Hi Christopher, I was out last week, but we were able to gather a few more pieces of data.
1) We ran some tcpdumps, and nothing odd popped up at all. Given our traffic levels, we were actually surprised how few RST, etc. we had. Any TCP issues that did occur did *not* coincide with the SD-- log lines. 2) We've confirmed this also happens with similar rates on publicly served traffic, as well (which mostly uses varnish as the backend). The symptom is a slightly truncated response body. That left us with the upgrade to 3.0.5 as the culprit, so we're going to need to rollback to 3.0.4 for now. If we have the time, we'll run some more tests on a limited subset of servers to see if we can find out anything more. Best, Luke — Luke Seelenbinder Stadia Maps | Founder & CEO stadiamaps.com > On Sep 27, 2024, at 10:30, Christopher Faulet <cfau...@haproxy.com> wrote: > > Le 26/09/2024 à 17:59, Luke Seelenbinder a écrit : >> Hi again, >> Sorry, it's the opposite direction of that. Varnish is the client in this >> case (it sits behind a few other pieces, but that's the relevant for the >> bug). Full flow: >> Client -> HAProxy -> Varnish -> HAProxy -> HAProxy Upstream application. >> That make more sense? :) > > > It make more sens indeed. Thanks. > > So HAproxy detects a server abort for an unknown reason since the 3.0.5. On > Varnish side, the response appears as truncated because it reports a 503 (At > least I guess it is what happens). At this stage, I guess it could be easier > to track the issue on Github. > > Outside this bug, have you many unsuccessful termination states in your logs > (other than '----' in fact) ? If not, it could be possible to enable H1 > traces on error level. Let me know. In the mean time, I will review the 3.0.5 > changes... > > -- > Christopher Faulet > > >
