Leaking pipes in Hgweb

[Stephen Lee]

On Sat, Dec 14, 2013 at 8:26 AM, Matt Mackall <mpm at selenic.com> wrote:
> On Fri, 2013-12-13 at 10:43 +1100, Stephen Lee wrote:
>> We have deployed hgweb using WSGI in Apache. Several other services
>> are also being hosted on the same server.
>>
>> Over time we see open file handles leaking - eventually the limit is
>> reached and Apache gives "Internal server errors".
>> According to lsof the majority of the files are pipes.
>
> How many file descriptors are you seeing? hgweb internally only uses
> pipes for hooks, which won't be invoked in normal browsing, so this
> might be a property of how your WSGI wrapper works.

The ulimit initially was 512 and it was crashing after about 3 days.
We increased it to ~8000 and it has been going up for about a week now.

>
>> We can see the number increase when accessing the hgweb pages, but not
>> when accessing any other pages.
>>
>> Any ideas? Or any other information that could be of help in debugging?
>
> Which Apache MPM are you using? With what settings?
>
> Most people probably don't hit this because they run Apache with finite
> lifetimes per subprocess/thread.
>

I'm not the Apache admin, I just installed hgweb into an existing WSGI setup.
I will check these settings on Monday at work and report back.

Thanks for the suggestions.

Steve