Adrien Faure (1e6fbc07) at 19 Mar 09:31

... and 1 more commit

Adrien Faure (fda70afc) at 19 Mar 09:31

Summary This MR fixes the per-action performance override path in the runner placement scorer.

The previous implementation had two problems:

• it applied a performance override incorrectly inside the generic objective loop,
• and after the first fix, the scheduler no longer consumed PendingExecutionConnection.node_pool_objective_score in the normal schedule_one() path.

This MR keeps the scorer fix and restores compatibility with scheduler-side pending-execution overrides, without changing schema, ORM mapping, or repository structure.

What changed

ryax/runner/application/site_placement_service.py

• _score_on_objectives() now accepts an explicit performance_overrides map.
• Performance overrides are applied only to the PERFORMANCE objective.
• Override values are now multiplied by the performance weight, like any other objective contribution.
• Fallback is now per pool:
  • if a pool has an override, use it for performance,
  • otherwise fall back to the generic node-pool performance score.
• The old all-or-nothing behavior was removed:
  • we no longer discard all overrides if one pool is missing.
• The old triple-counting behavior was removed:
  • the same performance override is no longer added once per objective.
• Compatibility was restored for the actual scheduling path:
  • if no explicit performance_overrides argument is provided,
  • _score_on_objectives() now reuses to_schedule.node_pool_objective_score when it is a real dict[str, float].
• Invalid non-dict node_pool_objective_score values are ignored safely with a debug log instead of crashing.

tests/runner/application/test_site_placement_service.py:

• Added small helpers to build node pools with objective scores and pending executions with objective weights.
• Added regression tests for:
  • generic scoring without overrides,
  • performance-only override behavior,
  • partial override fallback per pool,
  • 0.0 override values,
  • missing generic objective scores,
  • no-objectives behavior,
  • schedule_one() using pending-execution override scores in the real path,
  • explicit override argument taking precedence over pending-execution scores,
  • invalid stored override values being ignored safely.

Summary This MR fixes the per-action performance override path in the runner placement scorer.

The previous implementation had two problems:

• it applied a performance override incorrectly inside the generic objective loop,
• and after the first fix, the scheduler no longer consumed PendingExecutionConnection.node_pool_objective_score in the normal schedule_one() path.

This MR keeps the scorer fix and restores compatibility with scheduler-side pending-execution overrides, without changing schema, ORM mapping, or repository structure.

What changed

ryax/runner/application/site_placement_service.py

• _score_on_objectives() now accepts an explicit performance_overrides map.
• Performance overrides are applied only to the PERFORMANCE objective.
• Override values are now multiplied by the performance weight, like any other objective contribution.
• Fallback is now per pool:
  • if a pool has an override, use it for performance,
  • otherwise fall back to the generic node-pool performance score.
• The old all-or-nothing behavior was removed:
  • we no longer discard all overrides if one pool is missing.
• The old triple-counting behavior was removed:
  • the same performance override is no longer added once per objective.
• Compatibility was restored for the actual scheduling path:
  • if no explicit performance_overrides argument is provided,
  • _score_on_objectives() now reuses to_schedule.node_pool_objective_score when it is a real dict[str, float].
• Invalid non-dict node_pool_objective_score values are ignored safely with a debug log instead of crashing.

tests/runner/application/test_site_placement_service.py:

• Added small helpers to build node pools with objective scores and pending executions with objective weights.
• Added regression tests for:
  • generic scoring without overrides,
  • performance-only override behavior,
  • partial override fallback per pool,
  • 0.0 override values,
  • missing generic objective scores,
  • no-objectives behavior,
  • schedule_one() using pending-execution override scores in the real path,
  • explicit override argument taking precedence over pending-execution scores,
  • invalid stored override values being ignored safely.

Adrien Faure (4ce8416a) at 19 Mar 07:44

A few questions.

• Is it intended that the function is never called with the new parameter performance_overrides (except in the tests)?

• Second question:

it applied a performance override incorrectly inside the generic objective loop,

Do you mean this loop (!412 (diffs)) ?

⭐

Adrien Faure (325c8795) at 18 Mar 14:13

Adrien Faure (85e8b081) at 18 Mar 14:13

Ok thank you.

Do we care about the failing tests ?

Adrien Faure (4d01ea97) at 17 Mar 15:34

@mercierm I tested it with workflows executing a lot of sleep action on my machine with this setup:

• emit every 1s (limited to 1000)
• sleep 5
• sleep 4
• sleep 3
• sleep 2
• sleep 1

All workflows finished successfully. Max deployment is reached, and the scheduler manages to reuse old deployments.

Adrien Faure (9f6246fb) at 17 Mar 15:26

... and 9 more commits

Adrien Faure (b3acb6a6) at 17 Mar 13:41