netfs

i think we should track GPU resource usage under `kGPU` to help differentiate from main. monitoring tools can aggregate (sum/rates etc.) as desired.

fixed in https://github.com/tensorflow/serving/commit/b15210ca076b11eaa2dfd0ac2fb125be780c5d40

what happens if your serving process restarts (while the underlying filesystem or directory has been deleted). that still will be a disaster. no? IOW this fix is only useful until...

Thanks for the bug report. Happy to review a PR to fix this problem.

@njhill do you want wrap this PR by adding unit-test as requested by the reviewer? thanks!

no worries @njhill. @astleychen do you want to help here and add tests?

@Wentong-DST can you help resolve the conflicts, so we can merge?

the fix looks correct, but i am puzzled why the instructions here: https://github.com/tensorflow/serving#serve-a-tensorflow-model-in-60-seconds do not yield any error. do you know why?

There is no reason to not support it. Happy to accept+review PRs to support DT_HALF.

there is https://github.com/tensorflow/serving/pull/1753 for this under review .