[SteNicholas]

Purpose

Linked issue: close #477

Paimon Java accepts TIME columns in Lance data files
(LanceFileFormat.LanceRowTypeVisitor.visit(TimeType)) and writes them as Arrow TIMEMILLI
(time32[ms]) through ArrowFieldTypeConversion. Paimon C++ has represented TIME as
time32[ms] since #378, but rejected it for Lance:

• SchemaValidation::ValidateLanceDataField() had no TIME32 case, so a TIME column, at top
  level or nested in ARRAY or ROW, failed with Lance file format does not support type time32[ms].
• LanceStatsExtractor::CreateEmptyStats() had no TIME32 case either, so lifting only the
  validation would still fail to finalize an append data file with a TIME column, with
  cannot create empty Lance statistics for type time32[ms].

This change:

• Accepts TIME32 in ValidateLanceDataField(). The existing recursion covers TIME nested in
  ARRAY and ROW, and TableSchema::Create() still rejects non-millisecond units.
• Returns empty integer statistics for TIME32 in LanceStatsExtractor::CreateEmptyStats(),
  as VortexStatsExtractor does.
• Needs no reader change: the declared precision comes from the paimon.time.precision field
  metadata of the table schema, and LanceFileBatchReader aligns each batch to the read type.

Tests

• SchemaValidationTest.TestLanceDataTypes: millisecond TIME at top level and in
  ROW<ARRAY<TIME>> moves from the rejected to the accepted fields; second-unit TIME, top level
  and nested, is still rejected.
• LanceFileFormatTest.WriteThenReadSupportedTypes: adds a top-level TIME column and
  ARRAY<TIME>, ROW<TIME> and ROW<ARRAY<TIME>> columns declared with precisions 0, 3, 6 and
  9. The data covers 0 and 86399999, null and empty arrays, null elements and null ROW children,
  and is read with batch sizes 1, 2 and 4. A batch imported directly from the reader checks that
  the returned Arrow type keeps the read schema's field metadata, including
  paimon.time.precision.
• LanceFileFormatTest.ExtractStatistics: a TIME column yields an empty statistics entry.
• TimeColumnWriteInteTest.TestAppendTimeColumnFinalize, generalized from
  VortexTimeColumnWriteInteTest: append write, finalize and read of a TIME column for both
  Lance and Vortex.
• ScanAndReadInteTest.TestLanceJavaCompatibility and TestLancePythonCompatibility: the
  lance/append_java_compat and lance/append_python_compat fixtures are regenerated with
  Paimon Java and Paimon Python at the same commit as before (63ffd40), adding
  f_time_0 TIME(0) and f_time_3 TIME(3) columns with 00:00:00, 23:59:59.999, 12:34:56.789
  and NULL values. All other columns, values, options and commits are unchanged. Both writers
  store TIME in Lance as time32[ms] without precision metadata, so C++ reads the declared
  precision from the table schema. The fixture READMEs list the new columns and tie every
  writer-specific statement to the writer commit in their header.

API and Format

No API changes. No storage format changes: TIME is written to Lance as time32[ms], as in
Paimon Java.

Documentation

Yes. docs/source/user_guide/data_types.rst no longer lists TIME as not supported. It describes
the time32[ms] mapping, how the declared precision is kept in paimon.time.precision while
values keep millisecond precision, and which file formats read or write TIME. A note states its
limitations: only append-only tables support TIME, primary-key tables support it neither as a
primary key nor as a value column, a TIME column cannot be a partition or bucket key, and
predicates cannot reference TIME columns.

Generative AI tooling

Generated-by: Claude Code 2.1.311 (Claude Opus 5.5)

🤖 Generated with Claude Code

[lxy-9602]

+1