[Telemetry] Change trace_data to hold a list of raw trace data for each trace part

telemetry/telemetry/value/trace.py

Index: telemetry/telemetry/value/trace.py
diff --git a/telemetry/telemetry/value/trace.py b/telemetry/telemetry/value/trace.py
index 8c30c0255cd1773fc02109488daf9c531f220e41..d974ecba938b1799ef4a5fff4256351cfe3ed470 100644
--- a/telemetry/telemetry/value/trace.py
+++ b/telemetry/telemetry/value/trace.py
@@ -46,9 +46,9 @@ class TraceValue(value_module.Value):
       return self._serialized_file_handle.GetAbsPath()
 
   def _GetTraceParts(self, trace_data):
-    return [(trace_data.GetTraceFor(p), p)
+    return [(trace_data.GetTracesFor(p), p)
             for p in trace_data_module.ALL_TRACE_PARTS
-            if trace_data.HasTraceFor(p)]
+            if trace_data.HasTracesFor(p)]
 
   @staticmethod
   def _DumpTraceToFile(trace, path):
@@ -66,12 +66,14 @@ class TraceValue(value_module.Value):
     counter = 0
     try:
       trace_size_data = {}
-      for trace, part in self._GetTraceParts(trace_data):
-        file_path = os.path.join(temp_dir, '%s.trace' % counter)
-        self._DumpTraceToFile(trace, file_path)
-        trace_size_data[part] = os.path.getsize(file_path)
-        trace_files.append(file_path)
-        counter += 1
+      for traces_list, part in self._GetTraceParts(trace_data):

[charliea (OOO until 10-5), 2017/01/12 21:10:17]

High-level question: do you know why it's necessary to combine these? IIUC each
of these is a valid stand-alone trace. Can't we just write each one as a
separate trace, hand that html file over to tracing, and let it import each one
separately and have things still work correctly?

[nednguyen, 2017/01/12 21:29:16]

We have a thread a long ago with a similar question by memory team but I lose
the pointer to it. The basic idea here is Telemetry's architecture is:
1 run ---> 1 set of metrics computed for that run.

So for a multi-browsers test, to make this model consistent, we combine the
traces to avoid: 1 run ---> N set of metrics computed for that run. If we do
this, we need to expand our metric & result system to include the concept of
"trace", which seems to break the layering between testing & metrics.

Besides those conceptual problem of computing metrics on a per traces basic,
another reason why we want to combine data is that because multi-browsers do
share memory (which makes them interesting test cases). By not combining the
tracing data, we lose the view to the total system when trying to compute
metrics.

[charliea (OOO until 10-5), 2017/01/13 16:36:47]

Ned and I talked about this offline, and I was misunderstanding what was
happening. I thought that we were concatenating the traces so that, for two
browsers X and Y, we were outputting something like the following in HTML:

<script>
trace for X
trace for Y
</script>

instead, what we're doing is outputting two <script> tags, like:

<script>
trace for X
</script>
<script>
trace for Y
</script>

which looks great to me.

+        for trace in traces_list:
+          file_path = os.path.join(temp_dir, '%s.trace' % counter)
+          self._DumpTraceToFile(trace, file_path)
+          trace_size_data.setdefault(part, 0)
+          trace_size_data[part] += os.path.getsize(file_path)
+          trace_files.append(file_path)
+          counter += 1
       logging.info('Trace sizes in bytes: %s', trace_size_data)
       tf = tempfile.NamedTemporaryFile(delete=False, suffix='.html')
       tf.close()