Repro:
bazel run //tensorboard -- dev upload --logdir /tmp/nope --verbosity 0
(where /tmp/nope is any path; a nonexistent or empty directory will
repro most quickly).
Expected behavior: We start an upload cycle and poll the logdir once
every 5 seconds.
Actual behavior (latest nightly): We poll the logdir with unbounded
frequency. Logs fly off the screen. I’m seeing about 6000 polls per
second on my workstation (SSD). This is bad for CPU, bad for disks, and
bad for your wallet if your logdir is on a remote filesystem.
git bisect says: 9077c10 is the first
bad commit. (This makes sense.)
We can’t release to PyPI stable in this state.