[lemire]

A timestamp field reads ISO 8601 (RFC 3339) strings, such as the created_at of GH Archive events, into an Arrow timestamp column, and the writer writes timestamps back as ISO 8601 strings.

make_timestamp("created_at", time_unit::microsecond, "UTC")  // Arrow format "tsu:UTC"

• Types: type::timestamp, with a unit (seconds to nanoseconds) and a timezone (empty for none), imported from and exported to the Arrow formats tss:, tsm:, tsu: and tsn:, with or without a time zone.
• Reading: a date (2015-01-01), then optionally T (or t, or a space) and a time (15, 15:00, 15:00:00, or 15:00:00.123456 with . or ,), then optionally Z or an offset (+05:00, +0500, +05). A time without Z or an offset is UTC. An integer is a count of units. Anything else is an error: an impossible date (2015-02-29, 24:00), a fraction with more digits than the unit (as Arrow's JSON reader does: no silent loss of precision), or a value out of the range of the unit. The message names the value: invalid timestamp "2015-02-30" (row 0, field /t).
• Writing: "2015-01-01T15:00:00", with Z when the timestamp has a time zone, and the fraction (3, 6 or 9 digits, by unit) when it is not zero. The writer also writes date32 and date64 columns, as "2015-01-01".
• Not changed: schema inference never infers timestamps (strings stay utf8).

The parsing and formatting are in src/timestamp.h (the date arithmetic follows Howard Hinnant's days_from_civil), which the amalgamation now includes.

Results

From Python (fastsimdjson with this branch vendored), loads_arrow returns the same table as pyarrow.json.read_json for the four units, with and without a time zone, on GH Archive events and on 2,000 random timestamps per unit (offsets, fractions, the date alone), and both reject the same invalid ones.

GH Archive (886 MB), the gh schema with created_at as a string, then as timestamp[us, UTC], GB/s, Xeon Gold 6548N, best of 3:

	string	timestamp
bench_ndjson, GCC 16, one thread	1.986	1.982
bench_ndjson, clang 21, one thread	1.935	1.941
fastsimdjson loads_arrow (manylinux wheel, Python 3.14), one thread	2.11	2.11
the same, 32 threads	29.7	28.3
pyarrow.json.read_json, one thread	0.45	0.45
polars.read_ndjson, one thread	0.55	0.21

Reading the timestamps costs next to nothing; with them, loads_arrow is 4.7 times as fast as pyarrow and 10 times as fast as polars on one thread.

benchmarks/schemas.h has a gh_ts schema (bench_ndjson gharchive.ndjson gh_ts).

Tests

New tests: the parser (valid and invalid timestamps, offsets, fractions, leap years, the limits of nanoseconds), the formatter, 200,000 random round trips in every unit, reading (strings, integers, nulls, error messages), schema import and export, writing (with and without a time zone), and date32 and date64. They pass with Apple clang (also with ASan and UBSan), and with GCC 16 and clang 21 on Linux, without new warnings under -Wall -Wextra.