frontend: use experiment ID from path, not query

[wchargin]

Summary:
This commit changes the frontend experiment ID plumbing from using the
query string to using a path component: experiment URLs are now of the
form http://localhost:6006/experiment/123/. This supersedes the
short-term changes from #2580, #2613, and #2627.

Test Plan:
Apply the below patch to the multiplexer data provider, then launch
TensorBoard with --generic_data=true and verify that the correct
experiment ID is printed for each call, at / and at /experiment/123.
Verify also that no request has an experiment query parameter. Patch:

diff --git a/tensorboard/backend/event_processing/data_provider.py b/tensorboard/backend/event_processing/data_provider.py
index 0b354aa8..1c00b930 100644
--- a/tensorboard/backend/event_processing/data_provider.py
+++ b/tensorboard/backend/event_processing/data_provider.py
@@ -59,2 +59,3 @@ class MultiplexerDataProvider(provider.DataProvider):
   def data_location(self, experiment_id):
+    logger.warn("data_location(%r)", experiment_id)
     del experiment_id   # ignored
@@ -63,2 +64,3 @@ class MultiplexerDataProvider(provider.DataProvider):
   def list_runs(self, experiment_id):
+    logger.warn("list_runs(%r)", experiment_id)
     del experiment_id  # ignored for now
@@ -74,2 +76,3 @@ class MultiplexerDataProvider(provider.DataProvider):
   def list_scalars(self, experiment_id, plugin_name, run_tag_filter=None):
+    logger.warn("list_scalars(%r, ...)", experiment_id)
     del experiment_id  # ignored for now
@@ -105,2 +108,3 @@ class MultiplexerDataProvider(provider.DataProvider):
   ):
+    logger.warn("read_scalars(%r, ...)", experiment_id)
     # TODO(@wchargin): Downsampling not implemented, as the multiplexer

wchargin-branch: eid-frontend

[stephanwlee]

I think I am missing one critical information. Does this change require the document to be served at localhost:6006/experiment/[:eid]/?

1. I think code makes sense.
2. currently TB FE is not that experiment aware. When it increasingly is aware of the experiment, I think it is inevitable that we will introduce experiment param back to the router.
3. there should be separate router implementation from different implementations of the server.

Because (1) and (2) is a future work, this change LGTM.

[wchargin]

I’m not sure that I fully understand the question. The frontend already
assumes that it’s served from /, because otherwise the relative URLs
will resolve incorrectly. After this change, we still support loading
from / (with empty experiment ID) and additionally support loading
from /experiment/EID/. Does this answer your question?

If you’re asking about caching of the large HTML blob of the document
itself: It’s true that this won’t have a shared cache across two
experiments, but that was also true for the query parameter version. In
either case, we can improve the caching by loading the bundle (or parts
of it) dynamically on the frontend.

When it increasingly is aware of the experiment, I think it is
inevitable that we will introduce experiment param back to the
router.

I agree that this is plausible, especially because you can’t just route
to ../456/ to change the experiment ID, because when you’re at /
instead of /experiment/123/ this would do the wrong thing.