Feature Request / Improvement
PyIceberg accepts table sort orders at write time but ignores them. Arrow table writes produce unsorted data files and hard-code sort_order_id=None, so manifests are inconsistent with the sort order the user declared.
Related umbrella issue: #271
Use case / motivation
Without a truthful sort_order_id and physically ordered data files, readers can't use sort-order-aware pruning and the manifest disagrees with the table's sort metadata. Users who declare a sort order expect their writes to honor it.
Proposed change
When every sort field uses an identity transform and null placement is consistent, sort materialized Arrow table writes. Unpartitioned tables are sorted before bin packing and each final partition is sorted independently. The table's sort-order ID is recorded on the data files.
Unsupported transforms, nested or missing fields, mixed null placement, and streaming RecordBatchReader writes keep the current behavior. Files are marked unsorted and sort_order_id stays null, with a warning explaining why.
Implementation
PR #3830 sorts the writes, carries the sort-order ID through WriteTask, writes the truthful DataFile.sort_order_id, and includes unit and integration coverage.
Tooling note
I developed this with assistance from DS v4 Pro and reviewed the changes myself.
References
Feature Request / Improvement
PyIceberg accepts table sort orders at write time but ignores them. Arrow table writes produce unsorted data files and hard-code
sort_order_id=None, so manifests are inconsistent with the sort order the user declared.Related umbrella issue: #271
Use case / motivation
Without a truthful sort_order_id and physically ordered data files, readers can't use sort-order-aware pruning and the manifest disagrees with the table's sort metadata. Users who declare a sort order expect their writes to honor it.
Proposed change
When every sort field uses an identity transform and null placement is consistent, sort materialized Arrow table writes. Unpartitioned tables are sorted before bin packing and each final partition is sorted independently. The table's sort-order ID is recorded on the data files.
Unsupported transforms, nested or missing fields, mixed null placement, and streaming RecordBatchReader writes keep the current behavior. Files are marked unsorted and sort_order_id stays null, with a warning explaining why.
Implementation
PR #3830 sorts the writes, carries the sort-order ID through WriteTask, writes the truthful DataFile.sort_order_id, and includes unit and integration coverage.
Tooling note
I developed this with assistance from DS v4 Pro and reviewed the changes myself.
References