Re: Mongrel 2.0

["Evan Weaver" <evan-72XWLPH10WVXUHR/Jj/Uug@public.gmane.org>]

Whether the mongrel is overloaded such that responding further
degrades service depends on whether you are blocked on local CPU or on
a shared resource. In my experience, shared resource problems are more
common.

Evan

On 11/18/08, Kirk Haines <wyhaines-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> On Mon, Nov 17, 2008 at 10:36 PM, Zed A. Shaw <zedshaw-dd7LMGGEL7NBDgjK7y7TUQ@public.gmane.org> wrote:
>> Very cool, nice work.  One small comment...
>>
>> On Mon, Nov 17, 2008 at 03:32:33PM -0500, Evan Weaver wrote:
>>> Hi Mongrels,
>>>
>>> - Error code instead of connection-close, as discussed earlier
>>
>> With this do you mean returning one of the many error codes in an HTTP
>> response when the server is overloaded?
>>
>> I would recommend against that if that's the case.  In theory it's nice
>> to return an error code to clients that could be waiting.  In practice,
>> your queue is already overloaded, so taking more time to send a response
>> to clients only adds to the problem.  In a modern system this isn't such
>> a big deal, but in Ruby it becomes a mess because of the file id limits
>> in the select loop.
>
> In an overload situation, I think that just dropping the connection is
> an acceptable thing to do, though.  The mongrel is overloaded, so it
> is a correct and beneficial response on the part of the proxy to
> assume something is wrong with it, and to stop sending traffic to it
> for a while.
>
> The use case for sending a response instead of just dropping the
> connection is when the HTTP parser throws an exception because of
> malformed HTTP.
>
> Currently that results in an immediate closed connection.  The problem
> there is that upstream proxies can assume that getting the connection
> cut unexpectedly means that the mongrel itself has a problem.  If that
> proxy responds to this by removing that mongrel from the proxy
> rotation for a period of time, one misbehaving client, whether
> intentionally or unintentionally, can DOS a whole cluster.
>
> The fix is to just return a canned 400 response before closing.
>
>
> Kirk Haines
> _______________________________________________
> Mongrel-development mailing list
> Mongrel-development-GrnCvJ7WPxnNLxjTenLetw@public.gmane.org
> http://rubyforge.org/mailman/listinfo/mongrel-development
>


-- 
Evan Weaver