Caideyipi commented on PR #17871:
URL: https://github.com/apache/iotdb/pull/17871#issuecomment-6848124071

   Independent re-review completed: one full round of the 86-file diff and 
relevant implementation, API, and test context.
   
   - Reviewed head: `111fd493b7a4ab63437394e50beaa09ae72b7c94`
   - Comparison base: `cca1dbbde6d0dffa38495c92c92a48b2626d2f99`
   - Final local commit, confirmed source branch head, and confirmed PR head: 
`32f25d9cedd8af43529c3c0ea26487b2b7344c6a`
   
   Two behavioral issues were found and fixed in this round. Locations below 
refer to the reviewed head.
   
   1. **[P2] DTW raw and continued input did not invalidate cached matches when 
timestamps went backwards or an earlier value was replaced.** 
[DTWState.java:108](https://github.com/Caideyipi/iotdb/blob/111fd493b7a4ab63437394e50beaa09ae72b7c94/library-udf/src/main/java/org/apache/iotdb/library/match/model/DTWState.java#L108).
 With pattern `1,2,3` and threshold `0`, feeding timestamps `6,5,4,3,2,1` with 
corresponding values `3,2,1,3,2,1` missed both chronological matches. Likewise, 
after merging and observing times `1,3,4`, inserting time `2` did not rebuild 
the historical windows. Updating an earlier value could leave an obsolete 
match. Non-increasing timestamps now mark the state for the same deferred 
chronological rebuild used by merges. Regressions cover batches, late 
continuation after output, historical replacement, serialization before 
rebuilding, and repeated output.
   
   2. **[P2] Moving-average compensation could itself lose small values 
permanently.** 
[UDTFMvAvg.java:97](https://github.com/Caideyipi/iotdb/blob/111fd493b7a4ab63437394e50beaa09ae72b7c94/library-udf/src/main/java/org/apache/iotdb/library/dprofile/UDTFMvAvg.java#L97).
 For `window=3` and input `[1e300,1e150,1,1,1]`, the final window `[1,1,1]` 
returned `0` instead of `1`; all sums remained finite, so overflow recovery did 
not help. The fix detects lost compensation bits or overflowing additions, 
reconstructs the active window sum exactly once, and then updates that exact 
sum incrementally. The regression covers both signs, subsequent windows, and 
resetting the function for another run.
   
   The review also covered MAD null handling and INT64 extremes, sampling Row 
reuse and typed INT64 output, DWT levels, Histogram normal/subnormal/extreme 
ranges, MinMax extremes, Spline timestamps/coordinates/endpoints/duplicate 
timestamps and fixtures, AR/DTW validation, LOF valid-window coverage, and DTW 
history, cross-partition matches, interleaved/reversed/multilevel merges, 
serialization, continuation, and repeated output.
   
   Validation:
   
   - All three new regressions failed against the reviewed production 
implementation, then passed after repair.
   - Local offline `mvn -o clean test -pl library-udf -am`: **142 tests, zero 
failures/errors/skips**.
   - Local offline Chinese `clean test-compile -pl library-udf -am -P 
with-zh-locale -DskipTests`: passed.
   - Spotless with `with-integration-tests`, lifecycle Checkstyle (zero 
violations), and `git diff --check`: passed.
   - Required IoTDB dependencies were rebuilt from clean source; both clean 
reactors disabled local and remote build caches. All newly added tests were run 
locally. The local offline fallback retained the previously established remote 
Maven plugin-download limitation; no new external dependencies were introduced.
   - No full-project build or performance benchmark was run. CI was neither 
waited on nor polled.
   
   Resource assessment: DTW still retains O(N + R) state and serialized data 
for N effective input points and R matches. A pending final rebuild costs O(N * 
L^2) for pattern length L, with O(L) DTW scratch space. Combines insert 
incoming points without recalculating all historical DTW windows; serialization 
and multilevel transfer still traverse retained input. This is an algorithmic 
assessment, not a measured performance result. The moving-average exact 
fallback retains the active window and an incremental sum, with a one-time 
window reconstruction.
   
   The source branch was checked immediately before pushing and had no 
concurrent update. The source branch and PR head were both confirmed at the 
final SHA above, and the PR remains **Draft** because this round found 
behavioral issues. A fresh independent full-diff re-review is required before 
Ready. No second review round was started after these repairs.
   


-- 
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