works

05_scheduling/custom_retries.py

@@ -0,0 +1,95 @@
+# ---
+# cmd: ["modal", "run", "05_scheduling.custom_retries"]
+# ---
+
+# # Custom retries by exception type
+# There are two types of retries in Modal:
+# 1. When a function execution is interrupted by [preemption](https://modal.com/docs/guide/preemption#preemption), the input will be retried. This behavior is not configurable at this time.
+# 2. When a function execution fails, ie by a raised Exception, Modal will retry the function call if you have [`modal.Retries`](https://modal.com/docs/reference/modal.Retries) configured.
+# This example is about customizing the latter to only retry on certain exception types.
+# For example, you may only want to retry on certain expected errors (e.g. timeouts, or
+# transient network errors) and crash immediately on others (e.g. OOM, bad input).
+
+# The trick is to:
+# 1. Raise retryable errors in the usual way to trigger `modal.Retries`
+# 2. Catch and `return` non-retryable errors.
+# For #2, Modal will see a successful function call execution and return the exception
+# to your client/server to handle as desired.
+
+import modal
+
+app = modal.App("example-custom-retries")
+
+# ## Define retryable vs. crashable exceptions
+
+retry_exceptions = (
+    TimeoutError,
+    ConnectionError,
+    # transient CUDA errors, network blips, etc.
+)
+
+crashable_exceptions = (
+    MemoryError,
+    ValueError,
+    # OOM, bad input — retrying won't help
+)
+
+# ## Use a Dict to track call count across retries
+#
+# Each retry runs in a new container invocation, so we use a
+# [`modal.Dict`](https://modal.com/docs/reference/modal.Dict) to share
+# state and make the demo deterministic.
+
+call_counter = modal.Dict.from_name(
+    "custom-retries-demo-counter", create_if_missing=True
+)
+
+# ## Demo App
+#
+# This function follows a scripted sequence to demonstrate the behavior:
+#
+# 1. **Call 1** — raises `TimeoutError` (retryable → Modal retries)
+# 2. **Call 2** — raises `ConnectionError` (retryable → Modal retries)
+# 3. **Call 3** — raises `MemoryError` (crashable → returned, no more retries)
+#
+# So you'll see two retries, then a clean stop on the third attempt.
+
+
+@app.function(retries=modal.Retries(max_retries=5, initial_delay=1.0))
+def flaky_task():
+    call_count = call_counter.get("calls", 0) + 1
+    call_counter["calls"] = call_count
+    print(f"Attempt {call_count}")
+
+    # Scripted error sequence
+    errors = [
+        TimeoutError("GPU timed out"),  # attempt 1: retryable
+        ConnectionError("lost connection to data server"),  # attempt 2: retryable
+        MemoryError("CUDA out of memory"),  # attempt 3: crashable
+    ]
+    error = errors[min(call_count, len(errors)) - 1]
+
+    print(f"  Hit: {error!r}")
+
+    if isinstance(error, retry_exceptions):
+        print("  -> retryable, re-raising so Modal retries")
+        raise error
+
+    # Return instead of raise — Modal sees success, stops retrying
+    print("  -> non-retryable, returning error to stop retries")
+    return error
+
+
+# ## Entrypoint
+#
+# The caller checks whether the return value is an exception.
+
+
+@app.local_entrypoint()
+def main():
+    call_counter["calls"] = 0  # reset counter
+    result = flaky_task.remote()
+    if isinstance(result, Exception):
+        print(f"Stopped with non-retryable error: {result!r}")
+    else:
+        print(f"Result: {result}")