[5k4nd]

Fix RuntimeError: dictionary changed size during iteration in NodePool.get()

Fixes #302

NodePool.get() iterates over _alive_nodes while other threads sharing the client may mutate it through
mark_dead(), mark_live() or resurrect(), which raises:

  File "elastic_transport/_transport.py", line 338, in perform_request
    node = self.node_pool.get()
  File "elastic_transport/_node_pool.py", line 361, in get
    nodes = [
RuntimeError: dictionary changed size during iteration

We hit it in production with a client shared by the threads of a multi-threaded WSGI server.

The fix iterates over a copy of _alive_nodes. dict.copy() is atomic under the GIL, so no lock is needed.
The added test is deterministic: it mutates _alive_nodes during the iteration instead of relying on threads.

[cla-checker-service]

💚 CLA has been signed

[github-actions]

This pull request is stale because it has been open 7 days with no activity. Remove stale label or comment or this will be closed in 2 days.

[margaretjgu]

Thanks for the contribution. Please open up an issue first so I can validate the context of the problem and for our tracking.

[5k4nd]

@margaretjgu thanks for your reply, I just opened an issue #302
I also signed the CLA a week ago.