Add support for timestamp (and duration) types in leaf data hashing

[kurodo3]

Summary

ArrowDigester.hash_table() raises NotImplementedError: Unsupported leaf type: timestamp[us, tz=UTC] when a table contains any timestamp column. This blocks using datetime.datetime values as data or tag values in projects that rely on starfix for content-addressable hashing.

Reproduction

import pyarrow as pa
from starfix import ArrowDigester
from datetime import datetime, timezone

dt = datetime(2024, 1, 15, 12, 0, 0, tzinfo=timezone.utc)
table = pa.table({"ts": pa.array([dt], type=pa.timestamp("us", tz="UTC"))})
ArrowDigester.hash_table(table)
# NotImplementedError: Unsupported leaf type: timestamp[us, tz=UTC]

Also fails for naive timestamps:

table = pa.table({"ts": pa.array([dt], type=pa.timestamp("us"))})
ArrowDigester.hash_table(table)
# NotImplementedError: Unsupported leaf type: timestamp[us]

Root cause

In arrow_digester.py, _element_size_for_type includes fixed-size types like date32 (4 bytes), date64 (8 bytes), time32 (4 bytes), and time64 (8 bytes), but is missing timestamp variants:

_sizes = {
    pa.int8(): 1, pa.uint8(): 1,
    pa.int16(): 2, pa.uint16(): 2, pa.float16(): 2,
    pa.int32(): 4, pa.uint32(): 4, pa.float32(): 4, pa.date32(): 4,
    pa.int64(): 8, pa.uint64(): 8, pa.float64(): 8, pa.date64(): 8,
}
# ← missing: timestamp (all units and timezones)
# ← missing: duration (all units)

Note that _data_type_to_value (used for schema hashing) already handles timestamp correctly — only the leaf data hashing step is missing.

Arrow storage layout

All timestamp variants and duration variants are stored as signed 64-bit integers in Arrow's physical layout (8 bytes per element). The unit (s, ms, us, ns) and timezone are schema-level metadata, not part of the data buffer. This means the fix is straightforward: treat timestamps and durations as fixed-size 8-byte values, just like int64.

Proposed fix

In _element_size_for_type, add handlers after the existing time32/time64 checks:

if pa.types.is_timestamp(dt):
    return 8  # int64 physical storage, unit/tz in schema metadata

if pa.types.is_duration(dt):
    return 8  # int64 physical storage, unit in schema metadata

Expected behavior

• timestamp[us, tz=UTC] → hashed using 8-byte fixed-size data hashing (same as int64)
• timestamp[us] → same
• timestamp[s], timestamp[ms], timestamp[ns] → same
• duration[us] etc. → same

Impact

Affects any use case where datetime (Python) values appear in Arrow tables — a very common pattern for tagging data with experiment dates, timestamps, dates of birth, etc. For example, orcapod (which depends on starfix) cannot hash tag values that are datetime.datetime objects as a result of this gap.

Version

starfix==0.3.1