roboto.formats.parquet

ParquetParser.extract_timestamp_info()

ParquetParser.find_timestamp_field_by_type()

ParquetParser.get_data_for_column()

ParquetParser.get_timestamp_field_by_name()

ParquetParser.is_parquet_file()

ParquetParser.requires_rewrite()

ParquetParser.rewrite()

The timestamp field of a Parquet file, the field holding each row’s timestamp. Serves as both a descriptor of that field and as a utility for projecting its values to other time units.

Timestamp.to_epoch_nanoseconds()

Timestamp.unit()

Compute a boolean mask indicating which rows fall within the specified time range. Returns None if no time filtering is needed (both start_time and end_time are None).

Find a Parquet file’s timestamp field and describe it as a Timestamp.

The field is found by walking timestamp_field.path_in_schema one component at a time, first among the schema’s top-level fields and then through struct children, so a nested field and a top-level column whose name contains dots are never confused. schema must be the file’s ParquetFile.schema_arrow: the timestamp’s column index is counted over the schema’s leaves, which match the file’s leaf columns one for one.

unit_hint is the unit of the stored values, used when the field’s Arrow type does not carry one (an integer, floating-point or decimal field). Callers take it from the Unit metadata of the timestamp’s MessagePathRecord, or from unit.

Extract timestamps in nanoseconds since Unix epoch from the table’s timestamp field.

The field is found by walking timestamp.path through the table’s struct columns. A row whose enclosing struct is null has a null timestamp.

Generate AddMessagePathRequest objects for all fields in a Parquet schema.

Traverses the schema recursively to generate message paths for nested types (structs, lists) in addition to top-level fields.

Usage

For a schema with a struct column `position: struct<x: float, y: float>`: - Yields position (Object) - Yields position.x (Number) - Yields position.y (Number)

For a schema with `values: list<float64>`: - Yields values (NumberArray)

For a schema with `points: list<struct<x: float, y: float>>`: - Yields points (Array) - Yields points.x (Number) - Yields points.y (Number)

Prune list-of-struct columns to the projected leaves inside each element.

PyArrow’s prefix-based nested column selection cannot reach through list wrapper nodes, so resolve_columns() reads a list-nested leaf’s whole list ancestor column — every element keeps all of its struct fields. This Arrow-native post-read pass narrows each such element down to the requested leaves, leaving every other read path byte-identical.

A top-level root is narrowed iff at least one of its projected paths has a list ancestor; otherwise the table is returned unchanged (pure struct, scalar, and scalar-list reads never enter the rebuild). Per root, a trie is built from its paths with the root component stripped so non-list-nested siblings the projection also keeps are preserved. Every struct keeps its fields in the order the file stores them.

No field of fields may lie inside another: the outer one would be narrowed to the inner one rather than kept whole.

Open a remote Parquet file under a cache policy, from the cheapest available source.

Dispatches on choose_fetch_mode(): an already-cached copy is reused, a download is performed (concurrency-safe, atomic) when the policy calls for one, and otherwise the file is streamed over HTTP range requests without touching disk.

Open a Parquet file over HTTP via a signed URL (no local download).

A single ranged GET over fsspec’s shared HTTP session probes the first _STREAM_WHOLE_FILE_PROBE_BYTES of the file. A file smaller than the probe arrives whole in that one request and is read from an in-memory buffer; a larger file (or a failed probe) falls back to HTTP range-request streaming through pyarrow’s filesystem layer.

When size_bytes is known and at least _STREAM_WHOLE_FILE_PROBE_BYTES, the file is known-large up front: the whole-file probe could never win (a BufferReader read is taken only for sub-threshold files), so it is skipped and the file is range-streamed directly, avoiding a wasted 16 MiB GET. size_bytes of None (an older server omits the size) preserves the probe-then-decide behavior.

Build a deduplicated list of column names safe for read_row_group(columns=...).

Children of list-type columns are replaced by their list ancestor’s column name because PyArrow’s prefix-based nested column selection does not work through list wrapper nodes in the physical Parquet schema. Selecting the parent list column already returns its full nested structure.

This is important because the projected fields contain only leaf paths. For a column like points: list<struct<x, y>>, only points.x and points.y are selected — the parent points field is absent. This function derives the correct parent column name from the child’s path_in_schema.

Children of struct-type columns are preserved because PyArrow can resolve them via dot-separated prefix matching (e.g. "position.x" selects the x child of the position struct).

Each name joins path components with dots, the form in which read_row_group takes columns. Two fields whose paths join to the same name are both selected by it: "header.stamp" selects both a stamp field nested in a header struct and a top-level column of that name. select_fields() removes the fields that were not projected.

Keep only the fields fields names, found by walking their path components, in the table’s order.

Reading columns by their dot-joined names (resolve_columns()) can bring in more than was projected, such as a top-level column named header.stamp read along with a header struct’s stamp field, or a timestamp field read only to filter rows. This removes them:

• A top-level column that no field’s path starts with is dropped.
• A field that names a column or struct keeps it whole, even when other fields name some of its children.
• Otherwise a struct keeps only the children that some field’s path runs through, with its own null rows.
• Lists and maps, and everything below them, are kept as read (narrow_list_nested_fields() narrows the structs inside a list).
• A field whose path the table lacks is skipped.

A column from which nothing is removed is returned as read. With no fields at all, the result has no columns and keeps the table’s row count.

Return whether narrow_list_nested_fields() would change the table.

True iff at least one projected field addresses a leaf inside a list (its path has a list ancestor). When False, every projected field resolves through structs and scalars alone, so PyArrow’s column selection already returns the narrowed shape and the post-read prune is a no-op — callers can skip it.

Cheap enough to evaluate once per file and hoist the per-row-group narrowing decision out of the decode loop.

Determine whether a Parquet row group contains data within the requested time range. Used to short-circuit requesting column chunks from the given row group if not relevant.

The statistics of the timestamp’s column chunk in a row group, or None when there are none to use.

A nested field and a top-level column whose name contains dots can share a column chunk’s path_in_schema (header.stamp names both a header struct’s stamp field and a column of that name), so the chunk is found by its position, timestamp.column_index. Returns None when the row group has no chunk at that index, when that chunk’s path_in_schema is not the timestamp’s path joined with dots (timestamp was found in a schema other than this file’s), or when the chunk has no statistics.