[Xapian-devel] Problems with /bin/cat and flintlock?

[Samuel Williams]

Hi,

I just did a synthetic test on my server and it didn't cause a problem, so there is something else contributing to the problem.

I will investigate further.

Kind regards,
Samuel

On 24/05/2011, at 4:09 PM, Olly Betts wrote:

> On Tue, May 24, 2011 at 02:49:56PM +1200, Samuel Williams wrote:
>> In my case, I've certainly noticed that the bug is intermittent, it
>> doesn't always lock up passenger - maybe it is something to do with
>> trying to create multiple instances of WritableDatabase in different
>> processes? I will do further investigation.
> 
> I don't really see how that could matter, unless you're trying to
> use the same WritableDatabase object in more than one process
> (which isn't supported).
> 
>> Using SIGKILL will definitely do the trick, but it does feel to me to
>> be heavy handed - on the other hand, it might also be the correct
>> approach because if for some reason, the child process gets stuck, no
>> amount of SIGHUP is going to help. Well, I've seen OS bugs where even
>> SIGKILL doesn't work but obviously that isn't something we need to be
>> concerned with hopefully.
> 
> We want to make sure the child process exits, and I think SIGKILL is
> the appropriate way to do that.  If it fails, chances are you're
> not going to be able to get it to exit in any other way.
> 
>> So:
>> (1) I can spend some time trying to figure out if that test case
>> reproduces the bug
>> OR
>> (2) We can change to SIGKILL and be done with it.
>> 
>> Well, I'd like to have a solid test case first, so I'd go for (1), but
>> also my time is limited, which means that to fix the problem soon (2)
>> is the best approach. What do you think?
> 
> The testcase should work by my understanding of the bug, but it doesn't
> work, so that means I don't understand the bug correctly.  Since it
> appears I don't understand the bug, it's hard to know if the fix really
> is right.  And even if it is, without a regression test for it we won't
> get a test failure if this issue is reintroduced in the future.
> 
> So I'd really like to get a working testcase for it, though changing to
> SIGKILL is pretty clearly at least as good as what we currently do, so
> indefinitely blocking that change on having a testcase doesn't make much
> sense either.
> 
> Cheers,
>    Olly