Bug: dynamic_partition_overwrite silently skips spec-0 manifests after partition spec evolution

Bắt đầu trong table/init.py tại Table.dynamic_partition_overwrite, sau đó lần theo _build_partition_projection và thay đổi manifest-pruning từ #3011. Tái hiện snapshot spec-0/spec-1 hỗn hợp được mô tả trong issue và so sánh với fix liên quan trước đó trong #1108. Công việc được xem là hoàn tất khi một regression test xác nhận rằng thao tác ghi đè loại bỏ các hàng khớp khỏi cả hai spec lịch sử mà không xóa dữ liệu không liên quan.

Summary

dynamic_partition_overwrite produces incorrect results when a table has undergone
partition spec evolution. Manifests written under older specs are silently skipped
by the manifest pruning logic introduced in #3011, leaving stale data files that
should have been deleted.

Root cause

In Table.dynamic_partition_overwrite (table/__init__.py), the delete predicate
is built using only the current partition spec:

delete_filter = self._build_partition_predicate(
    partition_records=partitions_to_overwrite,
    spec=self.table_metadata.spec(),       # always current spec
    schema=self.table_metadata.schema()
)

A snapshot with mixed partition_spec_ids (spec-0 and spec-1 manifests) passes
this single predicate to _DeleteFiles. The manifest evaluator in _build_partition_projection
uses inclusive_projection(schema, spec) — when projecting a spec-1 predicate
(e.g. category=A AND region=us) through spec-0 (which only has category), the
region reference has no corresponding partition field, causing the evaluator to
incorrectly skip spec-0 manifests entirely.

Reproduction

import tempfile, pyarrow as pa
from pyiceberg.catalog import load_catalog
from pyiceberg.schema import Schema
from pyiceberg.types import NestedField, StringType, LongType
from pyiceberg.partitioning import PartitionSpec, PartitionField
from pyiceberg.transforms import IdentityTransform

schema = Schema(
    NestedField(1, "category", StringType(), required=False),
    NestedField(2, "region",   StringType(), required=False),
    NestedField(3, "value",    LongType(),   required=False),
)
spec_v0 = PartitionSpec(
    PartitionField(source_id=1, field_id=1000, transform=IdentityTransform(), name="category")
)

with tempfile.TemporaryDirectory() as warehouse:
    catalog = load_catalog("test", **{"type": "sql", "uri": f"sqlite:///{warehouse}/catalog.db", "warehouse": f"file://{warehouse}"})
    catalog.create_namespace("default")
    table = catalog.create_table("default.test", schema=schema, partition_spec=spec_v0)

    # Write under spec 0
    table.append(pa.table({"category": ["A","A","B"], "region": [None,None,None], "value": [1,2,10]}))

    # Evolve spec
    with table.update_spec() as u:
        u.add_field("region", IdentityTransform(), "region")
    table = catalog.load_table("default.test")

    # Write under spec 1
    table.append(pa.table({"category": ["A","B"], "region": ["us","us"], "value": [100,200]}))

    # Overwrite category=A — should delete ALL A rows (both specs)
    table.dynamic_partition_overwrite(
        pa.table({"category": ["A"], "region": ["us"], "value": [999]})
    )

    result = table.scan().to_arrow().to_pydict()
    a_values = [v for c,v in zip(result["category"], result["value"]) if c == "A"]
    print(a_values)  # BUG: prints [1, 2, 100, 999] — stale rows from spec-0 not deleted
                     # EXPECTED: [999]

Fix

Build the delete predicate per historical spec present in the snapshot, projecting
the new data files' partition values into each spec's coordinate space before evaluating.
PR with fix and regression tests to follow.

Related

• #3011 (introduced the manifest pruning optimization)
• #1108 (prior related fix by @Fokko for spec evolution in manifest rewriting)