data: expose data_location integration point (#2627)

Summary:
Generic data providers can now control the “data location” string in the
TensorBoard UI, which appears at the bottom of the runs selector:

![Screenshot of the runs selector and its data location string][ss]

As this is only a cosmetic change, it’s provided as an optional method
on the `DataProvider` interface.

[ss]: https://user-images.githubusercontent.com/4317806/64572837-cf7b7e00-d31d-11e9-8fd6-d81d36774172.png

Test Plan:
TensorBoard still seems to work properly with `--generic_data` set to
`false`, `auto`, or `true`, in both logdir and DB modes. Conspicuously
changing the `MultiplexerDataProvider` to always `return "tada"` induces
the appropriate change in the frontend, so the wiring is hooked up.

wchargin-branch: generic-data-location

Author: wchargin

tensorboard/backend/application.py

@@ -138,7 +138,9 @@ def standard_tensorboard_wsgi(flags, plugin_loaders, assets_zip_provider):
         max_reload_threads=flags.max_reload_threads,
         event_file_active_filter=_get_event_file_active_filter(flags))
     if flags.generic_data != 'false':
-      data_provider = event_data_provider.MultiplexerDataProvider(multiplexer)
+      data_provider = event_data_provider.MultiplexerDataProvider(
+          multiplexer, flags.logdir
+      )
 
   if reload_interval >= 0:
     # We either reload the multiplexer once when TensorBoard starts up, or we

tensorboard/backend/event_processing/data_provider.py

@@ -29,14 +29,17 @@
 
 
 class MultiplexerDataProvider(provider.DataProvider):
-  def __init__(self, multiplexer):
+  def __init__(self, multiplexer, logdir):
     """Trivial initializer.
 
     Args:
       multiplexer: A `plugin_event_multiplexer.EventMultiplexer` (note:
         not a boring old `event_multiplexer.EventMultiplexer`).
+      logdir: The log directory from which data is being read. Only used
+        cosmetically. Should be a `str`.
     """
     self._multiplexer = multiplexer
+    self._logdir = logdir
 
   def _test_run_tag(self, run_tag_filter, run, tag):
     runs = run_tag_filter.runs
@@ -53,6 +56,10 @@ def _get_first_event_timestamp(self, run_name):
     except ValueError as e:
       return None
 
+  def data_location(self, experiment_id):
+    del experiment_id   # ignored
+    return str(self._logdir)
+
   def list_runs(self, experiment_id):
     del experiment_id  # ignored for now
     return [

tensorboard/backend/event_processing/data_provider_test.py

@@ -75,7 +75,13 @@ def create_multiplexer(self):
     return multiplexer
 
   def create_provider(self):
-    return data_provider.MultiplexerDataProvider(self.create_multiplexer())
+    multiplexer = self.create_multiplexer()
+    return data_provider.MultiplexerDataProvider(multiplexer, self.logdir)
+
+  def test_data_location(self):
+    provider = self.create_provider()
+    result = provider.data_location(experiment_id="unused")
+    self.assertEqual(result, self.logdir)
 
   def test_list_runs(self):
     # We can't control the timestamps of events written to disk (without
@@ -102,7 +108,7 @@ def FirstEventTimestamp(multiplexer, run):
           return result
 
     multiplexer = FakeMultiplexer()
-    provider = data_provider.MultiplexerDataProvider(multiplexer)
+    provider = data_provider.MultiplexerDataProvider(multiplexer, "fake_logdir")
     result = provider.list_runs(experiment_id="unused")
     self.assertItemsEqual(result, [
         base_provider.Run(run_id=run, run_name=run, start_time=start_time)
@@ -164,7 +170,7 @@ def test_list_scalars_filters(self):
 
   def test_read_scalars(self):
     multiplexer = self.create_multiplexer()
-    provider = data_provider.MultiplexerDataProvider(multiplexer)
+    provider = data_provider.MultiplexerDataProvider(multiplexer, self.logdir)
 
     run_tag_filter = base_provider.RunTagFilter(
         runs=["waves", "polynomials", "unicorns"],

tensorboard/components/tf_backend/environmentStore.ts

@@ -22,7 +22,9 @@ namespace tf_backend {
     private environment: Environment;
 
     load() {
-      const url = tf_backend.getRouter().environment();
+      const url = tf_backend
+        .getRouter()
+        .environment(tf_backend.getExperimentId());
       return this.requestManager.request(url).then((result) => {
         const environment = {
           dataLocation: result.data_location,

tensorboard/components/tf_backend/router.ts

@@ -14,7 +14,7 @@ limitations under the License.
 ==============================================================================*/
 namespace tf_backend {
   export interface Router {
-    environment: () => string;
+    environment: (experiment?: string) => string;
     experiments: () => string;
     pluginRoute: (
       pluginName: string,
@@ -39,7 +39,11 @@ namespace tf_backend {
       dataDir = dataDir.slice(0, dataDir.length - 1);
     }
     return {
-      environment: () => createDataPath(dataDir, '/environment'),
+      environment: (experiment?: string) => {
+        const searchParams = new URLSearchParams();
+        searchParams.set('experiment', experiment || '');
+        return createDataPath(dataDir, '/environment', searchParams);
+      },
       experiments: () => createDataPath(dataDir, '/experiments'),
       pluginRoute: (
         pluginName: string,

tensorboard/components/tf_backend/test/backendTests.ts

@@ -97,7 +97,7 @@ namespace tf_backend {
         });
 
         it('returns correct value for #environment', () => {
-          assert.equal(router.environment(), 'data/environment');
+          assert.equal(router.environment(), 'data/environment?experiment=');
         });
 
         it('returns correct value for #experiments', () => {

tensorboard/data/provider.py

@@ -33,6 +33,22 @@ class DataProvider(object):
   downsampling strategies or domain restriction by step or wall time.
   """
 
+  def data_location(self, experiment_id):
+    """Render a human-readable description of the data source.
+
+    For instance, this might return a path to a directory on disk.
+
+    The default implementation always returns the empty string.
+
+    Args:
+      experiment_id: ID of enclosing experiment.
+
+    Returns:
+      A string, which may be empty.
+    """
+    return ""
+
+
   @abc.abstractmethod
   def list_runs(self, experiment_id):
     """List all runs within an experiment.

tensorboard/plugins/core/core_plugin.py

@@ -120,10 +120,15 @@ def _serve_environment(self, request):
       database (depending on which mode TensorBoard is running in).
     * window_title is the title of the TensorBoard web page.
     """
+    if self._data_provider:
+      experiment = request.args.get('experiment', '')
+      data_location = self._data_provider.data_location(experiment)
+    else:
+      data_location = self._logdir or self._db_uri
     return http_util.Respond(
         request,
         {
-            'data_location': self._logdir or self._db_uri,
+            'data_location': data_location,
             'window_title': self._window_title,
         },
         'application/json')

tensorboard/plugins/scalar/scalars_plugin_test.py

@@ -124,7 +124,7 @@ def wrapper(self, *args, **kwargs):
           fn(self, scalars_plugin.ScalarsPlugin(ctx), *args, **kwargs)
         with self.subTest('generic data provider'):
           flags = argparse.Namespace(generic_data='true')
-          provider = data_provider.MultiplexerDataProvider(multiplexer)
+          provider = data_provider.MultiplexerDataProvider(multiplexer, logdir)
           ctx = base_plugin.TBContext(
               flags=flags,
               logdir=logdir,