Christopher Faulet

Captures are somewhat old and some parts are not really usable. For instance, It is not really handy to mix captures via TCP/HTTP rule and declared captures , or captures...

Indeed, functionally speaking, captures are not explicitly related to HTTP. However, the sample fetches suggest it is only usable in an HTTP context. We can capture anything from a tcp-request...

That smells bad. You may indeed try to disable the zero-copy forwarding. For now, I have no idea. Are the scrambled data in response payload or in the headers part...

Ok, I'm not surprised. Zero-copy only concerns the payload transfer. Do you have any http rules manipulating the headers ? You have changed the issue title to specify all 3.0/2.9...

There is no obvious change on this area between 2.9 and 2.8 that may explain this bug. At least, at first glance. So, it may a side effect of something...

Just something strange, if the request is perform in HTTP/1.0, a "Connection" header should be added by Haproxy in the response with "close" or "keep-alive" value depending on the request....

Hum nothing fancy ...

Could you confirm it still happens on 3.0.4/2.9.10 ?

@awlx, Have you had the opportunity to test the 3.0.4 or better, the 3.0-HEAD, to check if the issue is still there ?

Following the @awlx's comment on #2749, is the bug still there on the 3.0.4/3.0.5 ?