Hi all, I'd like to ask for a review of PR #18725:
https://github.com/apache/iotdb/pull/18725 "Add LTTB downsampling as a built-in Table Model table function" Closes #18532 I have addressed all the review comments received so far. The latest head is 3660c706. Summary of the review rounds: - Duplicate input timestamps within a partition now fail with semantic error 701, including duplicates supplied by different regions and duplicates whose participant values are NULL (issue point 4). - Non-finite values (NaN, +/-Infinity) are eligible and are returned, but they never serve as triangle vertices A or C. One lookahead rule is now shared by target-count and window mode: C averages only the finite points of the next bucket that has one, falls back to the last finite eligible point, and a bucket without a finite candidate returns its earliest point without moving the anchor. - SIZE mode returns the first and last eligible points as buckets of their own, keeps one anchor chain across overlapping windows, and a window with no eligible point emits no row, consistent with M4. - LTTB is excluded from the per-fragment split and runs once over the partition inputs merged in order; planner tests assert the resulting plan shape (TableFunctionProcessor over MergeSort). - The buffer budget moved out of udf-api into an engine-internal interface, is resolved once per processor provider, and is charged to the operator's memory estimate, so a hot reload cannot make the reservation and the enforced limit diverge. The configuration is documented in iotdb-system.properties.template. - Added integration coverage for SLIDE, ORIGIN, invalid TIMECOL, and duplicate timestamps across devices of one partition, for the direct scan and for the window-function and SESSION inputs. Since the last review, commit 3660c706 additionally fixes a window emission stall: when a participant column has exactly one eligible point in the partition, `isReady` compared against the default value of `lastEligiblePositions` and the first window never became ready, so the whole partition was buffered and the query failed with the buffer limit error. Emission now streams in that case. A regression test covers both count and time windows. Local verification: the 46 LTTB unit tests in node-commons pass (LTTBNumericBoundaryTest 15, LTTBRegressionTest 10, LTTBTableFunctionTest 13, LTTBWindowMemoryTest 8), and spotless passes. CI has not run on this PR yet; a run approval would be appreciated. Best， DaZuiZui
