Count server items per run instead of per process

app.py

@@ -9,6 +9,7 @@
 # After the fix: every job completes.
 
 import inspect
+import threading
 import time
 
 import gradio as gr
@@ -18,12 +19,14 @@ from gradio_client import client as client_module
 
 HAS_HELPER_POOL = "helper_executor" in inspect.getsource(client_module.Client.__init__)
 
+items_lock = threading.Lock()
 items_handled = 0
 
 
 def echo(text: list[str]) -> list[list[str]]:
     global items_handled
-    items_handled += len(text)
+    with items_lock:
+        items_handled += len(text)
     return [text]
 
 
@@ -35,27 +38,39 @@ def self_url() -> str:
 def run_load_test(count, max_workers):
     count = int(count)
     max_workers = int(max_workers)
+    with items_lock:
+        # Counted per run: the global keeps accumulating for as long as the
+        # Space's process lives, and a stalled run leaves items behind.
+        baseline = items_handled
     client = Client(self_url(), verbose=False, max_workers=max_workers)
 
     started = time.time()
     jobs = [client.submit(str(i), api_name="/echo") for i in range(count)]
 
     done = 0
+    handled = 0
     while time.time() - started < 90:
         done = sum(1 for job in jobs if job.done())
+        with items_lock:
+            handled = items_handled - baseline
         elapsed = time.time() - started
         if done == count:
-            yield f"FINISHED: {count}/{count} jobs completed in {elapsed:.0f}s"
+            yield (
+                f"FINISHED: {count}/{count} jobs completed in {elapsed:.0f}s\n"
+                f"server handled {handled} items for this run"
+            )
             return
         yield (
             f"{done}/{count} jobs completed after {elapsed:.0f}s\n"
-            f"server has handled {items_handled} items"
+            f"server handled {handled} items for this run"
         )
         time.sleep(1)
 
     yield (
         f"STALLED: {done}/{count} jobs completed and then nothing for the rest of "
-        f"the 90s window, while the server handled {items_handled} items.\n"
+        f"the 90s window.\n"
+        f"The server handled {handled} items for this run: the {done} completed "
+        f"jobs, plus the ones still stuck in flight.\n"
         f"The client's worker threads are wedged for good now, so restart the "
         f"Space before running this again."
     )
@@ -70,7 +85,11 @@ with gr.Blocks() as demo:
         "Submitting many jobs in a tight loop through `Client.submit()` used to "
         "wedge the client permanently: jobs stopped completing well before the "
         "last one, and `job.done()` never became True for the rest even though "
-        "the server had already done the work."
+        "the server had already done the work.\n\n"
+        "That is why the item count runs ahead of the completion count during a "
+        "stalled run. The client has POSTed the jobs it is holding, so the server "
+        "handles the completed ones plus up to `max_workers` more whose results "
+        "the client can no longer read."
     )
     with gr.Row():
         with gr.Column():