Overhaul swarming_task_cost.py

It was previously named cost.py. It's useful for short snapshots to do quick
analysis. It will eventually be superseeded with internal tooling but it is
still useful in some cases.

Fix two bugs:
- when task with no OS specified, causing the script to crash.
- percentile calculation was broken

[email protected]
BUG=

Review-Url: https://codereview.chromium.org/2909693003

client/tools/cost.py → client/tools/swarming_task_cost.py

@@ -50,8 +50,9 @@ def do_bucket(items, bucket_type):
         gpu_tag = t[4:]
 
     if bucket_type in (MAJOR_OS, MAJOR_OS_ASAN):
-      os_tag = os_tag.split('-')[0]
-    tag = os_tag
+      if os_tag:
+        os_tag = os_tag.split('-')[0]
+    tag = os_tag or ''
     if bucket_type == MINOR_OS_GPU and gpu_tag and gpu_tag != 'none':
       tag += ' gpu:' + gpu_tag
     if bucket_type == MAJOR_OS_ASAN and is_asan:
@@ -126,7 +127,7 @@ def percentile(items, percent):
   rank = percent * .01 * (len(items) + 1)
   rank_int = int(rank)
   rest = rank - rank_int
-  if rest and rank_int <= len(items) - 1:
+  if rest and rank_int < len(items) - 1:
     return items[rank_int] + rest * (items[rank_int+1] - items[rank_int])
   return items[min(rank_int, len(items) - 1)]
 
@@ -139,7 +140,9 @@ def sp(dividend, divisor):
 
 
 def fetch_data(options):
-  """Fetches data from options.swarming and writes it to options.json."""
+  """Fetches TaskResultSummary as JSON from options.swarming and writes it to
+  options.json.
+  """
   if not options.start:
     # Defaults to 25 hours ago.
     options.start = datetime.datetime.utcnow() - datetime.timedelta(
@@ -160,23 +163,21 @@ def fetch_data(options):
     'query',
     '-S', options.swarming,
     '--json', options.json,
-    # Start chocking at 1b tasks. The chromium infrastructure is currently at
-    # around 200k tasks/day.
-    '--limit', '1000000000',
+    '--limit', '0',
     '--progress',
     url,
   ]
   if options.verbose:
-    cmd.append('--verbose')
-    cmd.append('--verbose')
-    cmd.append('--verbose')
+    cmd.extend(('--verbose', '--verbose', '--verbose'))
   logging.info('%s', ' '.join(cmd))
   subprocess.check_call(cmd)
   print('')
 
 
 def stats(tasks, show_cost):
-  """Calculates and prints statistics about the tasks."""
+  """Calculates and prints statistics about the tasks as a list of JSON encoded
+  TaskResultSummary.
+  """
   # Split tasks into 3 buckets.
   # - 'rn' means ran, not idempotent
   # - 'ri' means ran, idempotent
@@ -207,14 +208,14 @@ def stats(tasks, show_cost):
   cost_ri = sum(sum(i.get('costs_usd') or [0.]) for i in ri)
   cost_dd = sum(i.get('cost_saved_usd', 0.) for i in dd)
   cost_total = cost_rn + cost_ri + cost_dd
-  pendings = [
+  pendings = sorted(
     (parse_time(i['started_ts']) - parse_time(i['created_ts'])).total_seconds()
     for i in tasks if i.get('started_ts') and not i.get('deduped_from')
-  ]
-  pending_total = datetime.timedelta(seconds=round(sum(pendings)))
-  pending_avg = datetime.timedelta(seconds=round(average(pendings)))
-  pending_med = datetime.timedelta(seconds=round(median(pendings)))
-  pending_p99 = datetime.timedelta(seconds=round(percentile(pendings, 99)))
+  )
+  pending_total = datetime.timedelta(seconds=round(sum(pendings), 2))
+  pending_avg = datetime.timedelta(seconds=round(average(pendings), 2))
+  pending_med = datetime.timedelta(seconds=round(median(pendings), 2))
+  pending_p99 = datetime.timedelta(seconds=round(percentile(pendings, 99), 2))
 
   # Calculate percentages to understand load relativeness.
   percent_rn_nb_total = sp(len(rn), len(tasks))
@@ -302,7 +303,7 @@ def present_task_types(items, bucket_type, show_cost):
   for index, (bucket, tasks) in enumerate(sorted(buckets.iteritems())):
     if index:
       print('')
-    print('%s:' % (bucket))
+    print('%s:' % (bucket or '<None>'))
     stats(tasks, show_cost)
   if buckets:
     print('')