Carries the two fixes from v2.17.10. If you are on :beta with an ssh:// host, take this one.
Fixed
An install with an ssh:// host runs itself out of processes. Every host reports could not list containers (rc=2): runtime/cgo: pthread_create failed: Resource temporarily unavailable — the tcp:// ones and the local socket included, which have nothing to do with ssh. A live four-host install held 12074 defunct ssh processes and could not fork at all.
The connection reuse added in 2.17.6 leaves an ssh master in the background; 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. It runs under tini now, which reaps them. The reuse stays: it was the trigger, not the cause. The full reasoning, including why this is not a signal handler of our own, is in the 2.17.10 notes.
Restarting clears what has piled up; updating stops it coming back.
A host that is down was reported as the local machine in /api/status and /metrics, while the one that actually failed was left out — an alert about the wrong box on exactly the multi-host installs where a host being off is normal.