You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Summary:
TensorBoard can now start RustBoard if `--load_fast` is given at runtime
and `--define=link_data_server=true` was given at build time. The
default configuration still has no Rust code, so our Pip packages are
still portable. When we actually deploy this, we can distribute the Rust
binary in a separate Pip package that TensorBoard imports, but the UX
can stay the same: “just add `--load_fast`.”
Test Plan:
Test in the following configurations:
- with just a `--logdir`, everything works as normal;
- with `--define=link_data_server=true -- --logdir ... --load_fast`,
the subprocess is spawned and cleaned up on a successful exit or a
SIGKILL to TensorBoard, and INFO logs are shown iff `--verbosity 0`
is specified;
- with `--logdir ... --load_fast` but no `--define`, TensorBoard fails
to start the server and prints a message before exiting; and
- with `--grpc_data_provider localhost:6806`, TensorBoard connects to
an existing server without needing `--logdir` or `--define`.
To test the “data server died” case, comment out the `--logdir=%s` flag,
which will cause the server to fail with a usage message. That message
should appear in the logs.
To test the polling, add `thread::sleep(Duration::from_secs(3))` before
the server writes its port file, and run with `--verbosity 0` to note
the “Polling for data server port” messages.
This also works after syncing into Google, in all relevant
configurations; see <http://cl/344900410> and <http://cl/344955833>.
wchargin-branch: grpc-ingester
0 commit comments