Caideyipi opened a new pull request, #18861:
URL: https://github.com/apache/iotdb/pull/18861

   ## Description
   
   Schema mutations can leave alias indexes and statistics inconsistent. 
Deleting wide devices also repeatedly scans their children, and repeated 
table-device creation/deletion retains every historical attribute slot.
   
   - Validate aliases against other aliases and measurement names in an aligned 
batch before changing the tree, in both Memory and PBTree modes. Preserve 
aliases equal to their own measurement name and reject conflicting merge 
batches before mutation.
   - Keep matched-device counts on blacklist retries, while decrementing device 
statistics only on the first state transition.
   - Maintain direct measurement and non-view measurement child counts to avoid 
a child scan for each deleted measurement. Snapshot loading rebuilds the 
counters; nested devices, templates, and view-only alignment cleanup retain 
their behavior. The extra 8 bytes per internal node are included in memory 
estimates.
   - Account table attribute memory using value-size changes, independently of 
the full map-entry changes used by region memory accounting.
   - Reuse deleted attribute slots using a bitmap, including after snapshot 
recovery. Live pointers and the snapshot format are unchanged; retained list 
references and allocated bitmap storage are accounted for, and store operations 
use the same monitor.
   
   ## Validation
   
   - 191 tests passed across `DeviceAttributeStoreTest`, 
`DeviceBlackListConstructorTest`, `MTreeBelowSGMemoryImplTest`, 
`SchemaRegionAliasAndTagTest`, `SchemaRegionTableDeviceTest`, 
`SchemaStatisticsTest`, `SchemaRegionBasicTest`, `SchemaRegionTemplateTest`, 
and `MNodeTest`. Selected test sources and helpers were compiled separately 
against fresh reactor output and run through JUnitCore, with one JVM per test 
class as configured by Surefire.
   - `mvn -nsu package -pl distribution -am -Dmaven.test.skip=true`: all 39 
modules passed. The all-in-one and Edge distribution server jars match the 
fresh DataNode jar by SHA-256 and contain all six modified production classes.
   - `mvn spotless:apply -pl iotdb-core/datanode` and the reactor's 
Checkstyle/Spotless checks passed.
   - `mvn -nsu test-compile -Dmaven.test.skip=true -P with-zh-locale`: all 53 
reactor modules passed.
   
   The usual Maven test invocation is blocked by an unchanged master test, 
`AlignedWritableMemChunkTest`, which still calls 
`AlignedWritableMemChunk(List<IMeasurementSchema>, boolean)` at four sites 
although that constructor no longer exists. Packaging therefore skips test 
compilation; the 191 selected tests were compiled and run separately as 
described above.
   
   The original reproduction harness, with changed production classes freshly 
compiled, reports:
   
   - Repeated predelete: device count remains 0; rollback restores 1.
   - Clearing an attribute: table attribute memory returns to 0.
   - 100,000 empty-attribute create/delete cycles: 1 retained slot, 12 
accounted bytes, and an 8-byte snapshot (previously 100,000 slots and a 
400,004-byte snapshot).
   - Deleting 64,000 aligned measurements in blacklist traversal order: about 
19 ms in this local run (previously 5–8 seconds). This is a local reproduction 
benchmark, not a cluster throughput measurement.
   
   - [x] Self-reviewed.
   - [x] Added regression tests for alias conflicts, retries/rollback, 
attribute accounting/reuse, snapshot recovery, nested-device cleanup, view-only 
cleanup, and deletion without repeated child scans.
   


-- 
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