mirror of
https://github.com/GreptimeTeam/greptimedb.git
synced 2026-09-13 00:42:14 +00:00
Lei, HUANG
c3ea022de5
* perf(mito2): optimize flat merge heap and primary-key interleave Replace the per-row BinaryHeap pop/push cycle in FlatMerge with an in-place root mutation plus a single sift-down repair on a custom RootHeap, keeping the cold heap and direct-batch fast path unchanged. Fallible or awaiting batch transitions move the hot node out of the heap first, preserving error and cancellation semantics. Exploit the globally sorted merge output to build the internal Dictionary<UInt32, Binary> primary-key column with a one-pass ordered gather: append a Binary value only when the PK changes and reuse the current key for adjacent equal PKs, bypassing Arrow dictionary masks, hash interning and key remapping. Non-PK columns still use Arrow interleave. Also cache the current primary-key byte range in RowCursor to avoid repeated dictionary range decoding during comparisons, and add a setup-free Criterion benchmark with exact output-row assertions. 32-way/1 row-per-series/40-tag improves 955.79ms -> 562.36ms (-41.2%); 0-tag -39.5%, 64 rows/series -79.8%, 8-way -56.1%, single-iterator control +0.3%. Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * test(mito2): add rows-per-series sweep to flat merge bench Add 32-way/40-tag shapes for 1, 10, 100, 1000 and 10000 rows per series, and allow FLAT_MERGE_BENCH_SHAPE to match shape name prefixes so the whole sweep can run in one invocation. Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * test(mito2): add oracle-based correctness tests for RootHeap Drive RootHeap and a std BinaryHeap oracle with the same seeded op sequence (push / pop / mutate-root + repair) and assert peek, len, best_child and the full drain order after every operation. A second run with a tiny value range makes duplicates dominate, covering the equal-key branches of sift_up/sift_down and best_child. Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * perf(mito2): replace hot heap with a tournament tree in flat merge Replace the hot RootHeap with a fixed-capacity tournament (winner) tree over per-node slots: every internal node caches the champion of its subtree, so advancing the winner only replays the ~log2(k) nodes on its leaf-to-root path with one compare per level, instead of the heap's two-compares-per-level sift that also re-compares the same node pairs on every row. Two fast paths keep dense shapes at O(1) per row: - champion retention: after mutating the winner in place, skip the replay entirely when it still beats the runner-up (its path caches are unchanged by construction); - a second-best slot cache, invalidated on any structural change, so the retention check costs a single compare without walking the tree. The cold heap, hot/cold overlap window, direct-batch fast path and the remove-before-fallible-fetch batch transition semantics are unchanged. Vs the RootHeap version: 1rps/32way/40tag -19.7%, 0tag -34.4%, 8way -15.9%, 64rps -30.5%, sweep 10/100/1000/10000rps -29~32%; vs the original BinaryHeap baseline the main shape is -52.8%. The single-iterator control is +8% (+50ns one-time construction allocation, no merge work). Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * fix(mito2): support generic schemas in flat merge Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * fix(mito2): satisfy clippy in flat merge benchmark Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * perf(mito2): cache flat merge primary key index Compute the internal primary-key column index once when constructing BatchBuilder and reuse it for every output batch. Preserve the column-name gate for generic schemas. Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> * test(mito2): benchmark high-fan-in flat merges Add sparse 64, 128, 256, and 512-way merge shapes while keeping the total input fixed at 3.2 million rows. Compared with the merge-base heap implementation, median time improves by 56.0%, 60.8%, 55.6%, and 56.1%, respectively. Signed-off-by: Lei, HUANG <ratuthomm@gmail.com> --------- Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>