* Balance controller node selection during capacity ties
When multiple control-plane instances have equal remaining capacity,
the instance selection algorithm always picks the first instance in
iteration order. This causes burst workloads to concentrate all job
management on a single controller pod while others remain idle.
Add a tie-breaking criterion: when would_be_remaining capacity is
equal, prefer the instance with fewer jobs_running. This distributes
the control overhead (event processing, callbacks, output streaming)
more evenly across available controller pods during burst scenarios.
The change is backwards-compatible: when capacity clearly differs,
the existing "most remaining capacity" logic dominates unchanged.
Signed-off-by: Alexey Masolov <amasolov@redhat.com>
Signed-off-by: Alexey Masolov <alexey.masolov@gmail.com>
Made-with: Cursor
* Track jobs_running on controller node for container-group tasks
Address review feedback: container-group tasks only set controller_node
(execution_node is empty), so the previous consume_capacity call on the
control path never incremented jobs_running. This meant the tie-breaker
could not differentiate between nodes.
Now pass job_impact=True on the control path when the controller is not
also the execution node (avoiding double-counting for hybrid nodes).
Also improve test coverage:
- Fix test case to prove capacity dominates over jobs_running
- Add dedicated tests for container-group controller distribution
- Add test verifying equal-capacity tie-breaking behaviour
Signed-off-by: Alexey Masolov <amasolov@redhat.com>
Signed-off-by: Alexey Masolov <alexey.masolov@gmail.com>
Made-with: Cursor
* Add docstrings to consume_capacity and fit_task_to_most_remaining_capacity_instance
Address CodeRabbit docstring coverage warning by documenting the two
methods modified in this PR. No functional changes.
Signed-off-by: Alexey Masolov <amasolov@redhat.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Signed-off-by: Alexey Masolov <amasolov@redhat.com>
Signed-off-by: Alexey Masolov <alexey.masolov@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Liam Allen <lallen@redhat.com>
