Add Parquet support to view, less, and a new parquet2tab

[mbreese]

Adds a parquet2tab subcommand (mirroring csv2tab) and teaches view and less to read Parquet directly.

Parquet is detected from its PAR1 magic bytes, in the same spirit as the existing gzip sniffing, so no flag is needed. An explicit --parquet flag is also available, mirroring --csv.

$ tabl view nested.parquet
id    | name    | price    | day        | tags      | addr.city    | addr.zip    | attrs
======+=========+==========+============+===========+==============+=============+===========
1     | alpha   | 123.45   | 2024-01-01 | ["x","y"] | Boston       | 02115       | {"a":"1"}
2     |         | -0.50    | 1970-01-01 |           |              |             |
3     | gamma   | 1.00     | 2024-10-04 | ["solo"]  | St. Louis    | 63110       | {"b":"2"}

Approach

Readers now sit behind a RecordReader interface. The viewer, pager, and CSV exporter only ever needed ReadLine/Close/header access, so they consume the interface instead of *DelimitedTextFile. exporter.go and sorter.go reach further into the text internals and stay on the concrete type — DelimitedTextFile keeps its exported Header field, so both are untouched.

Converting CSVExporter is what lets parquet2tab reuse it wholesale, including the existing tab-escaping, rather than duplicating a second writer.

Type mapping

This is the substance of the change. parquet.Value.String() renders the physical type only — it would print a DATE as a day count and a DECIMAL(10,2) as its raw unscaled integer. ParquetFile builds a per-column formatter from the logical type instead:

Logical type	Rendering
Decimal(p,s)	scale applied via big.Int (handles int32/int64/FLBA two's complement)
Date	days since epoch → 2024-01-01
Time / Timestamp	unit-aware; RFC3339, unzoned when not adjusted to UTC
Integer{BitWidth,IsSigned}	unsigned values don't wrap negative
UUID, INT96	decoded to canonical / RFC3339 form

Schema handling: nested structs flatten to dotted names (addr.city); lists and maps render as JSON in one cell; NULLs are empty by default, configurable with --na=STRING.

Pipe limitation

Parquet keeps its schema in a footer at the end of the file, so reading one requires random access. It cannot be read from a pipe — - is rejected with an explicit message rather than failing obscurely. Flags meaningless for Parquet (--no-header, --header-comment, --show-comments, --csv) are rejected when actually set.

Verification

• make test green; six new tests lock down the formatting that's easy to get wrong (123.45 not 12345, 2024-01-01 not 19723, 4000000000 not negative)
• 150-row round-trip against examples/iris.txt is identical
• All error paths checked: pipe rejection, missing file, flag conflicts
• CGO_ENABLED=0 cross-compiles for linux/macos/windows

Dependency note: parquet-go bumps mattn/go-runewidth v0.0.2 → v0.0.15, which underpins termbox's width calculations. Pager output was diffed before and after, including CJK text, and is byte-identical — no pin needed.

Binary size: 6.5 MB → 15.2 MB (brotli, klauspost/compress, lz4, protobuf). Worth a look if that matters for releases; -ldflags="-s -w" would claw some back.

Known limitation: an empty list and a NULL list both render as the NA string. Distinguishing them needs definition-level inspection, left out of scope.

🤖 Generated with Claude Code