Azure-devops-docs: AgentService.js does not reap zombie processes

[egrubbs]

For Linux docker containers, exec ./externals/node/bin/node ./bin/AgentService.js interactive becomes PID 1, but it does not reap zombie processes. Docker containers that run multiple pipeline jobs end up with a huge amount of zombie processes which makes it harder to troubleshoot.
The quick fix is to use docker run --init which causes PID 1 to be /sbin/docker-init, which will reap zombie processes. It will also forward signals, so AgentService.js should still get the TERM and INT signals that it needs to restart and self-update.
I think the documentation should be updated to tell the user either:

• Use docker run --init
• The Dockerfile should install an init program

Alternatively, AgentService.js could be updated to reap zombie processes.

https://ahmet.im/blog/minimal-init-process-for-containers/

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: b31d4f6c-d16d-2ba1-1294-6aa94e328046
• Version Independent ID: 88aec599-4410-697d-1002-1c41b1c17bb6
• Content: Running a self-hosted agent in Docker - Azure Pipelines
• Content Source: docs/pipelines/agents/docker.md
• Product: devops
• Technology: devops-cicd-agents
• GitHub Login: @steved0x
• Microsoft Alias: sdanie

[steved0x]

@egrubbs Thank you for the feedback, I am passing this along to the team.

[vtbassmatt]

We don't really expect containers to run for longer than the length of a single job. Can you share more about your use case?

[egrubbs]

Microsoft-hosted agents exist in containers that only run for a single job, but the documentation for setting up a docker self-hosted agent uses the default behavior of AgentService.js, which continually requests new jobs.

The documentation for self-hosted Linux (not specific to docker) mentions the ability to use ./run.sh once to only process a single job, but the documentation does not recommend this as the normal way to run an agent. It even has directions on using systemd to run the agent as a service.
https://docs.microsoft.com/en-us/azure/devops/pipelines/agents/v2-linux?view=azure-devops#run-once

Only processing a single job per container is not mentioned at all in the documentation for setting up a self-hosted Linux agents inside of docker.
https://docs.microsoft.com/en-us/azure/devops/pipelines/agents/docker?view=azure-devops#linux

[vtbassmatt]

Good point, I should update our recommendations there. For a normal, VM-based self-hosted agent, it's not common to want --once. For container-based agents running in some kind of outside orchestrator (Kubernetes, ACI, etc.), we do recommend --once so that the agent goes away and the orchestrator can reap the container.

[vtbassmatt]

A doc-based fix has been merged and should publish today. I'm closing this issue since I'm likely to forget to circle back to it.