[cdoern]

This proposes a bounded job queue in core, with a worker pool, a per-job timeout, and cancellation, so that a filter can enqueue work which outlives the request that created it. It covers the jobs half of praxis-proxy/praxis#807 and leaves probes to a separate proposal, since the two share nothing beyond supervision.

The research behind it narrowed the scope considerably. FilteredSubrequestExecutor already runs a request through a real filter pipeline with no client connection attached, and there is a passing test driving exactly that against a backend today. What is missing is not an execution engine but a queue, a worker pool, and a lifetime for them.

Per the process this first PR contains only What? and Why?. The How? section listing the implementing PRs will follow once the direction is agreed.

The section worth the most scrutiny is the relationship to ENH #1042. That proposal cancels background work when its pipeline drops, which is what a token refresher wants but would cancel every in-flight job on a config reload. I have argued that jobs need process lifetime instead, and listed @aslakknutsen as a stakeholder for that reason.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}