I asked
>> I saw in the list archives, discusions on allowing a simple
>> dot file approach, to mark servers as down, or otherwise unavailable.
>> Not sure whether that feature made it into current version or not?

Robert replied
> It didn't. One of the basic design principles for Pound is that it
> never, ever touches the file system. We take security seriously!

Reading through the man pages, I discovered the logic of pound loading
defaults on boot, and from then on, never reading or writing to disk
again, other than syslog, which is not really reading or writing to disk.

Question - is a text copy of the lastest pound man page available on the
website? If not, could I suggest it be made available there?




I asked
>> Also, it lead me to consider the issue of how pound can tell
>> the difference between a lightly loaded server, and one under
>> stress which should be avoided? I'm guessing there is no provision
>> for that by the opening definition of what pound is?

Robert replied
> Correct.

Reading the man pages more, it seems you could get a web server to
signal back to pound, through a HA port setup, that pound should consider
this web server offline, by not responding on the locally configured HA port.

This way, the main web server application can/could finish responding to
pending requests. The alternative setup would be to turn the web server
itself off, which would crash sessions etc. So the HA port setup is preferred.


I asked
>> I'm guessing that people doing load balancing would often
>> have servers of different capacity. That's probably one
>> of the reasons they need a load balancer! And as people add
>> more servers, it's likely that they would be newer, and be
>> able to handle more.

Robert replied
> Have a look at the "priority" parameter for the back-end definition. It
> does what you want.

From the man pages, I can now see how to do that. Was a little surprised
at the limitation of only 9 steps, (1 - 9) but I guess that works fine.



Bearing all those points in mind, I'm guessing that an overloaded
web server, could signal back to pound, that it should be ignored for
a while, by stopping answers to queries on the HA port?

That seems a little black and white to me, either on or off.
I'm wonding if the HA port actual response is currently  interpreted by pound
in any way, or does it only consider any response, or no response,
as the indicator on whether to use that server or not?

I'm wondering if a variably loaded server, couldn't signal back to
pound, the current status, via the actual response on the HA port?
So roughly speaking, if the HA port said back to a query on that
port, '5' then pound would _reset_ that server priority setting to
now be 5, rather than whatever it was before?

Or you could do something like a difference figure, so that '+1' would
add 1 to the default predefined priority setting, or '-2' would take
2 off.

I'm guessing that there's more to this than my initial readings of the
docs I've found :-)

It seems to me, that using the  HA port to indicate that a server should
be ignored, would certainly work fine. So I'll probably head down that
path to start with.

But it seems to me that a bunch of servers with pre-defined priority
settings, is a bit 'set in stone' so to speak. In my experience, often
server loads bounce all over the place, and a variable way of signalling
back to the load balancer (pound) that "I can handle more than normal"
or "I am overloaded a bit, please give me fewer requests" would be a
good thing.

I realise that this would add to the complexity of pound a little, and require
a more complicated HA port response, and nobody but me may see the
effort as worthwhile or not.

Any thoughts?

am looking forward to testing pound :-)

regards
mark