Change jetstream_slots_available_percentage to `jetstream_slots_use…

…d_percentage` (#102)

* initial_commit

* pylint

* updated example

docs/observability-prometheus-metrics-in-jetstream-server.md

@@ -45,7 +45,7 @@ Now that we configured `prometheus_port=9090` above, we can observe various Jets
 # HELP jetstream_prefill_backlog_size Size of prefill queue
 # TYPE jetstream_prefill_backlog_size gauge
 jetstream_prefill_backlog_size{id="SOME-HOSTNAME-HERE>"} 0.0
-# HELP jetstream_slots_available_percentage The percentage of available slots in decode batch
-# TYPE jetstream_slots_available_percentage gauge
-jetstream_slots_available_percentage{id="<SOME-HOSTNAME-HERE>",idx="0"} 0.96875
+# HELP jetstream_slots_used_percentage The percentage of decode slots currently being used
+# TYPE jetstream_slots_used_percentage gauge
+jetstream_slots_used_percentage{id="<SOME-HOSTNAME-HERE>",idx="0"} 0.04166666666666663
 ```

jetstream/core/metrics/prometheus.py

@@ -35,14 +35,14 @@ def __new__(cls):
       documentation="Size of prefill queue",
       labelnames=["id"],
   )
-  _slots_available_percentage = Gauge(
-      name="jetstream_slots_available_percentage",
-      documentation="The percentage of available slots in decode batch",
+  _slots_used_percentage = Gauge(
+      name="jetstream_slots_used_percentage",
+      documentation="The percentage of decode slots currently being used",
       labelnames=["id", "idx"],
   )
 
   def get_prefill_backlog_metric(self):
     return self._prefill_backlog.labels(id=self._id)
 
-  def get_slots_available_percentage_metric(self, idx: int):
-    return self._slots_available_percentage.labels(id=self._id, idx=idx)
+  def get_slots_used_percentage_metric(self, idx: int):
+    return self._slots_used_percentage.labels(id=self._id, idx=idx)

jetstream/core/orchestrator.py

@@ -597,9 +597,11 @@ def _generate_thread(self, idx: int):
       max_concurrent_decodes = generate_engine.max_concurrent_decodes
 
       if self._metrics_collector:
-        self._metrics_collector.get_slots_available_percentage_metric(
+        self._metrics_collector.get_slots_used_percentage_metric(
             idx
-        ).set_function(lambda: float(my_slots.qsize() / max_concurrent_decodes))
+        ).set_function(
+            lambda: float(1 - (my_slots.qsize() / max_concurrent_decodes))
+        )
 
       # Check if there are any free my_slots. We don't want to block here since
       # we can still generate if we can't insert. We do this in a while loop to