Retry flush metrics from ThreadStats to Datadog over RemoteDisconnected errors.

Author: tianchu

datadog_lambda/metric.py

@@ -91,6 +91,42 @@ def write_metric_point_to_stdout(metric_name, value, timestamp=None, tags=[]):
     )
 
 
+def flush_thread_stats():
+    """"Flush distributions from ThreadStats to Datadog.
+
+    Modified based on `datadog.threadstats.base.ThreadStats.flush()`,
+    to gain better control over exception handling.
+    """
+    _, dists = lambda_stats._get_aggregate_metrics_and_dists(float("inf"))
+    count_dists = len(dists)
+    if not count_dists:
+        logger.debug("No distributions to flush. Continuing.")
+
+    lambda_stats.flush_count += 1
+    logger.debug(
+        "Flush #%s sending %s distributions", lambda_stats.flush_count, count_dists
+    )
+    try:
+        lambda_stats.reporter.flush_distributions(dists)
+    except Exception as e:
+        # The nature of the root issue https://bugs.python.org/issue41345 is to be complex,
+        # but comprehensive tests suggest that it is safe to retry on this specific error.
+        if isinstance(e, api.exceptions.ClientError) and "RemoteDisconnected" in str(e):
+            logger.debug(
+                "Retry flush #%s due to RemoteDisconnected", lambda_stats.flush_count
+            )
+            try:
+                lambda_stats.reporter.flush_distributions(dists)
+            except Exception:
+                logger.debug(
+                    "Flush #%s failed after retry",
+                    lambda_stats.flush_count,
+                    exc_info=True,
+                )
+        else:
+            logger.debug("Flush #%s failed", lambda_stats.flush_count, exc_info=True)
+
+
 def are_enhanced_metrics_enabled():
     """Check env var to find if enhanced metrics should be submitted
 
@@ -165,3 +201,6 @@ def submit_errors_metric(lambda_context):
     "DATADOG_HOST", "https://api." + os.environ.get("DD_SITE", "datadoghq.com")
 )
 logger.debug("Setting DATADOG_HOST to %s", api._api_host)
+
+# Unmute exceptions from datadog api client, so we can catch and handle them
+api._mute = False

datadog_lambda/wrapper.py

@@ -12,7 +12,7 @@
 from datadog_lambda.cold_start import set_cold_start, is_cold_start
 from datadog_lambda.constants import XraySubsegment, TraceContextSource
 from datadog_lambda.metric import (
-    lambda_stats,
+    flush_thread_stats,
     submit_invocations_metric,
     submit_errors_metric,
 )
@@ -177,7 +177,7 @@ def _after(self, event, context):
                 )
 
             if not self.flush_to_log or should_use_extension:
-                lambda_stats.flush(float("inf"))
+                flush_thread_stats()
             if should_use_extension:
                 flush_extension()
 

tests/test_metric.py

@@ -6,7 +6,8 @@
 except ImportError:
     from mock import patch, call
 
-from datadog_lambda.metric import lambda_metric
+from datadog.api.exceptions import ClientError
+from datadog_lambda.metric import lambda_metric, flush_thread_stats
 from datadog_lambda.tags import _format_dd_lambda_layer_tag
 
 
@@ -36,3 +37,22 @@ def test_lambda_metric_flush_to_log(self):
         self.mock_metric_lambda_stats.distribution.assert_not_called()
 
         del os.environ["DD_FLUSH_TO_LOG"]
+
+
+class TestFlushThreadStats(unittest.TestCase):
+    def setUp(self):
+        patcher = patch(
+            "datadog.threadstats.reporters.HttpReporter.flush_distributions"
+        )
+        self.mock_threadstats_flush_distributions = patcher.start()
+        self.addCleanup(patcher.stop)
+
+    def test_retry_on_remote_disconnected(self):
+        # Raise the RemoteDisconnected error
+        self.mock_threadstats_flush_distributions.side_effect = ClientError(
+            "POST",
+            "https://api.datadoghq.com/api/v1/distribution_points",
+            "RemoteDisconnected('Remote end closed connection without response')"
+        )
+        flush_thread_stats()
+        self.assertEqual(self.mock_threadstats_flush_distributions.call_count, 2)

tests/test_wrapper.py

@@ -38,14 +38,6 @@ def setUp(self):
         self.mock_threadstats_flush_distributions = patcher.start()
         self.addCleanup(patcher.stop)
 
-        self.metric_lambda_stats_patcher = patch("datadog_lambda.metric.lambda_stats")
-        self.mock_metric_lambda_stats = self.metric_lambda_stats_patcher.start()
-        self.addCleanup(self.metric_lambda_stats_patcher.stop)
-
-        self.wrapper_lambda_stats_patcher = patch("datadog_lambda.wrapper.lambda_stats")
-        self.mock_wrapper_lambda_stats = self.wrapper_lambda_stats_patcher.start()
-        self.addCleanup(self.wrapper_lambda_stats_patcher.stop)
-
         patcher = patch("datadog_lambda.wrapper.extract_dd_trace_context")
         self.mock_extract_dd_trace_context = patcher.start()
         self.mock_extract_dd_trace_context.return_value = ({}, None)
@@ -101,10 +93,16 @@ def lambda_handler(event, context):
 
         lambda_handler(lambda_event, lambda_context)
 
-        self.mock_metric_lambda_stats.distribution.assert_has_calls(
-            [call("test.metric", 100, timestamp=None, tags=ANY)]
+        self.mock_threadstats_flush_distributions.assert_has_calls(
+            [call([{
+                'metric': 'test.metric', 'points': [[ANY, [100]]],
+                'type': 'distribution',
+                'host': None,
+                'device': None,
+                'tags': ANY,
+                'interval': 10
+            }])]
         )
-        self.mock_wrapper_lambda_stats.flush.assert_called()
         self.mock_extract_dd_trace_context.assert_called_with(
             lambda_event, lambda_context, extractor=None
         )
@@ -122,16 +120,11 @@ def lambda_handler(event, context):
         lambda_event = {}
         lambda_handler(lambda_event, get_mock_context())
 
-        self.mock_metric_lambda_stats.distribution.assert_not_called()
-        self.mock_wrapper_lambda_stats.flush.assert_not_called()
+        self.mock_threadstats_flush_distributions.assert_not_called()
 
         del os.environ["DD_FLUSH_TO_LOG"]
 
     def test_datadog_lambda_wrapper_flush_in_thread(self):
-        # stop patchers so mock_threadstats_flush_distributions gets called
-        self.metric_lambda_stats_patcher.stop()
-        self.wrapper_lambda_stats_patcher.stop()
-
         # force ThreadStats to flush in thread
         import datadog_lambda.metric as metric_module
 
@@ -158,15 +151,7 @@ def lambda_handler(event, context):
         metric_module.lambda_stats.stop()
         metric_module.lambda_stats.start(flush_in_thread=False)
 
-        # resume patchers
-        self.metric_lambda_stats_patcher.start()
-        self.wrapper_lambda_stats_patcher.start()
-
     def test_datadog_lambda_wrapper_not_flush_in_thread(self):
-        # stop patchers so mock_threadstats_flush_distributions gets called
-        self.metric_lambda_stats_patcher.stop()
-        self.wrapper_lambda_stats_patcher.stop()
-
         # force ThreadStats to not flush in thread
         import datadog_lambda.metric as metric_module
 
@@ -193,10 +178,6 @@ def lambda_handler(event, context):
         metric_module.lambda_stats.stop()
         metric_module.lambda_stats.start(flush_in_thread=False)
 
-        # resume patchers
-        self.metric_lambda_stats_patcher.start()
-        self.wrapper_lambda_stats_patcher.start()
-
     def test_datadog_lambda_wrapper_inject_correlation_ids(self):
         os.environ["DD_LOGS_INJECTION"] = "True"
 
@@ -445,5 +426,4 @@ def lambda_handler(event, context):
         lambda_handler_double_wrapped(lambda_event, get_mock_context())
 
         self.mock_patch_all.assert_called_once()
-        self.mock_wrapper_lambda_stats.flush.assert_called_once()
         self.mock_submit_invocations_metric.assert_called_once()