All follow-ups are merged and final verification is complete.
| Repository / PR |
Merge commit |
| pygraphistry #2077 |
f7a7253bc95d9cc0663bd130a53a03f2f355cbcd |
| pygraphistry #2078 |
d20c6ae1af73e64b05aa4b24459dfd3ba247edf9 |
| pygraphistry #2079 |
65c359b5d94231224010ce00f7bcbbfadcb1e4fb |
| tck-gfql #201 |
84b098c2784849a47e1c3c12bbf0766a2ffe3b4b |
| pyg-bench #261 |
26e90c745f3233578230a96923b27b6bd7505ca9 |
The actual merged product graphistry/ tree exactly matches the tested combined ffa52d8abaaf81f1a3508f2723a7f79bdcda2db0; only CHANGELOG differs. The final benchmark workflow pins product 65c359b5d94231224010ce00f7bcbbfadcb1e4fb.
| Priority / issue |
Root cause and impact |
Implementing PR |
| P1 — #2075 |
Explicit PolarsGPU row aggregation could run eagerly on CPU. Earlier GPU traversal did not prove aggregate execution. Predates2073. |
#2077: strict aggregate collectors/E110, native expressions/constants/global shapes, null/empty and ordered-DISTINCT semantics. |
| P2 — #2076 |
Parser failure became an empty dependency list, admitting malformed sources. Qualified operands and lexical shadowing exposed additional validation drift during amplification. |
#2078: literal precedence, structured syntax/type/dependency errors, visible dotted columns and correctly scoped bindings. Merged after #2077. |
| P2 — graphistry/tck-gfql#200 |
Flattened entities/properties and absent/present-null rows could not be distinguished from IDs or nullable values alone. |
#2079 plus graphistry/tck-gfql#201: typed independent kind/presence, snapshots and OPTIONAL/reentry alignment. Both merged in order. |
| P2 — #2082, explicitly deferred |
Polars1.21 endpoint is_in shape errors reproduce in all33 cases on exact unmodified master92. Current construction traces to Aug15/19, before2073; first affected release unknown. |
Separate old-version compatibility fix. Current Polars/RAPIDS26 succeeds. Failures remain visible. |
Priority was challenged against impact: silent execution on the wrong requested backend comes first; validation and entity correctness follow independently. The old-version traversal defect is independently reproducible baseline debt, not a reason to weaken follow-up oracles. #2075/#2076 were not skipped: #2077/#2078 implement them.
Final verification:
Final product CI34888725964 SUCCESS: 83 successful jobs, 2 conditional skips (no-plans-in-repo, changed-line-coverage). CodeQL SUCCESS.
Current six-module DGX evidence:
| Image/lane |
Result |
Limits |
| RAPIDS25/cuDF25.2.2/Polars1.21 |
1053passed,215skipped,33failed |
All33 failures are baseline CPU-Polars traversal; exact master92 control matches IDs/error types. |
| RAPIDS26/cuDF26.2.1/Polars1.35.2 |
1086passed,215skipped |
cuDF and CPU coverage. |
| PolarsGPU1.35.2/RAPIDS26 |
1301passed,0skipped |
Actual strict GPU execution and structured capability declines; PolarsGPU25 not claimed. |
Boundaries cover valid/invalid syntax, literal-looking columns, constants, missing/qualified operands, lexical scopes, schema changes, empty/all-null/global/grouped behavior, key-only versus zero-output forms, null keys, DISTINCT order/multiplicity, absent versus present-null entities, stale metadata, OPTIONAL alignment and caller immutability. Coalesce validation is valid, but current cudf-polars26 cannot lower it; explicit E110 tests preserve CPU/cuDF numeric oracles and do not silently substitute CPU execution.
Review found and fixed qualified-column rejection, operand-order-dependent lexical scope, overly optimistic GPU coalesce test expectations, and ignored thread-limit receipts. Earlier count dtype/null storage/DISTINCT and entity typing/sentinel/OPTIONAL-fixture fixes remain verified. Entity changes add no getattr/setattr/hasattr/cast calls or B010 suppressions. Final changes have two clean source-review waves; unchanged source reuse is supported by explicit equivalence, not relabeled historical runs.
Performance includes focused aggregate/entity/validation oracles and representative Q1–Q9 across20k/100k andfourengines. Original main matrix had three flags; two predeclared complete ABBA/BAAB controls (3warmups/11samples, cold bindings, explicit4Polars threads) repeat none. First control flags20kGPUQ9; second haszero flags. Actual Q3/Q4/Q9 traces prove equal paths and strict GPU collection on both sources, with unchanged fast-path code. Assessment: no demonstrated repeatable product regression; run variation is the likely explanation, not a proven exact environmental cause or a speedup guarantee. Original flags and the thread-setting difference remain disclosed.
pyg-bench#261 owns the durable main comparator and corrected structural guards. General-chain diagnostics require one edge-index materialization and one final node-universe read; same-result two-read faults fail and zero engagement is invalid. The obsolete whole-query10/14join calibration was not relaxed to15. Allfour full DGX structural/scaling ABBA runs pass6/6probes. Shared locks, verified idle/resource gates,4CPU quota, modest shares/nice and gentle retries protected colleague workloads.
#2075, #2076 and graphistry/tck-gfql#200 are closed. #2082 remains open as explicit P2 baseline compatibility debt; this completion does not claim those failures were fixed. No remaining maintainer action is needed for this follow-up batch.
Full commands, source/image/data/query manifests, XML, raw timing blocks, rejected hypotheses, reviews and performance-resolution.md are retained under local plans/gfql-2073-followups/. Cross-repository CI uploads portable evidence. Main canonical parity rounds floats to four decimal places; it is not an independent full-precision oracle. Independent focused aggregate/entity oracles supplement it.
All follow-ups are merged and final verification is complete.
f7a7253bc95d9cc0663bd130a53a03f2f355cbcdd20c6ae1af73e64b05aa4b24459dfd3ba247edf965c359b5d94231224010ce00f7bcbbfadcb1e4fb84b098c2784849a47e1c3c12bbf0766a2ffe3b4b26e90c745f3233578230a96923b27b6bd7505ca9The actual merged product
graphistry/tree exactly matches the tested combinedffa52d8abaaf81f1a3508f2723a7f79bdcda2db0; only CHANGELOG differs. The final benchmark workflow pins product65c359b5d94231224010ce00f7bcbbfadcb1e4fb.Priority was challenged against impact: silent execution on the wrong requested backend comes first; validation and entity correctness follow independently. The old-version traversal defect is independently reproducible baseline debt, not a reason to weaken follow-up oracles. #2075/#2076 were not skipped: #2077/#2078 implement them.
Final verification:
Final product CI34888725964 SUCCESS: 83 successful jobs, 2 conditional skips (no-plans-in-repo, changed-line-coverage). CodeQL SUCCESS.
Current six-module DGX evidence:
Boundaries cover valid/invalid syntax, literal-looking columns, constants, missing/qualified operands, lexical scopes, schema changes, empty/all-null/global/grouped behavior, key-only versus zero-output forms, null keys, DISTINCT order/multiplicity, absent versus present-null entities, stale metadata, OPTIONAL alignment and caller immutability. Coalesce validation is valid, but current cudf-polars26 cannot lower it; explicit E110 tests preserve CPU/cuDF numeric oracles and do not silently substitute CPU execution.
Review found and fixed qualified-column rejection, operand-order-dependent lexical scope, overly optimistic GPU coalesce test expectations, and ignored thread-limit receipts. Earlier count dtype/null storage/DISTINCT and entity typing/sentinel/OPTIONAL-fixture fixes remain verified. Entity changes add no getattr/setattr/hasattr/cast calls or B010 suppressions. Final changes have two clean source-review waves; unchanged source reuse is supported by explicit equivalence, not relabeled historical runs.
Performance includes focused aggregate/entity/validation oracles and representative Q1–Q9 across20k/100k andfourengines. Original main matrix had three flags; two predeclared complete ABBA/BAAB controls (3warmups/11samples, cold bindings, explicit4Polars threads) repeat none. First control flags20kGPUQ9; second haszero flags. Actual Q3/Q4/Q9 traces prove equal paths and strict GPU collection on both sources, with unchanged fast-path code. Assessment: no demonstrated repeatable product regression; run variation is the likely explanation, not a proven exact environmental cause or a speedup guarantee. Original flags and the thread-setting difference remain disclosed.
pyg-bench#261 owns the durable main comparator and corrected structural guards. General-chain diagnostics require one edge-index materialization and one final node-universe read; same-result two-read faults fail and zero engagement is invalid. The obsolete whole-query10/14join calibration was not relaxed to15. Allfour full DGX structural/scaling ABBA runs pass6/6probes. Shared locks, verified idle/resource gates,4CPU quota, modest shares/nice and gentle retries protected colleague workloads.
#2075, #2076 and graphistry/tck-gfql#200 are closed. #2082 remains open as explicit P2 baseline compatibility debt; this completion does not claim those failures were fixed. No remaining maintainer action is needed for this follow-up batch.
Full commands, source/image/data/query manifests, XML, raw timing blocks, rejected hypotheses, reviews and performance-resolution.md are retained under local plans/gfql-2073-followups/. Cross-repository CI uploads portable evidence. Main canonical parity rounds floats to four decimal places; it is not an independent full-precision oracle. Independent focused aggregate/entity oracles supplement it.