serving
serving copied to clipboard
tensorflow
Tensorflow Serving should handle SIGTERM correctly
TF Serving should terminate gracefully when SIGTERM is received.
This is especially important for docker / kubernetes use cases when a process is terminated gracefully or is killed have very different meanings for determinating a workload success status.
Related issues :- #799,#498,#356.
All closed and unsolved... I looked it up before
What kind of server doesn't handle signals, it's basic functionality of any server. I was really surprised to find out that tensorflow-serving just ignores it.
Ideally SIGTERM should do a graceful shut down where all existing connections are served by shutting down.
Otherwise the relevant signal should be set in the Dockerfile.
Bumping issue again. Are there any plans to support graceful shutdown for tensorflow-serving? As already been said - for hosting tensorflow_serving in k8s this is must have feature for easy & smooth deployments rollout and scale up/scale down events.
CC @nniuzft
