[Telemetry] Fix a possible source of flakiness on Android.

tools/telemetry/telemetry/multi_page_benchmark_runner.py

 #!/usr/bin/env python
 # Copyright (c) 2012 The Chromium Authors. All rights reserved.
 # Use of this source code is governed by a BSD-style license that can be
 # found in the LICENSE file.
 import csv
 import logging
 import os
 import sys
 
 from telemetry import all_page_interactions # pylint: disable=W0611
+from telemetry import connection_gone_exception
 from telemetry import browser_finder
 from telemetry import browser_options
 from telemetry import discover
 from telemetry import multi_page_benchmark
 from telemetry import page_runner
 from telemetry import page_set
 
 def Main(benchmark_dir):
   """Turns a MultiPageBenchmark into a command-line program.
 
@@ skipping 38 matching lines @@
 
   ps = page_set.PageSet.FromFile(args[1])
 
   benchmark.CustomizeBrowserOptions(options)
   possible_browser = browser_finder.FindBrowser(options)
   if not possible_browser:
     print >> sys.stderr, """No browser found.\n
 Use --browser=list to figure out which are available.\n"""
     sys.exit(1)
 
-  if options.output_format == 'csv':
-    results = multi_page_benchmark.CsvBenchmarkResults(csv.writer(sys.stdout))
-  elif options.output_format == 'terminal-block':
-    results = multi_page_benchmark.TerminalBlockBenchmarkResults(sys.stdout)
-  else:
-    raise Exception('Invalid --output-format value: "%s". Valid values are '
-                    '"csv" and "terminal-block".'
-                    % options.output_format)
-
-  with page_runner.PageRunner(ps) as runner:
-    runner.Run(options, possible_browser, benchmark, results)
-  # When using an exact executable, assume it is a reference build for the
-  # purpose of outputting the perf results.
-  results.PrintSummary(options.browser_executable and '_ref' or '')
+  retries = 3
+  while retries:

[nduca, 2012/12/04 04:07:47]

This shouldn't be done. We really need to robustly handle this somehow inside
telemetry. The point of page_runner is to do the retry logic again. Perhaps you
should raise BrowserGone. That'll cause us to recreate the browser. Then,
page_runner could track the number of tries and fail catastrophically if that
happens?

+    try:
+      if options.output_format == 'csv':
+        results = multi_page_benchmark.CsvBenchmarkResults(
+            csv.writer(sys.stdout))
+      elif options.output_format == 'terminal-block':
+        results = multi_page_benchmark.TerminalBlockBenchmarkResults(sys.stdout)
+      else:
+        raise Exception('Invalid --output-format value: "%s". Valid values are '
+                        '"csv" and "terminal-block".'
+                        % options.output_format)
+      with page_runner.PageRunner(ps) as runner:
+        runner.Run(options, possible_browser, benchmark, results)
+      # When using an exact executable, assume it is a reference build for the
+      # purpose of outputting the perf results.
+      results.PrintSummary(options.browser_executable and '_ref' or '')
+      retries = 0
+    except connection_gone_exception.ConnectionGoneException:
+      logging.warning('Lost connection to browser. Retrying.')
+      retries -= 1
 
   if len(results.page_failures):
     logging.warning('Failed pages: %s', '\n'.join(
         [failure['page'].url for failure in results.page_failures]))
 
   if len(results.skipped_pages):
     logging.warning('Skipped pages: %s', '\n'.join(
         [skipped['page'].url for skipped in results.skipped_pages]))
   return min(255, len(results.page_failures))