README.md

Blocking Event Loop Example

This example demonstrates how blocking the asyncio thread can prevent the worker from processing tasks.

As the official doc from the SKD states:

⚠️ WARNING: Do not block the thread in async def Python functions. This can stop the processing of the rest of the Temporal.

more information

Install dependencies

uv python install 3.13
uv sync
uv sync --group encryption

running the example

start grafana and prometheus

./start-grafana_prometheus.sh

• Grafana: http://localhost:3000
• Prometheus: http://localhost:9090

run the worker

uv run _worker.py

start workflows

uv run _starter.py

Blocking the event loop

The file codec.py in the function encode blocks "randomly" the asyncio thread.

The blocking is done as part of the codec, which is used to encrypt/decrypt the payloads.

  def sleep():
     if random.random() < 0.1:
     time.sleep(0.4)
  
  sleep()

This prevents the worker from processing task to the point that they time out (default timeout=10)

This is visible in the worker logs

2025-04-30T14:18:45.415499Z  WARN temporal_sdk_core::worker::workflow: Task not found when completing error=status: NotFound, message: "Workflow task not found.",...

and grafana dashboard (Workflow Task Execution Latency (p95))

The task is dispatched to the worker, but it cannot be run while the event loop is blocked.

To change the code to run with asyncio you can search for # Simulate blocking work and replace

        #asyncio.get_running_loop().run_in_executor(None, sleep)
        sleep()

for

        asyncio.get_running_loop().run_in_executor(None, sleep)
        #sleep()

then restart/rerun everything to review the performance

After changing the code to run with asyncio the performance is much better, in the graph above the change was done at 14:23:

• Workflow execution latency decreased from 3 minutes to 8 seconds (p95)
• Workflow task latency decreased from 1 minute to 7 seconds (p95)
  • It’s still relatively high due to latency added by the codec, but it’s no longer blocking the worker in the same way.
• Workflow schedule to start latency decreased from 1 minute to a few milliseconds (p95)

Changing the code to run purely with asyncio

        #asyncio.get_running_loop().run_in_executor(None, sleep)
        await asyncio.sleep(0.4)
        #sleep()

makes the real difference, but it is not always possible when we use external dependencies.