Windmill CE

Several Windmill features are Enterprise-only in a way that is not visible from the UI: the setting is accepted, stored, read back, and displayed correctly. Only the enforcement is missing.

Everything on this page was measured against the version this repository pins.

Per-key concurrency limits do nothing

The real implementation is in windmill-queue/jobs_ee.rs. The OSS build compiles a stub whose update_concurrency_counter returns Ok((true, None)) — unconditionally allowed. Settings are both saved and read back (get_tag_and_concurrency works), so the UI shows the limit and the gate passes everything through.

This is why politeness here is expressed as loop structure instead — see Following links. Worker groups are also Enterprise-only.

What does work is the for-loop’s parallel / parallelism: that is implemented in worker_flow.rs in the OSS build, via the suspend column on v2_job_queue.

stop_after_if did not stop a while loop

A minimal flow with stop_after_if ran 643 iterations before it was killed by hand. No mechanism that limits load on someone else’s server is built on top of a loop that can do that; crawl_level runs exactly one level and capture-ledger decides whether there is another.

Schema defaults are UI-only

A flow’s input schema can declare default: true. That default is applied for UI-initiated runs only. A webhook or API run receives nothing.

Measured — triggering crawl_level with only its required arguments:

crawl_id          = '00000000-…'
depth             = 0
host_parallelism  = (not passed)     ← schema says default: 4
per_host_delay_ms = (not passed)
respect_robots    = (not passed)     ← schema says default: true

Two consequences worth stating plainly:

1. Scripts must not rely on schema defaults. crawl_host reads its settings from Windmill variables for exactly this reason — an argument that arrives undefined produced “Channel target must be a string”.
2. An absent argument arrives as JSON null, not undefined. TypeScript default parameters only fire for undefined, so null passes straight through. respect_robots did this: robots.txt was never read, a disallowed page was captured, and the crawl reported success. The ?? true in plan_level.ts is the only thing preventing it.

wmill lint catches nothing useful

wmill lint validates flow YAML. Three deliberate breakages, all reported ✅ Lint passed:

broken	result
argument renamed crawl_id → crawlId	passed
path: pointing at a script that does not exist	passed
a nonsense key added to a step	passed

It is not a substitute for the tests.

The flow test UI hides this class of bug

Windmill’s flow editor can run a flow, run one step, or run up to a step. It is useful while writing a flow. It is not a guard:

• there is no CLI, so it cannot run in CI or assert anything
• it runs from the UI, so schema defaults are applied — which means the respect_robots bug above shows green there and broken in production
• “restart from step” is Cloud/Enterprise only

sync push and sync pull both delete

• wmill sync push deletes remote items that are absent locally. It has deleted a flow that was created through the API.
• wmill sync pull deletes local files absent remotely. It has deleted three uncommitted scripts.

Commit before syncing, and push before pulling. pnpm run windmill:diff is sync push --dry-run and shows what would change.

Variables have a size limit

Somewhere between 10,000 and 20,000 bytes. Measured when this repo still carried BrowserHive’s 16,315-byte .proto (until v0.22): it did not fit and lived in a resource instead. Nothing this size is stored any more.

Schedule error handlers are Enterprise-only

A failed nightly run leaves a red entry in the run history and notifies nobody. See the limits in Development.