httpd-bugs mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
Subject [Bug 57520] proxy: worker failure disrupts existing connections
Date Thu, 05 Feb 2015 13:26:59 GMT

--- Comment #13 from Yann Ylavic <> ---
(In reply to dk from comment #11)
> Just to confirm - if httpd
> encounters 503 from Jetty under this regime, would it retry a different
> balancer or return 503 to client (with failonstatus set)? The docs are not
> very clear on whether failonstatus would cause the requests to be retried
> and/or if this behavior could be configured...

Yes, any 503 is retried on another BalancerMember, provided at least one is not
in recovery state (forcerecovery=on may be used to force recovery of all the
members otherwise).

However failonstatus plays no role on retry, it only sets the statuses for
which the backend is put into recovery state, but only 503 (or connect() error)
will be retried (moreover 503 is implicit, so failonstatus=503 is redundant
With 502 for example (as per your configuration), it will cause the member to
be put into recovery state for retry seconds but the request will *not* be sent
to another one, and the 502 will reach the client.
Keep in mind that mod_proxy can't decide whether a request can be resent,
particularly non-idempotent ones (as per RFC), you wouldn't want a bank order
to be applied twice (it depends on which way though :) and hence mod_proxy
generally let the browser/user decide.

Finally retry=0 is probably playing a bad role here since when retried, all the
balancer members will be taken into account (and checked against the recovery
period for eligibility), hence retry=0 will force eligibility during the retry,
which may lead to the same member being elected.

> I am going to clean up our proxy settings per your advice. The ones I have
> now were result of trial and error with various suggestions from
> (e.g. I do not care
> about Expect header or HTTP 1.0). I am curious how much of the advice in
> that discussion thread is still applicable with 2.2.29 vs. may have been
> fixed / worked around?

Link between the two tickets is done now.

If needed, this discussion should be continued on the users@ mailing-list.
Bugzilla if for reporting bugs only, not for support.
I'll be happy to contribute, but we should stop here (I realize this is not
what I've done so far, by I guess it's enough).

You are receiving this mail because:
You are the assignee for the bug.

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message