Tracing and Observability

Seeing what actually happened

Beginner Lesson 8 introduced stream_mode="updates", seeing which node's output changed at each step. That's useful, but it doesn't show timing, or the exact order events occurred in when a graph has both a loop and a branch. stream_mode="debug" fills that in: it emits a "task" event the moment a node starts, and a "task_result" event the moment it finishes, each carrying a timestamp. This lesson reconstructs a full execution trace from those two event types, without LangSmith, without any external service, without anything beyond what every earlier lesson already required.

What a debug event looks like

{'step': 2, 'timestamp': '...', 'type': 'task',
 'payload': {'id': '...', 'name': 'increment', 'input': {'count': 1}, ...}}

{'step': 2, 'timestamp': '...', 'type': 'task_result',
 'payload': {'id': '...', 'name': 'increment', 'result': {'count': 2}, ...}}

Every event carries a step number, a type ("task" for a node starting, "task_result" for it finishing), and a payload with the node's name and, for results, what it returned. The id field is shared between a task and its matching result, that's how you know which task_result belongs to which task when several nodes might be involved.

Matching starts to finishes

started_at: dict[str, datetime] = {}

for event in app.stream({"count": 0}, stream_mode="debug"):
    payload = event["payload"]
    task_id = payload["id"]

    if event["type"] == "task":
        started_at[task_id] = datetime.fromisoformat(event["timestamp"])
        print(f"[step {event['step']}] -> entering {payload['name']} (input={payload['input']})")

    elif event["type"] == "task_result":
        elapsed = datetime.fromisoformat(event["timestamp"]) - started_at[task_id]
        print(f"[step {event['step']}] <- leaving  {payload['name']} (result=..., took {elapsed}ms)")

Record the start time keyed by id when a task begins, compute the elapsed time when its matching result arrives. This works correctly even if a real graph runs multiple nodes concurrently (Intermediate Lesson 17's fan-out), each has its own id, so their timings never get mixed up.

A graph with both a loop and a branch

The lesson's graph increments a counter in a loop (same shape as Beginner Lesson 5) until it reaches 3, then branches to even_path or odd_path depending on whether the final count is even or odd. Running it under stream_mode="debug" shows the loop's three passes through increment, each with its own timing, followed by exactly one branch taken, never both.