Repository navigation
perf(gfql): eager native seed lanes record their index engagement in gfql_explain - #2040
Conversation
46f95a6 to
d0c6880
Compare
daa7348 to
994f4bd
Compare
d0c6880 to
3806647
Compare
994f4bd to
479a421
Compare
3806647 to
d0859c1
Compare
479a421 to
530afa4
Compare
d0859c1 to
b32b52a
Compare
530afa4 to
059b2a8
Compare
b32b52a to
a55adf6
Compare
059b2a8 to
bf46a8a
Compare
a55adf6 to
033c1a7
Compare
bf46a8a to
86b7017
Compare
033c1a7 to
2f60508
Compare
86b7017 to
88ff625
Compare
2f60508 to
c4d2e9a
Compare
88ff625 to
6312e18
Compare
c4d2e9a to
ab74105
Compare
6312e18 to
0664c3a
Compare
ab74105 to
682c072
Compare
0664c3a to
4a80847
Compare
682c072 to
373647f
Compare
|
READY (373647f) — restacked on the new #2038 head after #2035 landed (patch-id identical to the previous head). Round-003 amplification (see #2037 / #2038) ran at the #2038 head, which is this PR minus the trace records; the pandas/cuDF served tags in that probe came from spying on the fast-path entry because the eager lanes only record Receipts at 373647f
|
4a80847 to
a3991bb
Compare
373647f to
a5774fe
Compare
|
CI at the rebased head a5774fe (on the rebased #2038): 75/75 check-runs success (tck-gfql, changed-line coverage, polars lane, hygiene guards included). Patch-id identical to the previous head 373647f, so the GPU receipt (1331 passed, 3 failures identical to master, #2043) carries over. Ready to merge after #2038; it auto-retargets to master on that merge and I rebase it within the hour. |
a5774fe to
f817f0b
Compare
…gfql_explain The pandas and cuDF single-node and seeded-hop lanes served through the resident indexes without a trace step, so gfql_explain reported no engagement for shapes the index answered. They now record the same indexed_traversal step the polars lane does; policy off and no-index runs record nothing and keep their answers. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA
f817f0b to
8f82a71
Compare
Owner stipulation on #2037/#2038: the eager (pandas/cuDF) native op-list lanes served seeded shapes through the resident indexes without recording an explain step, so
gfql_explainreportedused_index: Falsefor shapes the index answered while the polars lane (#2038) reports its engagement._single_node_rows_via_index_or_filter(both engines' node-only lane) and_seeded_typed_hop_pandas_cudfnow record anindexed_traversalstep (native_seed_lookupwith the resolver that served,native_seeded_hopwhen the index branch produced the candidates). Policy off and no-index runs record nothing and keep their answers.Pins (
test_native_seed_lane_explain.py, pandas/polars/cuDF): node-only served by the property index →used_index+ seam; seeded typed hop → seam +index_tracepath; policy off →policy_off, no index step, same answer; no resident index → no index step.Stacked on #2038. Related: #2027, #2033.
🤖 Generated with Claude Code
https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA