worker-specific-task-queues

Worker-Specific Task Queues

Use a unique Task Queue for each Worker in order to have certain Activities run on a specific Worker.

This is useful in scenarios where multiple Activities need to run in the same process or on the same host, for example to share memory or disk. This sample has a file processing Workflow, where one Activity downloads the file to disk and other Activities process it and clean it up.

The strategy is:

• Each Worker process creates two Worker instances:
  • One instance listens on the normal-task-queue Task Queue.
  • Another instance listens on a uniquely generated Task Queue (in this case, uuid is used, but you can inject smart logic here to uniquely identify the Worker, as Netflix did).
• The Workflow and the first Activity are run on normal-task-queue.
• The first Activity returns one of the uniquely generated Task Queues (that only one Worker is listening on—i.e. the Worker-specific Task Queue).
• The rest of the Activities do the file processing and are run on the Worker-specific Task Queue.

Activities have been artificially slowed with sleep(3000) to simulate doing more work.

Running this sample

1. temporal server start-dev to start Temporal Server.
2. npm install to install dependencies.
3. npm run start.watch to start the Worker.
4. In another shell, npm run workflow to run the Workflow.

Example output:

Downloading https://temporal.io and saving to /tmp/b15036de-dbc7-4bc9-b2c7-7c48635c5797

Did some work on /tmp/b15036de-dbc7-4bc9-b2c7-7c48635c5797, checksum: b3fc767460efa514753a75e6f3d7af97

Removing /tmp/b15036de-dbc7-4bc9-b2c7-7c48635c5797

Example history in Temporal Web

Things to try

You can try to intentionally crash Workers while they are doing work to see what happens when work gets "stuck" in a unique queue: currently the Workflow will scheduleToCloseTimeout without a Worker, and retry when a Worker comes back online.

Workflow history with crash

After the 5th attempt, it logs Final attempt 5 failed, giving up and exits. But you may wish to implement compensatory logic, including notifying you.