Symptom

I have some docker-in-docker containers (based on kubeadm-dind-cluster) as part of a CI system where, with Docker 17.12-ce:

$ docker exec -ti <container> echo hello
hello
<hang>

By <hang>, I mean that control is not returned to the original terminal. This in turn causes our CI to hang.

Reproducibility

This started happening after tens of such containers were created, run and destroyed.

Once the system gets into this state, it does not recover, i.e. further docker execs into affected containers will hang in a similar way.

The same docker exec on the same setup on Docker 17.06-ce does not cause hangs, which seems to be working more reliably.

Context

This is on Ubuntu 17.10 on a fast dedicated server from OVH.

I downgraded the system which was experiencing this issue to Docker 17.06, which is now in production as a CI runner, so it will be a bit of work to reproduce it (would have to set up a brand new machine and try to reproduce the issue there).

cc @justincormack who requested this issue