Caideyipi opened a new pull request, #18849:
URL: https://github.com/apache/iotdb/pull/18849

   When consensus subscription replay scans a wide WAL prefix or a fragmented 
request, a single `ProgressWALIterator.hasNext()` can read many physical 
entries before returning a logical request. The queue's logical-request limit 
does not bound this work, so one region can occupy a shared prefetch worker for 
minutes while other regions stop delivering data.
   
   Share a physical-entry scan budget and a cooperative 50 ms time slice across 
WAL catch-up, pending-gap recovery, readability probes, and end-of-round 
scheduling. A yielded scan preserves its reader position, pending fragments, 
and buffered retry request, publishes completed tablets, and requeues the task 
so other regions can run. Repeated realtime jumps continue a yielded gap scan 
instead of reopening it at the old cursor. A yield is distinct from WAL 
exhaustion and does not trigger a WAL roll or an incomplete-scan error. The 
public unbounded iterator API retains its existing behavior.
   
   Regression coverage checks skipped prefixes and replicated request fragments 
across several rounds, deadline expiry and buffered retries, and repeated 
pending gaps with exactly one delivery per request. Existing memory 
backpressure, seek, WAL corruption/retry, and queue tests also pass.
   
   Validation:
   
   - `mvn -nsu -Ddevelocity.off=true -pl iotdb-core/datanode spotless:apply`
   - `mvn -nsu -Ddevelocity.off=true -pl iotdb-core/datanode -am test-compile 
-DskipTests`
   - `mvn -nsu -Ddevelocity.off=true -pl iotdb-core/datanode -am test 
-Dtest=ProgressWALIteratorTest,ConsensusPrefetchingQueueTest,ConsensusPrefetchingQueueWalBackpressureTest,ConsensusPrefetchingQueueSeekTest
 -Dsurefire.failIfNoSpecifiedTests=false` — 76 tests, 0 failures/errors, 1 
optional benchmark skipped.
   - `mvn -nsu -Ddevelocity.off=true -pl iotdb-core/datanode -am test-compile 
-P with-zh-locale -DskipTests`
   - `git diff --check`
   
   The time slice is checked between WAL operations and cannot interrupt an 
individual file read. The original hour-long, wide-table workload has not been 
rerun, so its throughput and stall-duration improvement still need workload 
validation.
   


-- 
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]

Reply via email to