Straight to stable, no beta: on an install with an ssh:// host this takes about a day and then nothing works at all.
Fixed
Every host suddenly "could not be checked" — including the ones that do not use ssh.
The message was could not list containers (rc=2): runtime/cgo: pthread_create failed: Resource temporarily unavailable, on all four hosts of a live install at once, the two tcp:// ones and the local socket among them. Nothing was wrong with any of those hosts. The container held 12074 defunct ssh processes, 9839 of them from a single hour, and could no longer start a process at all.
They come from the connection reuse added in 2.17.6. ControlPersist leaves an ssh master in the background so the next call does not pay the 355 ms handshake again; when the docker client exits, that master is handed to PID 1 — to Docksentry — and Python never waits for a child it did not start. One left-behind process per connection, each holding a slot forever. One ssh host filled the table in a little over a day.
The reuse is not the bug and stays. It only made visible that Docksentry has been running as PID 1 without ever doing PID 1's job. It now runs under tini, which reaps them. Measured against the built image with 200 orphans: 200 left behind before, none after.
Deliberately not a signal handler of our own: one calling waitpid(-1) races the update path for its own children and can take an exit status out from under it — an update would then report a success it never had, which is worse than the failure being fixed.
If you run an ssh:// host this affects you, and has since 2.17.6. Restarting the container clears what has piled up; updating stops it coming back. tcp://, context:// and single-host installs are not affected.
A host that is down was reported as the local machine. With one endpoint switched off, /api/status and /metrics said the local machine was unreachable — while listing its running containers — and left out the host that actually was. An alert about the wrong box and none about the right one, on exactly the multi-host installs where a host being off is normal. In every release since 2026-08-01.