On Feb 8, 2010, at 5:25 PM, Graham Dumpleton wrote:

I should add that you should also read:

 http://blog.dscpl.com.au/2009/03/load-spikes-and-excessive-memory-usage.html


I have read this a few times... I feel fairly enlightened, but it still doesn't explain how I'm losing daemons.


Because by default the Python interpreter still gets created in those
processes, is probably why you are getting confused. That is, those
processes wouldn't be replaced straight away, unlike the daemon mode
processes which would be as they are part of a static pool size,
whereas main Apache server child processes are effectively part of a
dynamic pool size.


But I guess what I'm seeing is that the daemons really are failing to recirculate, or something - I'm still in this world where we're suddenly at 2-3 daemons, when we started with 24. (I should add that we chose 24 because we do have big beefy machines with lots of RAM and cores, plus this particular application tends to be CPU-heavy)

Thus, setting WSGIRestrictEmbedded and disabling default behaviour
that sees Python interpreter still initialised in those processes may
clear things up.


I'll definitely give that a try... at least it should further reduce the number of log messages that may be confusing me.

Alec

Graham

On 9 February 2010 10:56, Graham Dumpleton <[email protected]> wrote:
If you are not using embedded mode, ie., only using daemon mode, then
add the directive:

 WSGIRestrictEmbedded On

This will tell mod_wsgi not to bother to initialise the Python
interpreter in the Apache server child processes, given it will not be
required.

This presumes mod_wsgi 3.X is being used as 2.X behaves differently.

That should eliminate those messages and make it clearer what is going on.

I will explain more later when have the time to catch up on all my email.

Graham

On 9 February 2010 10:42, Alec Flett <[email protected]> wrote:

Ok, I think I'm starting to get a handle on whats going on.

For background, we run in prefork mode. We currently have:
StartServers         5
MinSpareServers      5
MaxSpareServers     10
ServerLimit         600
MaxClients          600
MaxRequestsPerChild  1000

For mod_wsgi I've got maximum-requests=1000

For a bunch of PIDs, these are the mod_wsgi log messages I see:
pids: 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 434,
473,
Initializing Python.
Attach interpreter ''.
Destroying interpreters.
Cleanup interpreter ''.
Terminating Python.
Python has shutdown.

Now I did some exploring and it turns out those PIDs are apache children,
NOT mod_wsgi daemons.

I think that apache is quietly shutting down apache children, perhaps when they reach MaxRequestsPerChild, and this is taking the mod_wsgi children down with them, and mod_wsgi is not restarting those children. Could there possibly be some off-by-one bug where if we're on the 1000th request, mod_wsgi thinks "kill this child, and restart it" but then apache comes in
and kills the child just before it starts?

Alec

On Feb 8, 2010, at 11:38 AM, Alec Flett wrote:

Ok, I've now found wsgi_manage_process...

FWIW I haven't been able to reproduce the crash by calling
os.kill(os.getcwd(), signal.SIGBUS) and frankly I'm not even sure how specifically our children are crashing, if it's a SIGBUS or something else. all I know is the state I find the appserver in and there's little to
nothing from the logs

I'm going to keep digging...

Alec

On Feb 8, 2010, at 10:35 AM, Alec Flett wrote:

So I'm still seeing this problem - that our python processes are crashing for some reason (our problem, I'm sure) but mod_wsgi isn't restarting them.

I just perused the mod_wsgi.c source and I don't see anything that would restart children if they crashed? In particular, I don't see anything catching SIGCHLD but I'm willing to believe the the apr_ APIs are doing
this in a different way.

Also is there some kind of scoreboard telling which children are
available to recieve new requests? Because the server continues to serve requests except for the missing children, leading me to believe mod_wsgi has somehow figured out that the dead children are not allowed to handle new
requests.

Can you point me at the crash-recovery code?

Alec

On Jan 28, 2010, at 9:51 PM, Graham Dumpleton wrote:

2010/1/29 Alec Flett <[email protected]>:

On Jan 27, 2010, at 3:07 PM, Graham Dumpleton wrote:

Should restart on a crash automatically.

One cause of what you are seeing is Python threads being deadlocked
and over time causing available threads to be used up.

Are you using multithread daemons? Is your code and third party
modules thread safe?


nope, single-threaded! threads=1 on the WSGIDaemonProcess line.

Try setting 'inactivity-timeout=120' as option to WSGIDaemonProcess.


great, that seems like a good idea anyway.

I would also suggest setting LogLevel to 'info' so that additional
information printed out in error logs about process restarts.

That was going to be my next question ...:)


This way you might get an idea what request threads are actually
doing.

So none of this explains the "missing daemons" problem - where the
daemons
are not actually starting back up again... as you can see below, I set
the
display-name so that I can look at the daemons with "ps" - when I do a
ps ax
| grep <group> I only see a few processes

The extra level of logging may show if processes are doing some sort
of shutdown. If they are crashing, then you should already see
segmentation fault messages in main Apache error log, not virtual
host, so make sure you check both logs.

The processes should be restarted if they truly exit or crash. If it is an order process restart due to maximum requests or WSGI script
file being touched, there is also a fail safe which defaults to 5
seconds. If it doesn't die in that time a thread should cause it to kill itself. The only way this would work in that way is if some C extension module for Python had registered a competing C code level signal handler or blocked signals and it interfered with mod_wsgi. In that case though the process would still exist and you should still
see it.

If it was an Apache restart that triggered process restart, you
presumably would have known about that unless you have some automated system which does that. Even so, Apache will kill any daemon process
off which don't shut down in 3 seconds.

Can't also be case that processes are zombies, because that would mean
Apache isn't doing wait on their exit code, which it should be.

So, all quite confusing.

(in fact one of my servers in
production has dropped from the original 24 process, down to 7
yesterday,
and now only at 3 today!)

Unless you have long lived requests, 24 process is actually quite a
lot. Any well tuned system should manage with a lot less.

Even with that number of processes, since not multithreaded, unless you have a problem in your code with not releasing file descriptors, wouldn't expect to run out of resources. You might though use lsof or ofiles or other tool to work out if large number of file descriptors in use. Even then, if Apache/mod_wsgi can't restart processes because
of that, you should see error messages in main Apache error log.

Graham

Let me know what you find and also post your actual daemon mode
configuration.


Here's one of them:

#############################
# Project: client
##############################

WSGIDaemonProcess client-freebase.com processes=24 threads=1
display-name=%{GROU
P}
python-path=/mw/app/client_88277/_install/lib/python2.6/site- packages
maximum
-requests=1000

WSGIScriptAlias / /mw/app/client_88277/_install/bin/client.wsgi

# Server configuration for client
<Directory /mw/app/client_88277/_install/bin>
WSGIProcessGroup client-freebase.com
</Directory>



Graham

--
You received this message because you are subscribed to the Google
Groups
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.


--
You received this message because you are subscribed to the Google
Groups
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.



--
You received this message because you are subscribed to the Google
Groups "modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.


--
You received this message because you are subscribed to the Google Groups
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.


--
You received this message because you are subscribed to the Google Groups
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.


--
You received this message because you are subscribed to the Google Groups
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/modwsgi?hl=en.




--
You received this message because you are subscribed to the Google Groups "modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to [email protected] . For more options, visit this group at http://groups.google.com/group/modwsgi?hl=en .


--
You received this message because you are subscribed to the Google Groups 
"modwsgi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to 
[email protected].
For more options, visit this group at 
http://groups.google.com/group/modwsgi?hl=en.

Reply via email to