deardeng opened a new pull request, #68995:
URL: https://github.com/apache/doris/pull/68995
Problem Summary: Cloud tablet rebalancer rounds retained current and future
BE-to-tablet indexes at partition and table levels as well as the global
indexes, duplicating route membership across a round. Build partition and table
route views when each balance phase visits them, retain only per-table counts
during the global phase, and add a unit test for the index lifetime. Preserve
global routes, warmup rollback, scheduling order, and transfer thresholds.
Views built later in a round can observe concurrent route updates.
For an offline A/B test on the diandian metadata image, compile separate
full FE packages from the pre-L2 and L2 source with Maven build cache disabled.
Run fresh FE processes in A-B-B-A order with the same image, JDK 17.0.2, 128
GiB heap, configuration, and offline-startup patches. After 300 seconds of
warmup, record about 360 seconds of JFR and GC logs per run. Excluding each
first round, the two per-run median balance times are 159.273/168.325 seconds
before and 106.040/107.717 seconds after L2, a 34.75% reduction using the means
of those medians.
In separate A-B-B-A runs, trigger live class histograms at the same
partition-balance stage after route statistics are built. Live heap is
35.511/35.512 GiB before and 16.322/16.321 GiB after L2, a 19.190 GiB (54.04%)
reduction. The difference is almost entirely ConcurrentHashMap route indexes.
After a round completes, live heap is about 16.05-16.09 GiB in both versions
because those temporary indexes have been released.
Across the fixed JFR windows, Young GC pause totals are 8.096/7.651 seconds
before and 3.584/3.341 seconds after L2; Young GC counts are 19/19 and 18/18.
Rebalancer allocation per second does not materially decrease. Maximum total
heap occupied before GC over each full run is 104.983/105.215 GiB before and
99.195/100.434 GiB after L2; this includes reclaimable objects and is not live
heap. The offline FE had no real MS/BE or user query traffic, so these results
do not establish production memory requirements or route migration throughput.
### Release note
None
### Check List (For Author)
- Test: Unit Test: 62 related FE tests passed. Build: both full FE packages
passed `./build.sh --fe --clean` with Checkstyle and Maven build cache
disabled. Manual test: the A-B-B-A JFR/GC test and separate phase-aligned
live-heap A-B-B-A test described above. Regression test: cloud balance Docker
cases passed in the author's local run.
- Behavior changed: Yes (partition and table route views are rebuilt when
visited and may observe a concurrent route update within the same round)
- Does this need documentation: No
### What problem does this PR solve?
Issue Number: close #xxx
Related PR: #xxx
Problem Summary:
### Release note
None
### Check List (For Author)
- Test <!-- At least one of them must be included. -->
- [ ] Regression test
- [x] Unit Test
- [ ] Manual test (add detailed scripts or steps below)
- [ ] No need to test or manual test. Explain why:
- [ ] This is a refactor/code format and no logic has been changed.
- [ ] Previous test can cover this change.
- [ ] No code files have been changed.
- [ ] Other reason <!-- Add your reason? -->
- Behavior changed:
- [x] No.
- [ ] Yes. <!-- Explain the behavior change -->
- Does this need documentation?
- [x] No.
- [ ] Yes. <!-- Add document PR link here. eg:
https://github.com/apache/doris-website/pull/1214 -->
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label <!-- Add branch pick label that this PR should
merge into -->
--
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]
