gae_ts_mon: instrument Cloud Endpoint methods

appengine_module/gae_ts_mon/config.py

 # Copyright 2015 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 copy
 import datetime
+import functools
 import logging
 import os
 import sys
 import time
 import threading
 
+import endpoints
 import webapp2
 
 from google.appengine.api import modules
 from google.appengine.api.app_identity import app_identity
 from google.appengine.api import runtime
 from google.appengine.ext import ndb
 
 from infra_libs.ts_mon import handlers
 from infra_libs.ts_mon import shared
 from infra_libs.ts_mon.common import http_metrics
 from infra_libs.ts_mon.common import interface
 from infra_libs.ts_mon.common import metric_store
 from infra_libs.ts_mon.common import monitors
 from infra_libs.ts_mon.common import targets
 
 
 def _reset_cumulative_metrics():
   """Clear the state when an instance loses its task_num assignment."""
   logging.warning('Instance %s got purged from Datastore, but is still alive. '
                   'Clearing cumulative metrics.', shared.instance_key_id())
   for _target, metric, start_time, _fields in interface.state.store.get_all():
     if metric.is_cumulative():
       metric.reset()
 
 
 _flush_metrics_lock = threading.Lock()
 
 
-def flush_metrics_if_needed(time_fn=datetime.datetime.utcnow):
-  time_now = time_fn()
+def need_to_flush_metrics(time_fn=time.time):

[nodir, 2016/03/18 17:38:02]

I'd add a docstring that this function is not pure as the func name my falsly
imply

[Sergey Berezin, 2016/03/21 18:49:32]

Done.

+  time_now = datetime.datetime.utcfromtimestamp(time_fn())
   minute_ago = time_now - datetime.timedelta(seconds=60)
   with _flush_metrics_lock:
     if interface.state.last_flushed > minute_ago:
       return False
     interface.state.last_flushed = time_now

[nodir, 2016/03/18 17:38:02]

What happens if flushing fails after this value is changed?

[Sergey Berezin, 2016/03/21 18:49:32]

We simply retry a minute later. This in fact is a feature, as otherwise we'd be
needlessly affecting every user request e.g. during brief PubSub outages.

+  return True
 
-  return _flush_metrics(time_now)
+
+def flush_metrics_if_needed(time_fn=time.time):
+  if not need_to_flush_metrics(time_fn=time_fn):
+    return False
+  return _flush_metrics(interface.state.last_flushed)

[nodir, 2016/03/18 17:38:02]

passing last_flushed as time_now seems to heavily rely on the fact that
need_to_flush_metrics assigns it to time_fn(), which is an abstraction leak. I'd
save time_fn() to a variable, pass it to need_to_flush_metrics and
_flush_metrics

[Sergey Berezin, 2016/03/21 18:49:32]

Very good point, thanks. I refactored this part of the code.

 
 
 def _flush_metrics(time_now):
   """Return True if metrics were actually sent."""
   entity = shared.get_instance_entity()
   if entity.task_num < 0:
     if interface.state.target.task_num >= 0:
       _reset_cumulative_metrics()
     interface.state.target.task_num = -1
     interface.state.last_flushed = entity.last_updated
@@ skipping 104 matching lines @@
       shared.INTERNAL_CALLBACK_NAME, _internal_callback)
 
   logging.info('Initialized ts_mon with service_name=%s, job_name=%s, '
                'hostname=%s', service_name, job_name, hostname)
 
 
 def _instrumented_dispatcher(dispatcher, request, response, time_fn=time.time):
   start_time = time_fn()
   response_status = 0
   interface.state.store.initialize_context()
-  flush_thread = threading.Thread(target=flush_metrics_if_needed)
-  flush_thread.start()
+  flush_thread = None
+  if need_to_flush_metrics(time_fn=time_fn):
+    flush_thread = threading.Thread(
+        target=_flush_metrics, args=(interface.state.last_flushed,))
+    flush_thread.start()
   try:
     ret = dispatcher(request, response)
   except webapp2.HTTPException as ex:
     response_status = ex.code
     raise
   except Exception:
     response_status = 500
     raise
   else:
     if isinstance(ret, webapp2.Response):
       response = ret
     response_status = response.status_int
   finally:
-    flush_thread.join()
+    if flush_thread:
+      flush_thread.join()

[nodir, 2016/03/18 17:38:02]

please join after elapsed_ms is computed

[Sergey Berezin, 2016/03/21 18:49:32]

This is on purpose: I want to track the total request time including the
monitoring overhead. Otherwise we may miss ts_mon regressions.

     elapsed_ms = int((time_fn() - start_time) * 1000)
 
-    fields = {'status': response_status, 'name': '', 'is_robot': False}
-    if request.route is not None:
-      # Use the route template regex, not the request path, to prevent an
-      # explosion in possible field values.
-      fields['name'] = request.route.template
-    if request.user_agent is not None:
-      # We must not log user agents, but we can store whether or not the
-      # user agent string indicates that the requester was a Google bot.
-      fields['is_robot'] = (
-          'GoogleBot' in request.user_agent or
-          'GoogleSecurityScanner' in request.user_agent)
+    # Use the route template regex, not the request path, to prevent an
+    # explosion in possible field values.
+    name = request.route.template if request.route is not None else ''
 
-    http_metrics.server_durations.add(elapsed_ms, fields=fields)
-    http_metrics.server_response_status.increment(fields=fields)
-    if request.content_length is not None:
-      http_metrics.server_request_bytes.add(request.content_length,
-                                            fields=fields)
-    if response.content_length is not None:  # pragma: no cover
-      http_metrics.server_response_bytes.add(response.content_length,
-                                             fields=fields)
+    http_metrics.update_http_server_metrics(
+        name, response_status, elapsed_ms,
+        request_size=request.content_length,
+        response_size=response.content_length,
+        user_agent=request.user_agent)
+
   return ret
 
 
 def instrument_wsgi_application(app, time_fn=time.time):
   # Don't instrument the same router twice.
   if hasattr(app.router, '__instrumented_by_ts_mon'):
     return
 
   old_dispatcher = app.router.dispatch
 
   def dispatch(router, request, response):
     return _instrumented_dispatcher(old_dispatcher, request, response,
                                     time_fn=time_fn)
 
   app.router.set_dispatcher(dispatch)
   app.router.__instrumented_by_ts_mon = True
 
 
+def instrument_endpoint(time_fn=time.time):
+  """Decorator to instrument Cloud Endpoint methods."""
+  def decorator(fn):
+    method_name = fn.__name__
+    assert method_name
+    @functools.wraps(fn)
+    def decorated(service, *args, **kwargs):
+      service_name = service.__class__.__name__
+      endpoint_name = '/_ah/spi/%s.%s' % (service_name, method_name)
+      start_time = time_fn()
+      response_status = 0
+      interface.state.store.initialize_context()
+      flush_thread = None
+      if need_to_flush_metrics(time_fn=time_fn):
+        flush_thread = threading.Thread(
+            target=_flush_metrics, args=(interface.state.last_flushed,))
+        flush_thread.start()
+      try:
+        ret = fn(service, *args, **kwargs)
+        response_status = 200
+        return ret
+      except endpoints.ServiceException as e:
+        response_status = e.http_status
+        raise
+      except Exception:
+        response_status = 500
+        raise
+      finally:
+        if flush_thread:
+          flush_thread.join()

[nodir, 2016/03/18 17:38:02]

same here

[Sergey Berezin, 2016/03/21 18:49:32]

Same response as above.

+        elapsed_ms = int((time_fn() - start_time) * 1000)
+        http_metrics.update_http_server_metrics(
+            endpoint_name, response_status, elapsed_ms)
+    return decorated
+  return decorator
+
+
 def reset_for_unittest():
   shared.reset_for_unittest()
   interface.reset_for_unittest()