Hello,

Is there any option to reload configuration in pound without sending
to program SIGTERM ?? We are currenlty using pound as load balancer
between 5 servers and over 2k sites. We are using header filtering to
decide to which backend request should be send, and we really need to
reload configuration ever 10minutes. killall -9 pound && pound does
the work, but by doing this we have nearly 1,2sec downtime which is
unacceptable. Have you got and ideas how to solve this problem, or how
to modify source of pound to allow reloading without restarting ??
Virtual host on the backend is impossible in our case.



-- 
Pozdrowienia,
 ForAll.pl - Firma

Pound really doesn't do this.

Take a look at haproxy.


On 10/26/06 11:03 AM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

> Hello,
> 
> Is there any option to reload configuration in pound without sending
> to program SIGTERM ?? We are currenlty using pound as load balancer
> between 5 servers and over 2k sites. We are using header filtering to
> decide to which backend request should be send, and we really need to
> reload configuration ever 10minutes. killall -9 pound && pound does
> the work, but by doing this we have nearly 1,2sec downtime which is
> unacceptable. Have you got and ideas how to solve this problem, or how
> to modify source of pound to allow reloading without restarting ??
> Virtual host on the backend is impossible in our case.
> 
> 

> Pound really doesn't do this.
>   
Will the feature "dynamic service configuration" be added?
How much can the new tool "poundctl" help in this use case?


> Take a look at haproxy.
>   
Is there a little competion running between free load balancers?

Regards,
Markus

>> Pound really doesn't do this.
>>   
> Will the feature "dynamic service configuration" be added?
> How much can the new tool "poundctl" help in this use case?


>> Take a look at haproxy.
>>   
> Is there a little competion running between free load balancers?

> Regards,
> Markus


poundctl didn't help - it can only manage backends that was already
readed from the config file.

But I've found another solution - we can run two instances of pound
(one, primary on port 80 which will redirect to the second one on port
81)
It could look like this:

POUND Primary
(port 80)
      |
      |
      |
      |
POUND Secondary
(port 81, restart every 10minutes)
      |   |
      |   |
      |   |
      |   |
      BACKENDS

And now I have to change sources of primary Pound to wait 2 seconds,
when connection to backend fails(restarting secondary pound). And by
this way client won't see "connection refused", but simple will have
to wait 2 seconds to load page.
So My question is... how to change pound to wait for example 2 seconds
on backend error and then repeat connection without sending any error
to client ??

On 10/26/06, ForAll.pl - Firma <firma(at)forall.pl> wrote:
> And now I have to change sources of primary Pound to wait 2 seconds,
> when connection to backend fails(restarting secondary pound). And by
> this way client won't see "connection refused", but simple will have
> to wait 2 seconds to load page.
> So My question is... how to change pound to wait for example 2 seconds
> on backend error and then repeat connection without sending any error
> to client ??

Using your diagram, your frontend pound on port 80 can forward
requests to 2 pounds, say on port 81 and 82.

When you are ready to restart one to reload the configuration, disable
backend pound 81, wait for all requests to finish then restart it, and
enable it, disable pound 82, wait for all requests to finish and
enable it.

Repeat when you are ready to reload again. You should be able to
automate this with a shell script...

-Dave

> On 10/26/06, ForAll.pl - Firma <firma(at)forall.pl> wrote:
>> And now I have to change sources of primary Pound to wait 2 seconds,
>> when connection to backend fails(restarting secondary pound). And by
>> this way client won't see "connection refused", but simple will have
>> to wait 2 seconds to load page.
>> So My question is... how to change pound to wait for example 2 seconds
>> on backend error and then repeat connection without sending any error
>> to client ??

> Using your diagram, your frontend pound on port 80 can forward
> requests to 2 pounds, say on port 81 and 82.

> When you are ready to restart one to reload the configuration, disable
> backend pound 81, wait for all requests to finish then restart it, and
> enable it, disable pound 82, wait for all requests to finish and
> enable it.

> Repeat when you are ready to reload again. You should be able to
> automate this with a shell script...

> -Dave

Nice idea, but running 3 pounds on one machine, which process over
20mln requests per day is not very efficient.

Also one more thing - how can I check in my shell script that all requests to
pound are
finished ??

On 10/26/06, ForAll.pl - Firma <firma(at)forall.pl> wrote:
> Nice idea, but running 3 pounds on one machine, which process over
> 20mln requests per day is not very efficient.

Sorry, best I can come up with right now! I also wish that you could
gracefully reload pound configs as well. :-)

> Also one more thing - how can I check in my shell script that all requests to
pound are
> finished ??

Check for established connections to the port that pound is bound to.
For example, on Linux:

netstat -an | grep "$IP_ADDR:81" | grep ESTABLISHED

will help.

-Dave

Or just wait 30 seconds if your longest request can't be more than 10
seconds.

On 10/26/06 3:02 PM, "David Rees" <drees76(at)gmail.com> wrote:

>> Also one more thing - how can I check in my shell script that all requests
to
>> pound are
>> finished ??
> 
> Check for established connections to the port that pound is bound to.

> Or just wait 30 seconds if your longest request can't be more than 10
> seconds.

> On 10/26/06 3:02 PM, "David Rees" <drees76(at)gmail.com> wrote:

>>> Also one more thing - how can I check in my shell script that all requests
to
>>> pound are
>>> finished ??
>> 
>> Check for established connections to the port that pound is bound to.


If somebody is downloading 100mb file it could take a long time :)
With about 200conn/sec there are nearly 50connections to big files, so
it is almost impossible to restart pound without breaking connections.

The best option will be to reload config without restarting pound, but
I have no idea on how to do this.

On 10/26/06, ForAll.pl - Firma <firma(at)forall.pl> wrote:
> If somebody is downloading 100mb file it could take a long time :)
> With about 200conn/sec there are nearly 50connections to big files, so
> it is almost impossible to restart pound without breaking connections.

Yes, long established connections will make it nearly impossible to
restart pound in a timely manner, without a large pool of pounds to
rotate through. For example, say you expect the longest connection to
stay open 60 minutes, you would need to have 6 pounds available to
rotate through if you are reloading the configuration every 10
minutes.

A bit of a hack, but luckily pound doesn't use much memory when running.

-Dave

Just checking on assumptions here, but why is it that you need to reload the
pound configuration every 10 minutes?  That seems kind of excessive.  I
still think that pound should behave better in this regard, but I am still
surprised that you need to reload that often.

Is there something that you could generalize in your configuration, perhaps,
that would make much of this reloading unnecessary?


On 10/26/06 3:52 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>> Or just wait 30 seconds if your longest request can't be more than 10
>> seconds.
> 
>> On 10/26/06 3:02 PM, "David Rees" <drees76(at)gmail.com> wrote:
> 
>>>> Also one more thing - how can I check in my shell script that all requests
>>>> to
>>>> pound are
>>>> finished ??
>>> 
>>> Check for established connections to the port that pound is bound to.
> 
> 
> If somebody is downloading 100mb file it could take a long time :)
> With about 200conn/sec there are nearly 50connections to big files, so
> it is almost impossible to restart pound without breaking connections.
> 
> The best option will be to reload config without restarting pound, but
> I have no idea on how to do this.
> 

> Just checking on assumptions here, but why is it that you need to reload the
> pound configuration every 10 minutes?  That seems kind of excessive.  I
> still think that pound should behave better in this regard, but I am still
> surprised that you need to reload that often.

> Is there something that you could generalize in your configuration, perhaps,
> that would make much of this reloading unnecessary?


> On 10/26/06 3:52 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>>> Or just wait 30 seconds if your longest request can't be more than 10
>>> seconds.
>> 
>>> On 10/26/06 3:02 PM, "David Rees" <drees76(at)gmail.com> wrote:
>> 
>>>>> Also one more thing - how can I check in my shell script that all
requests
>>>>> to
>>>>> pound are
>>>>> finished ??
>>>> 
>>>> Check for established connections to the port that pound is bound to.
>> 
>> 
>> If somebody is downloading 100mb file it could take a long time :)
>> With about 200conn/sec there are nearly 50connections to big files, so
>> it is almost impossible to restart pound without breaking connections.
>> 
>> The best option will be to reload config without restarting pound, but
>> I have no idea on how to do this.
>> 


I'm doing Virtual hosts managing in pound, thats why I have to reload
config every 10minutes. I know that pound is not dedicated to manage
VH, but I have no other way to do this at backends.

Do hosts move from backend to backend that often?

If not, why not give each host a range of first letters and split things
evenly using regexps?

For that matter you could even go a bit overboard and have extra patterns
available that would assign different first letters differently.

For instance if you have 6 backends, you could define 10 backends for each
possible first letter in the host name.  Then by enabling or disabling the
correct hosts, you have just about all the control you might like with room
for nearly 70% expansion in the 4 backends that do not yet exist.

You could then write a simple front-end for poundctl that would manipulate
all of this for you.


On 10/26/06 11:22 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>> Just checking on assumptions here, but why is it that you need to reload the
>> pound configuration every 10 minutes?

> I'm doing Virtual hosts managing in pound, thats why I have to reload
> config every 10minutes. I know that pound is not dedicated to manage
> VH, but I have no other way to do this at backends.
> 

> Do hosts move from backend to backend that often?

> If not, why not give each host a range of first letters and split things
> evenly using regexps?

> For that matter you could even go a bit overboard and have extra patterns
> available that would assign different first letters differently.

> For instance if you have 6 backends, you could define 10 backends for each
> possible first letter in the host name.  Then by enabling or disabling the
> correct hosts, you have just about all the control you might like with room
> for nearly 70% expansion in the 4 backends that do not yet exist.

> You could then write a simple front-end for poundctl that would manipulate
> all of this for you.


> On 10/26/06 11:22 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>>> Just checking on assumptions here, but why is it that you need to reload
the
>>> pound configuration every 10 minutes?

>> I'm doing Virtual hosts managing in pound, thats why I have to reload
>> config every 10minutes. I know that pound is not dedicated to manage
>> VH, but I have no other way to do this at backends.
>> 


It is not so simple. I have 500users per one backend, and each user
have different domain, but also domain can ba shared:
for example user on backend 1 has assigned: subdomain1.domain.com
user on backend 2 has assigned subdomain2.domain.com
user on backend 3 has assigned subdomain3.domain.com
etc
In this case we cannot manage backend based on domain patterns, but
only on full domain name.

Write an extension to pound that reads from a sql-lite database which domain
goes to which backend.

That would give you everything you need.


On 10/27/06 12:32 AM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>> Do hosts move from backend to backend that often?
> 
>> If not, why not give each host a range of first letters and split things
>> evenly using regexps?
> 
>> For that matter you could even go a bit overboard and have extra patterns
>> available that would assign different first letters differently.
> 
>> For instance if you have 6 backends, you could define 10 backends for each
>> possible first letter in the host name.  Then by enabling or disabling the
>> correct hosts, you have just about all the control you might like with room
>> for nearly 70% expansion in the 4 backends that do not yet exist.
> 
>> You could then write a simple front-end for poundctl that would manipulate
>> all of this for you.
> 
> 
>> On 10/26/06 11:22 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:
> 
>>>> Just checking on assumptions here, but why is it that you need to reload
>>>> the
>>>> pound configuration every 10 minutes?
> 
>>> I'm doing Virtual hosts managing in pound, thats why I have to reload
>>> config every 10minutes. I know that pound is not dedicated to manage
>>> VH, but I have no other way to do this at backends.
>>> 
> 
> 
> It is not so simple. I have 500users per one backend, and each user
> have different domain, but also domain can ba shared:
> for example user on backend 1 has assigned: subdomain1.domain.com
> user on backend 2 has assigned subdomain2.domain.com
> user on backend 3 has assigned subdomain3.domain.com
> etc
> In this case we cannot manage backend based on domain patterns, but
> only on full domain name.
> 

> Write an extension to pound that reads from a sql-lite database which domain
> goes to which backend.

> That would give you everything you need.


> On 10/27/06 12:32 AM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:

>>> Do hosts move from backend to backend that often?
>> 
>>> If not, why not give each host a range of first letters and split things
>>> evenly using regexps?
>> 
>>> For that matter you could even go a bit overboard and have extra patterns
>>> available that would assign different first letters differently.
>> 
>>> For instance if you have 6 backends, you could define 10 backends for each
>>> possible first letter in the host name.  Then by enabling or disabling the
>>> correct hosts, you have just about all the control you might like with room
>>> for nearly 70% expansion in the 4 backends that do not yet exist.
>> 
>>> You could then write a simple front-end for poundctl that would manipulate
>>> all of this for you.
>> 
>> 
>>> On 10/26/06 11:22 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:
>> 
>>>>> Just checking on assumptions here, but why is it that you need to reload
>>>>> the
>>>>> pound configuration every 10 minutes?
>> 
>>>> I'm doing Virtual hosts managing in pound, thats why I have to reload
>>>> config every 10minutes. I know that pound is not dedicated to manage
>>>> VH, but I have no other way to do this at backends.
>>>> 
>> 
>> 
>> It is not so simple. I have 500users per one backend, and each user
>> have different domain, but also domain can ba shared:
>> for example user on backend 1 has assigned: subdomain1.domain.com
>> user on backend 2 has assigned subdomain2.domain.com
>> user on backend 3 has assigned subdomain3.domain.com
>> etc
>> In this case we cannot manage backend based on domain patterns, but
>> only on full domain name.
>> 

Nice idea.

Hi

I think pound creates a list of precompiled regular expressions at 
startuptime, kept in memory for speed. These REs are applied to the URL 
and headers of each request to determine which Service is to be used. 
Fetching this from some kind of sql server in an equally efficient manner 
may be a bit involved, I guess.

Extending poundctl to being able to alter the URL, HeadRequire and 
HeadDeny RE's will probably be much simpler, but you will probably have 
some synchronisation and session problems there as well.

If session persistence is of little importance I think the simplest fix 
would be making pound close down the listening sockets immediately when it 
is signalled to shut down. Then a new pound instance can start 
immediately, even if the old one is still processing existing tcp 
connections. A restart of pound would then require just milliseconds of 
downtime.

Regards
Rune

---
Rune Sætre <rune.saetre(at)netcom-gsm.no>
NetCom as
..

On Fri, 27 Oct 2006, ForAll.pl - Firma wrote:

>
>> Write an extension to pound that reads from a sql-lite database which domain
>> goes to which backend.
>
>> That would give you everything you need.
>
>
>> On 10/27/06 12:32 AM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:
>
>>>> Do hosts move from backend to backend that often?
>>>
>>>> If not, why not give each host a range of first letters and split things
>>>> evenly using regexps?
>>>
>>>> For that matter you could even go a bit overboard and have extra patterns
>>>> available that would assign different first letters differently.
>>>
>>>> For instance if you have 6 backends, you could define 10 backends for each
>>>> possible first letter in the host name.  Then by enabling or disabling the
>>>> correct hosts, you have just about all the control you might like with
room
>>>> for nearly 70% expansion in the 4 backends that do not yet exist.
>>>
>>>> You could then write a simple front-end for poundctl that would manipulate
>>>> all of this for you.
>>>
>>>
>>>> On 10/26/06 11:22 PM, "ForAll.pl - Firma" <firma(at)forall.pl> wrote:
>>>
>>>>>> Just checking on assumptions here, but why is it that you need to reload
>>>>>> the
>>>>>> pound configuration every 10 minutes?
>>>
>>>>> I'm doing Virtual hosts managing in pound, thats why I have to reload
>>>>> config every 10minutes. I know that pound is not dedicated to manage
>>>>> VH, but I have no other way to do this at backends.
>>>>>
>>>
>>>
>>> It is not so simple. I have 500users per one backend, and each user
>>> have different domain, but also domain can ba shared:
>>> for example user on backend 1 has assigned: subdomain1.domain.com
>>> user on backend 2 has assigned subdomain2.domain.com
>>> user on backend 3 has assigned subdomain3.domain.com
>>> etc
>>> In this case we cannot manage backend based on domain patterns, but
>>> only on full domain name.
>>>
>
> Nice idea.
>
>
>
> -- 
> To unsubscribe send an email with subject 'unsubscribe' to pound(at)apsis.ch.
> Please contact roseg(at)apsis.ch for questions.
>
http://www.apsis.ch/pound/pound_list/archive/2006/2006-10/1161885823000/1161941495000
>