fix(gfql): polars plain single-hop chains collapse duplicate node rows (#2051) - #2060
fix(gfql): polars plain single-hop chains collapse duplicate node rows (#2051)#2060lmeyerov wants to merge 2 commits into
Conversation
|
CI receipt at 95d03e2: 40 check-runs success, 17 skipped by the CI path filter as whole jobs (umap/spark/full-ai/neo4j/graphviz/dgl/docs/build matrices and the cypher-frontend gates). The GFQL lanes all ran and passed: tck-gfql, gfql-benchmarks, test-gfql-core (3.12, 3.14), test-pandas-compat-gfql (legacy py3.9, latest py3.14), test-polars (3.9–3.14), python-lint-types (3.8–3.14). Local at the same head: 1374 pass on the polars chain suites with the #2051 pins flipped from strict xfail to pass (see the PR body). Stacked order for landing: #2055 → #2056 → this PR. The dgx cuDF + polars-gpu lane runs after the step-7 window on dgx closes (a timed benchmark lane is running there now); I will post it here. |
|
CI receipt at b48f87b (the #2043 dedup marker dropped): 62 check-runs success, 7 skipped by the path filter as whole jobs; tck-gfql, test-polars ×6, test-gfql-core, gfql-benchmarks and the lint/guard lanes all green. Ready for owner merge after #2055 → #2056 (this PR sits on master directly, so it can also go first). |
…e id (#2051) The unnamed, untyped single-hop chain shape is served by the polars chain's skip-combine branch, which built its node frame with a semi-join against the node table and so kept a node table's duplicate rows for the same id, where the full polars chain, pandas and cuDF collapse them. Root cause confirmed by route-off amplification: with the branch forced to decline, the #2051 strict expected failures pass. The branch now applies the same one-row-per-id rule as the full chain's node materialization. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA
…real-GPU receipt), marker dropped
b48f87b to
347d975
Compare
|
Folded into #2056 (owner asked for a shallower stack); the commits, receipts and CHANGELOG entries live there unchanged. |
1 participant
Closes #2051. Independent of the #2055/#2056 stack (polars chain only).
Defect. The unnamed, untyped single-hop chain shape on polars is served by the chain's skip-combine branch, whose node frame is a semi-join against the node table, so a node table carrying the same id twice came back with both rows. pandas, cuDF, the full polars chain,
hop(), and every named/typed/multi-hop/undirected polars shape collapse the duplicate.Root cause, confirmed by amplification. Replaying the existing suites with
GFQL_ROUTES_OFF=polars-plain(the route-off switch from #2054) made the #2051 strict expected failures XPASS: with the branch out of the way the full chain collapses the rows. One line: the branch now applies the sameunique(subset=[node])rule as the full chain's node materialization.Pins.
graphistry/tests/compute/gfql/lazy/engine/polars/test_chain_duplicate_node_rows_2051.py: the two leaking shapes flip from strict expected failure to green against pandas; the five collapsing shapes andhop()stay green.Receipts (local, polars 1.42, cudf 25.10): 1374 passed / 176 skipped / 17 xfailed across the polars mirror suites, residual-polars, native seed resolution, polars conformance matrix and binding rows, chain, endpoint-closure matrix, collision matrix and index suites; guards OK; mypy identical to master; ruff clean. dgx GPU lane after the step-7 window.
🤖 Generated with Claude Code
https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA