Add singleflight logic to the metrics endpoint cache

Prevents the scenario where requests for /metrics start taking longer
than the Prometheus scraping interval, and the whole process would
soon OOM due to the concurrent cache refreshes.

mailman_api/mailman.py

@@ -6,6 +6,7 @@ import ssl
 import subprocess
 import sys
 import time
+import threading
 from functools import wraps
 from flask import Flask, request, abort, make_response, jsonify, g
 
@@ -132,14 +133,24 @@ def regenerate_archives():
 ### Instrumentation.
 
 def cache(timeout):
-    _store = {'deadline': 0, 'value': None}
+    """Decorator for caching expensive results.
+
+    Allows a single thread to recompute the cached value even in
+    presence of concurrent / overlapping requests (will serve stale
+    data in the meantime).
+
+    """
+    _store = {'deadline': 0, 'value': '', 'lock': threading.Lock()}
     def _cache_wrapper(fn):
         @wraps(fn)
         def _cache(*args, **kwargs):
             now = time.time()
-            if now > _store['deadline']:
+            if now > _store['deadline'] and _store['lock'].acquire(False):
+                try:
                     _store['value'] = fn(*args, **kwargs)
                     _store['deadline'] = now + timeout
+                finally:
+                    _store['lock'].release()
             return _store['value']
         return _cache
     return _cache_wrapper