On Wed, Sep 23, 2026 at 11:43:34AM +0200, Amaury Denoyelle wrote: > On Wed, Sep 23, 2026 at 06:20:18PM +0900, ygkat wrote: > > Hi Amaury, > > Thanks for taking a look. > > > I'm not sure to agree yet with the proposed fix as this could have > > > unwanted consequences, most notably with the backend support (which > > > the proposed patch does not support here). However, a localized fix > > > in qcc_recv_reset_stream() is at least mandatory here. > > That's a fair point, and honestly the reason our patch doesn't cover > > qcc_be_is_reusable()/backend semantics is simply that we don't run any > > QUIC/HTTP3 backends ourselves, so we hadn't considered that case at > > all. > > We're currently carrying a live leak in production from this (around > > 1.22M stuck connections on one node before we restarted it), so we're > > planning to apply both what we submitted before (the qcc_is_dead() > > reorder, frontend-only as submitted) and the change below internally, > > regardless of the outcome of the wider upstream discussion. Since > > we don't run QUIC/HTTP3 backends, the backend concern you raised > > doesn't affect our own deployment either way. > I understand your concern given the impact of the leak on your side.
One additional question, what is the uptime of your process which suffers from 1 million stuck connections ? Just to know if this grows rapidly or slowly over time. -- Amaury Denoyelle
