Description

Dask LightGBM will sometimes try to bind to ports that were previously free, but are now used by a different program.

Specifically, it seems that the python code in LGBMDaskRegressor [1] finds open ports, saves the port number and then immediately closes them. After that the C++ layer will try reopening the port [2]. This can go wrong when another program binds to the port between these two steps.

I would say most of my runs succeed but I've ran into the 'LightGBMError: Binding port blah failed' error a handful of times now, and I'm fairly confident the above race condition is the issue.

[1]

Reproducible example

It's kind of hard to reproduce this reliably since it's effectively a race condition. I hope my description of the issue suffices, let me know if I can do more to help.

Environment info

I'm using LightGBM 3.3.2 on Dask.