Skip to main content

Menu

Research

Accepting a Parquet delivery for evaluation

Language: English

Before a score is reported, check which delivery arrived, which cases the scoring reader selected, and which labels it used for those cases. Choose the checks against the delivery agreement. A fixed release may require exact files at exact paths. A permitted rewrite may allow the files and row order to change while preserving the case IDs and their labels.

Verify the fixed delivery and the scored roster

Pair each relative Parquet path with its SHA-256. Use the path spelling as delivered, including percent-encoded characters. Verify the complete mapping, including missing and extra paths. Then run the scoring filter with the intended reader, version and partition options. Compare the returned IDs with the approved roster and check uniqueness if the agreement calls for one row per case.

In a three-shard fictional delivery, the intended roster is case-a and case-b, with two correct outcomes. Both layouts use identical file checksum inventories and every tested reader selects two cases.

Observed resultWhich condition needs attention
Original delivery, PyArrow URI decoding selects case-a and case-bThe path mapping and roster both match.
Original delivery, PyArrow literal segments selects case-c and case-dThe path mapping matches. Hold scoring for the roster mismatch.
Swapped shards, PyArrow URI decoding selects case-c and case-dHold for both the new path-to-file assignments and the roster mismatch.
Swapped shards, PyArrow literal segments selects case-a and case-bThe roster matches. Hold for the new path-to-file assignments.

In the original, PyArrow 26.0.1 with URI decoding interprets split=test%252Fgold as the split value test%2Fgold. With literal segments, it selects the directory spelled split=test%2Fgold for that same query. The literal scan records one correct outcome out of two on the original. DuckDB 1.4.4 with Hive partitioning matches the URI scan in both layouts in this fixture.

If the delivery stores split membership in both the path and a physical column, check both declarations. In a separate four-case fixture, swapping two intact shards produces four conflicts. Under explicit Hive partitioning, both tested readers select the cases indicated by the paths. The physical-column-only scans use the stored split. Resolve conflicting declarations against the agreement before scoring.

Check labels by case ID when a rewrite is allowed

Compare the rewrite with an approved case-to-label mapping, using the scorer's decoded values. In a separate eight-row fictional example, a complete-row reorder and a two-label exchange both change the full file bytes and preserve identical serialized footer metadata. Exact comparison by ID finds zero label changes in the intact reorder and two in the exchange. The reorder passes the label check when row order may change. The label exchange fails it even though the footer and label totals match.

Record the delivery reference, scoring reader and filter, missing or unexpected IDs, and any changed case-to-label pairs. This gives the next reviewer the inputs needed to replay each acceptance decision.