Fly-Style opened a new pull request, #18936:
URL: https://github.com/apache/druid/pull/18936
## Core ideas around the change
- Scale up more aggressively when per-partition lag is meaningful.
- Relax the PPT (partitions-per-task) increase limit based on lag severity
and headroom.
- Keep behavior conservative near `taskCountMax` and avoid negative headroom
effects.
## Key changes
- **Lag-proportional PPT relaxation**: add `rawExtra` when lag/partition
exceeds thresholds, plus a step function for higher lag.
- **Headroom-aware scaling**: multiply by `max(0, 1 -
currentTasks/maxTasks)` to reduce aggressiveness near max and avoid negative
scaling.
- **No fractional PPT**: PPT remains integer; when PPT is already 1,
scale-up is limited only by `taskCountMax`.
- **Lag-amplified idle decay**: when lag/partition exceeds the activation
threshold, the lag-based busy factor is multiplied
by a ramped multiplier (1.0x at 50K to 2.0x at 500K) to reduce predicted
idle more aggressively during high lag.
## Expected behavior (for example, 30 partitions, `taskCountMax`=30)
- 3 tasks: 50K lag/partition → 8 tasks; 300K → 15 tasks; 500K+ → 30 tasks.
- 10 tasks: 100K → 15; 300K → 30; 500K → 30.
- 20/25 tasks (PPT=1): scale-up capped only by `taskCountMax`.
Note: examples assume lag-dominant conditions (low/mide processing rate, low
idle).
## Open questions / risks
- **Lag skew**: using `aggregateLag / partitionCount` can under-react if a
few partitions are far behind. Consider max/percentile lag if available.
## Potential (further) calibration plan
- Replay historical metrics from `CostBasedAutoScaler#collectMetrics`
(aggregate lag, partition count, current/max tasks, poll idle ratio).
- Sweep parameters with aggressive bias:
- `LAG_ACTIVATION_THRESHOLD` in [25K, 50K]
- `BASE_RAW_EXTRA` in [5, 7, 9]
- `LAG_STEP` in [50K, 100K]
- set `LAG_AGGRESSIVE_THRESHOLD = 2 × LAG_ACTIVATION_THRESHOLD` unless
replay shows early maxing out
- Choose the most aggressive config that keeps overshoot and oscillation
acceptable.
- Guardrails during rollout:
- cap per-cycle task increase, e.g. `min(10, 0.2 * (taskCountMax -
currentTaskCount))`
- minimum dwell time before another scale-up when lag barely exceeds
activation
- alert on `currentTaskCount > taskCountMax`
This PR has:
- [ ] been self-reviewed.
- [ ] using the [concurrency
checklist](https://github.com/apache/druid/blob/master/dev/code-review/concurrency.md)
(Remove this item if the PR doesn't have any relation to concurrency.)
- [ ] added documentation for new or modified features or behaviors.
- [ ] a release note entry in the PR description.
- [ ] added Javadocs for most classes and all non-trivial methods. Linked
related entities via Javadoc links.
- [ ] added or updated version, license, or notice information in
[licenses.yaml](https://github.com/apache/druid/blob/master/dev/license.md)
- [ ] added comments explaining the "why" and the intent of the code
wherever would not be obvious for an unfamiliar reader.
- [ ] added unit tests or modified existing tests to cover new code paths,
ensuring the threshold for [code
coverage](https://github.com/apache/druid/blob/master/dev/code-review/code-coverage.md)
is met.
- [ ] added integration tests.
- [ ] been tested in a test Druid cluster.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
