Support periodic pruning of the most expensive worker process.

examples/example_config.py

@@ -71,6 +71,39 @@
 timeout = 30
 keepalive = 2
 
+#
+#   prune_function
+#       A function that is passed a process ID of a worker and returns a
+#       score (such as total memory used).  Once every prune seconds, the
+#       worker with the highest score is killed (unless the score is below
+#       the prune floor).
+#
+#   prune_seconds
+#       How many seconds to wait between killing the worker with the highest
+#       score from the prune function.  If set to 0 (the default), then no
+#       pruning is done.  The actual time waited is a random value between
+#       90% and 100% of this value.
+#
+#   prune_floor
+#       When the score from the prune function is at or below this value, the
+#       worker will not be killed even if it has the highest score.
+#
+
+import psutil
+
+def proc_vmsize(pid):
+    # Return how many MB of virtual memory is being used by a worker process
+    try:
+        p = psutil.Process(pid)
+        mb = p.memory_info().vms/1024/1024
+        return mb
+    except psutil.NoSuchProcessError:
+        return 0
+
+prune_seconds = 5*60            # Prune largest worker every 4.5-5m
+prune_function = proc_vmsize    # Measure worker size in MB of VM
+prune_floor = 300               # Don't kill workers using <= 300 MB of VM
+
 #
 #   spew - Install a trace function that spews every line of Python
 #       that is executed when running the server. This is the

gunicorn/arbiter.py

@@ -63,6 +63,7 @@ def __init__(self, app):
         self.reexec_pid = 0
         self.master_pid = 0
         self.master_name = "Master"
+        self.next_prune = None
 
         cwd = util.getcwd()
 
@@ -203,6 +204,13 @@ def run(self):
 
             while True:
                 self.maybe_promote_master()
+                if 0 < self.cfg.prune_seconds:
+                    if self.next_prune is None:
+                        self.next_prune = time.monotonic() + self.cfg.prune_seconds
+                    elif self.next_prune <= time.monotonic():
+                        self.prune_worker()
+                        self.next_prune +=
+                            self.cfg.prune_seconds*(0.9 + 0.1*random.random())
 
                 sig = self.SIG_QUEUE.pop(0) if self.SIG_QUEUE else None
                 if sig is None:
@@ -486,6 +494,22 @@ def reload(self):
         # manage workers
         self.manage_workers()
 
+    def prune_worker(self):
+        """\
+        Kill the worker with highest prune score
+        """
+        workers = list(self.WORKERS.items())
+        max = self.cfg.prune_floor
+        victim = 0
+        for (pid, worker) in workers:
+            score = self.cfg.prune_function(pid)
+            if max < score:
+                max = score
+                victim = pid
+        if victim != 0:
+            self.log.info(f"Pruning worker (pid: {victim}) with score {score}")
+            self.kill_worker(pid, signal.SIGTERM)
+
     def murder_workers(self):
         """\
         Kill unused/idle workers
@@ -579,16 +603,18 @@ def manage_workers(self):
         active_worker_count = len(workers)
         if self._last_logged_active_worker_count != active_worker_count:
             self._last_logged_active_worker_count = active_worker_count
-            self.log.debug("{0} workers".format(active_worker_count),
-                           extra={"metric": "gunicorn.workers",
-                                  "value": active_worker_count,
-                                  "mtype": "gauge"})
+            self.log.debug(
+                "{0} workers".format(active_worker_count),
+                extra={
+                    "metric": "gunicorn.workers",
+                    "value": active_worker_count,
+                    "mtype": "gauge"})
 
     def spawn_worker(self):
         self.worker_age += 1
-        worker = self.worker_class(self.worker_age, self.pid, self.LISTENERS,
-                                   self.app, self.timeout / 2.0,
-                                   self.cfg, self.log)
+        worker = self.worker_class(
+            self.worker_age, self.pid, self.LISTENERS, self.app,
+            self.timeout / 2.0, self.cfg, self.log)
         self.cfg.pre_fork(self, worker)
         pid = os.fork()
         if pid != 0:

gunicorn/config.py

@@ -724,6 +724,54 @@ class WorkerConnections(Setting):
         """
 
 
+class PruneFunction(Setting):
+    name = "prune_function"
+    section = "Worker Processes"
+    cli = ["--prune-function"]
+    validator = validate_callable(1)
+    type = callable
+
+    def prune_score(pid):
+        return 0
+    default = staticmethod(prune_score)
+    desc = """\
+        A function that is passed a process ID of a worker and returns a
+        score (such as total memory used).  Once every prune seconds, the
+        worker with the highest score is killed (unless the score is below
+        the prune floor).
+        """
+
+
+class PruneSeconds(Setting):
+    name = "prune_seconds"
+    section = "Worker Processes"
+    cli = ["--prune-seconds"]
+    meta = "INT"
+    validator = validate_pos_int
+    type = int
+    default = 0
+    desc = """\
+        How many seconds to wait between killing the worker with the highest
+        score from the prune function.  If set to 0 (the default), then no
+        pruning is done.  The actual time waited is a random value between
+        90% and 100% of this value.
+        """
+
+
+class PruneFloor(Setting):
+    name = "prune_floor"
+    section = "Worker Processes"
+    cli = ["--prune-floor"]
+    meta = "INT"
+    validator = validate_pos_int
+    type = int
+    default = 0
+    desc = """\
+        When the score from the prune function is at or below this value, the
+        worker will not be killed even if it has the highest score.
+        """
+
+
 class MaxRequests(Setting):
     name = "max_requests"
     section = "Worker Processes"