Check versions on connection

[mrocklin]

A decent fraction of the errors that we see (15%?) are due to mismatched versions. We might consider silently running client.get_versions(check=True) on startup and issuing a warning if things don't match well.

[jonnor]

Added a comment about one occurrence of this issue on the PR: #1720 (comment)

[costrouc]

Yes please this would be extremely helpful. When running dask on Kubernetes it is too easy to get a version mismatch from separate conda solves for each docker image (dask worker, dask scheduler, jupyterlab user) without pinning.

[jrbourbeau]

I think this was added in #2627 and is included in distributed >= 2.9.2

[jrbourbeau]

Just confirmed that a warning is logged when there are package version mismatches (when using the current distributed master branch). For example:

distributed.worker - WARNING - Mismatched versions found

distributed
+---------------------------+---------------------------+
|                           | version                   |
+---------------------------+---------------------------+
| This Worker               | 2.11.0+21.g05ca3fb1.dirty |
| scheduler                 | 2.12.0                    |
| tcp://192.168.0.102:64357 | 2.11.0+21.g05ca3fb1.dirty |
+---------------------------+---------------------------+

msgpack
+---------------------------+---------+
|                           | version |
+---------------------------+---------+
| This Worker               | 0.6.1   |
| scheduler                 | 1.0.0   |
| tcp://192.168.0.102:64357 | 0.6.1   |
+---------------------------+---------+

[jrbourbeau]

Closed by #2627