODBC connections, Perl
clients? I thought it was mainly web-server stuff.

Apache/PHP isn't the only problem. The problem isn't solved by
telling others to fix their software, either... is this something
that can be done _within_ postmaster?

-Bop

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC machine,
which is currently in MacOS land. Your bopping may vary.

Tom Samplonius wrote:

On Sun, 26 Nov 2000, Alain Toussaint wrote:

"I have all sorts of client apps, connecting in different ways, to
my server. Some of the clients are leaving their connections open,
but unused. How can I prevent running out of backends, and boot
the inactive users off?"

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??

I don't see it solving anything. You just move the connection
management problem from the database to the middleman (in the industry
such a thing would be called a query multiplexor). Multiplexors have
often been used in the past to solve this problem, because the database
could not be extended or protected.

And I'm requesting protection. Because the database isn't capable of dynamically
detroying temporary backends. (Which would be another solution to this
problem)

Besides, if you are an n-tier developer, this isn't a problem as your
middle tier not does connection management, but some logic as well. At
the end of the day, PHP/Apache is just not suitable for complex
applications.

Is it dump on PHP day?

Okay, pretend the problem is left-open Perl connections. Slam that for
a while. Move over to left open Access connections. Bag on that for
a few posts. Errant C code for a few days. Still have a problem. :-)

How does a db admin close connections that are idle, and unwanted, without
shutting the postmaster down?

-Bop

--
Brought to you from iBop the iMac, a MacOS, Win95, Win98, LinuxPPC machine,
which is currently in MacOS land. Your bopping may vary.

At 12:38 AM 11/27/00 -0700, Ron Chmara wrote:

Don Baccus wrote:

At 12:07 AM 11/26/00 -0500, Alain Toussaint wrote:

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??
that middleman could be configured to have 16 persistant connections,every
clients would deal with the middleman instead of going direct to the
database,this would be an advantage where multiple PostgreSQL server are
used...

Well, this is sort of what AOLserver does for you without any need for
middlemen.

What if you have a server farm of 8 AOL servers, and 12 perl clients, and
3 MS Access connections, leaving things open? Is AOLserver parsing the
Perl DBD/DBI, connects, too? So you're using AOLserver as (cough) a
middleman? <g>

Well, no - we'd use the built-in Tcl, Python or nsjava (still in infancy)
modules which interface natively to AOLserver's built-in database API.

You don't NEED the various connection implementations buried in various
languages because they're provided directly in the server. That's the
point. That's the main reason people use it.

If you're going to run CGI/Perl scripts using its database connectivity
stuff, don't use AOLserver. They'll run since AOLserver supports CGI,
but they'll run no better than under Apache and probably worse, since
no one doing serious AOLserver work uses CGI and therefore the code which
implements it has languished - there's no motivation to improve something
that no one uses.

If you're willing to use a language module which exposes the AOLserver
API to your application, then AOLserver's a great choice.

Again, reading stuff like this makes me think "ugh!"
This stuff is really pretty easy, it's amazing to me that the Apache/db
world talks about such kludges when they're clearly not necessary.

How does AOL server time out access clients, ODBC connections, Perl
clients? I thought it was mainly web-server stuff.

Well, for starters one normally wouldn't use ODBC since AOLserver
includes drivers for PostgreSQL, Oracle and Sybase. There's one for
Solid, too, but no one seems to use Solid since they raised their
prices drastically a couple of years ago (if you're going to spend
lots of money on a database, Oracle and Sybase are more than willing
to help you). Nor does nsjava use JDBC, it encapsulates the AOLserver
API into a database API class(es?).

AOLserver manages the database pools in about the same way it manages
threads, i.e. if a thread can't get the handles it needs (usually only
one, sometimes two, more than that usually indicates poorly written
code) it blocks until another thread releases a handle. When a thread
ends (returns a page) any allocated handles are released. Transactions
that haven't been properly committed are rolled back as well (lesser of
two evils - the event's logged since it indicates a bug).

For each pool you provide the name of the driver (which of course serves
to select which RDMBS that pool will use - you can use as many different
RDBMSs as you have, and have drivers for), a datasource, the maximum
number of connections to open for that pool, minimum and maximum lifetimes
for connections, etc.

- Don Baccus, Portland OR <dhogaza@pacifier.com>
Nature photos, on-line guides, Pacific Northwest
Rare Bird Alert Service and other goodies at
http://donb.photo.net.

Uh, Don?
Not all the world's a web page, you know. Thatkind of thinking is _so_
mid 90's ;-) Dedicated apps that talk directly the user seem to be making
a comeback, due to a number of factors. They can have much cleaner user
interfaces, for example.

Which brings us back around to the point of why this is on Hackers:
PostgreSQL currently has no clean method for dropping idle connections.
Yes, some apps handle this themselves, but not all. A number of people
seem to feel there is a need for this feature. How hard would it be to
implement?

Probably not too hard: we've already got an 'idle' state, suring which we
block on the input. Add a timeout to hat, and we're pretty much there.

<goes and looks at code for a bit>

Hmm, we're down in the bowels of libpq, doing a recv() on the socket
to the frontend, about 4 layers down from backend's blocking call to
ReadCommand(). I seem to recall someone working on creating an async
version of the libpq API, but Tom not being happy with the approach.
So, it's not a simple change.

Ross

On Mon, Nov 27, 2000 at 07:18:48AM -0800, Don Baccus wrote:

At 12:38 AM 11/27/00 -0700, Ron Chmara wrote:

Don Baccus wrote:

At 12:07 AM 11/26/00 -0500, Alain Toussaint wrote:

how about having a middle man between apache (or aolserver or any other
clients...) and PosgreSQL ??
that middleman could be configured to have 16 persistant connections,every
clients would deal with the middleman instead of going direct to the
database,this would be an advantage where multiple PostgreSQL server are
used...

Well, this is sort of what AOLserver does for you without any need for
middlemen.

What if you have a server farm of 8 AOL servers, and 12 perl clients, and
3 MS Access connections, leaving things open? Is AOLserver parsing the
Perl DBD/DBI, connects, too? So you're using AOLserver as (cough) a
middleman? <g>

Note that only the AOL servers here are web client/servers, the rest are
dedicated apps.

<snip Don missing the point>

--
Open source code is like a natural resource, it's the result of providing
food and sunshine to programmers, and then staying out of their way.
[...] [It] is not going away because it has utility for both the developers
and users independent of economic motivations. Jim Flynn, Sunnyvale, Calif.

You could set MaxRequestsPerChild in apache's httpd.conf. This controls
how many requests each apache process is allowed to serve. After it
serves this many the process dies which should close the postgres process
as well (if it isn't, you have other problems).

I know that for a long time Apache recommened setting this fairly low on
Solaris due to a memory leak in solaris...ideally you'd want to set this
really high, but setting it low will make the processes die...

-philip

On Fri, 24 Nov 2000, jmcazurin wrote:

Is it possible to set something like a timeout for persistent connctions?
(Er, would that be something that someone would want
to do? A Bad Thing?)

see my other email about apache's MaxRequestsPerChild...

What happens when the httpd process that held a persistent connection
dies? Does "its" postgres process drop the connection and wait for
others? When the spare apache processes die, the postgres processes
remain.

On my server (freebsd 4.x, php 4.0.2, postgresl 7.0.3) when I kill the
httpd processes the postgres processes die as well...

"Ross J. Reedstrom" <reedstrm@rice.edu> writes:

Which brings us back around to the point of why this is on Hackers:
PostgreSQL currently has no clean method for dropping idle connections.
Yes, some apps handle this themselves, but not all. A number of people
seem to feel there is a need for this feature.

I'm still not following exactly what people think would happen if we did
have such a "feature". OK, the backend times out after some interval
of seeing no activity, and disconnects. How is the client going to
react to that, exactly, and why would it not conclude that something's
gone fatally wrong with the database?

Seems to me that you still end up having to fix the client, and that
in the last analysis this is a client issue, not something for the
backend to hack around.

regards, tom lane

On Mon, Nov 27, 2000 at 12:09:00PM -0500, Tom Lane wrote:

I'm still not following exactly what people think would happen if we did
have such a "feature". OK, the backend times out after some interval
of seeing no activity, and disconnects. How is the client going to
react to that, exactly, and why would it not conclude that something's
gone fatally wrong with the database?

Because a lot of commercial (and other) databases have this "feature",
a lot of well behaved apps (and middleware packages) already know how
to deal with it: i.e. try to reconnect, and continue. If that fails,
throw an error.

Seems to me that you still end up having to fix the client, and that
in the last analysis this is a client issue, not something for the
backend to hack around.

It's already fixed, see above. In addition, your assuming the same
administrative entity has control over the clients and the backend.
This is not always the case. For example, in a web hosting environment.
Then, the DBA has the responsibiltiy to ensure minimal interference
between different customers.

As it stands, the client that causes the problem sees no problem to
fix: other clients get 'that damn PostgreSQL backend quits accepting
connections', and yell at the DBA. So, the DBA wants a way to propagate
the 'problem' to the clients that cause it, by timing out the idle
connections. Then, those clients _will_ fix their code, if it doesn't
already do it for them, as per above.

Basically, PostgreSQL is being too polite: it's in the clients interest to
keep the connection open, since it minimizes response time, regardless
of how this might affect other backends. It's cooperative vs. hard
multitasking, all over again.

Clients and servers optimize for different parameters: the client wants
minimum response time for it's requests. The backend wants minimum
_average_ response time, over all requests.

Ross
--
Open source code is like a natural resource, it's the result of providing
food and sunshine to programmers, and then staying out of their way.
[...] [It] is not going away because it has utility for both the developers
and users independent of economic motivations. Jim Flynn, Sunnyvale, Calif.

At 10:46 AM 11/27/00 -0600, Ross J. Reedstrom wrote:

Uh, Don?
Not all the world's a web page, you know. Thatkind of thinking is _so_
mid 90's ;-) Dedicated apps that talk directly the user seem to be making
a comeback, due to a number of factors. They can have much cleaner user
interfaces, for example.

Of course. But the question's been raised in the context of a web server,
and I've answered in context.

I've been trying to move the discussion offline to avoid clogging
the hackers list with this stuff but some of the messages have escaped
my machine with my forgetting to remove pg_hackers from the distribution
list. I'll try to be more diligent if the discussion continues.

- Don Baccus, Portland OR <dhogaza@pacifier.com>
Nature photos, on-line guides, Pacific Northwest
Rare Bird Alert Service and other goodies at
http://donb.photo.net.