Caideyipi opened a new pull request, #18812:
URL: https://github.com/apache/iotdb/pull/18812

   ## Description
   
   In a mixed large-field and numeric subscription workload, one queue retained 
most of a DataNode's subscription materialization budget while other queues 
also rejected real-time entries. The old warning called every rejection "queue 
full", even when capacity remained. The reported enterprise run (case 1392, 
result 27524) consumed 18,919,210 of 24,693,910 written rows before its 
consumer timed out. The source counts matched the writes. This PR isolates the 
materialized-tablet budget by queue and makes admission failures observable; it 
does not claim that this allocation change alone explains or resolves the 
entire observed delivery stall.
   
   ### Admission and recovery
   
   - Give each active consensus subscription queue an equal share of the 
DataNode materialization budget. A queue cannot reserve another queue's share. 
Existing allocations drain through delivery and ACK if a newly active queue 
reduces the share.
   - Preserve already accepted pending requests during ordinary backpressure. 
Cleanup and seek still clear stale pending requests, and rejected real-time 
requests remain eligible for WAL replay.
   - Return a critical poll error containing `SUBSCRIPTION_OVERSIZED_ENTRY` 
when one materialized WAL entry exceeds its queue's entire share. Do not 
advance that entry's WAL or commit progress. This is an intentional behavior 
change from the empty-budget oversized-entry exception; operators must increase 
the materialization budget, reduce concurrent queues, or narrow the topic to 
consume such an entry.
   
   ### Diagnostics and configuration
   
   - Report a stable `reasonCode` for real-time admission rejection, 
distinguishing subscription quota/limit, consensus request memory, writer 
backlog, queue capacity, and lifecycle states. Add queue-level memory 
used/quota and rejection counters to metrics and core reports. Capture offer 
rejection reasons per writer thread so concurrent offers cannot mislabel one 
another.
   - Explain in the config template and a Chinese user-guide page that 
`subscription_cache_memory_usage_percentage` controls serialized poll-response 
caching, not the materialized-tablet budget. Start at the default `0.2`; 
compare `0.3`/`0.4` only if response-cache eviction or reserialization is high 
and pipe memory has headroom. Size the ninth 
`chunk_timeseriesmeta_free_memory_proportion` share from the largest 
materialized entry, active queue count, and desired buffering depth. The 
repository has no published documentation navigation, so the new page still 
needs integration into the documentation site.
   
   ### Validation and limits
   
   - 56 focused tests passed: 54 DataNode and 2 consensus tests, including 
mixed-size queues, pending preservation, WAL recovery, oversized-entry errors, 
lifecycle cleanup, and concurrent rejection reporting.
   - Full English reactor `test` build and Chinese reactor `test-compile 
-DskipTests -P with-zh-locale` passed; Spotless and `git diff --check` passed.
   - The enterprise 3c1d/2-writer workload has not been rerun on this change. 
The 960 MB retained-tablet snapshot in result 27524 was taken after the 
consumer exited, so it does not by itself prove the cause of the earlier 
67-second stop. Re-run that workload while collecting rejection codes, memory 
pools, prefetch-task durations, and ACK/GC timing.
   
   ### Related PR
   
   #18808 adds delivery-stall and node-memory diagnostics and retains the 
previous single oversized-entry exception. These PRs overlap in 
`ConsensusPrefetchingQueue`, metrics, memory manager, and localized messages. 
Their metric and oversized-entry semantics need reconciliation before merging 
both.
   
   <hr>
   
   This PR has:
   - [x] been self-reviewed for concurrent reads and writes.
   - [x] added documentation and Javadocs for the changed behavior.
   - [x] added focused unit tests for new paths.
   - [ ] been tested in a test IoTDB cluster with the reported enterprise 
workload.
   
   <hr>
   
   ##### Key changed classes
   
   `SubscriptionMemoryManager`, `ConsensusPrefetchingQueue`, 
`SubscriptionQueueRegistry`, and `ConsensusSubscriptionPrefetchingQueueMetrics`.
   


-- 
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