mostlyright.core.formats.parquet

Parquet format — Arrow-backed lossless serialization.

Uses pyarrow as the engine and zstd compression. Lossless for the schemas in this project (string, int, Int64 nullable, float, bool, tz-aware Timestamp[ns, tz], object-of-strings).

Documented exceptions:

pandas Categorical dtypes roundtrip as Categorical with pyarrow’s dictionary encoding, but the category ordered flag is preserved while the underlying dtype may rebuild on read.
pandas extension dtypes that pyarrow does not natively map (rare in the canonical schemas — see design doc §A) may degrade to their object representation.
Object columns holding mixed Python types are written using pyarrow’s best-effort type inference; columns mixing dicts/lists and primitives will raise at write time rather than silently coerce.

Decompression-bomb protection is NOT enforced in this layer — see the docstring on loads and design doc §E. The MCP server layer applies the per-payload uncompressed-size cap before invoking loads.

Functions

`dumps`(df)	Serialize a DataFrame to a parquet byte payload.
`loads`(data)	Parse a parquet byte payload back into a DataFrame.

mostlyright.core.formats.parquet.dumps(df)

Serialize a DataFrame to a parquet byte payload.

Engine: pyarrow. Compression: zstd. The full DataFrame schema — column names, dtypes (including nullable Int64, boolean, and tz-aware Timestamp), and the index unless the caller drops it — is preserved on the wire.

Return type: bytes
Parameters: df (DataFrame)

mostlyright.core.formats.parquet.loads(data)

Parse a parquet byte payload back into a DataFrame.

Engine: pyarrow. No decompression-bomb protection at this layer — the MCP server enforces a per-payload uncompressed-size cap (1 GB default per design doc §E) before invoking this function, and the SDK caller is trusted to pass parquet they themselves produced. Adding a size cap here would push policy into the format layer; it lives at the trust boundary instead.

Return type: DataFrame
Parameters: data (bytes)