yggdrasil.parquet.parquet_file

parquet_file

Parquet Tabular leaf over the new :class:IO substrate.

:class:ParquetFile is an :class:IO subclass that auto-registers under :data:MimeTypes.PARQUET. The Parquet file format is footer-indexed: readers parse the metadata block at the end of the file once and use it to plan column reads. Writes buffer row groups and finalize the footer on close.

The reworked memory-management model lets us pass self (or a :meth:view) directly to :class:pyarrow.parquet.ParquetWriter and :class:pyarrow.parquet.ParquetFile, so reads and writes don't have to bounce through a :class:pyarrow.BufferReader / :class:pyarrow.BufferOutputStream adapter.

Native engine dispatch

When the holder is a local path, :meth:_read_arrow_dataset / :meth:_scan_polars_frame / :meth:_read_polars_frame short-circuit to the format-aware scanners (pds.dataset(format="parquet"), pl.scan_parquet, pl.read_parquet) — those push projection and predicate filters into the Parquet reader at plan time, which the generic Arrow batch shim can't do.

ParquetOptions dataclass

ParquetOptions(
    source: "Field | None" = None,
    target: "Field | None" = None,
    safe: bool = False,
    checked_cast: bool = False,
    mode: Mode = Mode.AUTO,
    schema_mode: Mode = Mode.AUTO,
    row_size: int | None = None,
    byte_size: int | None = None,
    row_limit: int | None = None,
    use_threads: bool = True,
    match_by: list["Field"] | None = None,
    unique_by: list["Field"] | None = None,
    time_sample_by: list["Field"] | None = None,
    fill_strategy: str = "ffill",
    predicate: Predicate | None = None,
    wait: WaitingConfig = WaitingConfig.default(),
    spark_session: "SparkSession | None" = None,
    arrow_memory_pool: MemoryPool | None = None,
    update_column_names: list[str] | None = None,
    zorder_by: list[str] | None = None,
    optimize_after_merge: bool = False,
    vacuum_hours: int | None = None,
    retry: WaitingConfigArg | None = None,
    return_data: bool = False,
    safe_merge: bool = False,
    sync_metadata: bool = True,
    compression: "str | None" = "snappy",
    compression_level: "int | None" = None,
    use_dictionary: bool = True,
    write_statistics: bool = True,
    row_group_size: "int | None" = None,
)

Bases: CastOptions

:class:CastOptions extended with Parquet-specific knobs.

merged property

merged: Field | None

Target reconciled with source under :attr:schema_mode.

With the default :attr:Mode.AUTO, a target field that matches a source field by name is merged — so variant (ObjectType / NullType) target slots adopt the source dtype — while the target's field set is preserved (source-only columns are not pulled in). This is the schema casts coerce to: see the cast_* dispatchers, which run against merged rather than the raw target so a bare columns= projection autotypes against the bound source.

column_names property

column_names: list[str] | None

The target field's column names, if a target field is bound.

match_by_keys property

match_by_keys: list[str] | None

Resolved key column names to dedup on.

Pulls the :attr:Field.name of each entry in :attr:match_by. Returns None when no keys are set so callers can branch on "keys vs no-keys" with a single truthiness check.

select_source_column_names

select_source_column_names() -> list[str] | None

The source field's column names, if a source field is bound.

read_columns

read_columns() -> list[str] | None

Columns a source reader must keep — the projection plus the predicate's columns.

:attr:column_names is what the read should end up with, but the predicate row-filter runs before the cast projects down to it, so any column the predicate touches has to survive the read even when the caller didn't ask for it (columns=["a"] + predicate on b). None means no projection — read everything.

check classmethod

check(options: CastOptionsArg = None, /, **overrides: Any) -> T

Canonical entry point — coerce anything into a :class:CastOptions.

Dispatch by what options is:

• None — construct fresh :class:CastOptions(**overrides).
• :class:CastOptions — if no overrides given, return it unchanged; if overrides given, .copy(**overrides).
• :class:Mapping (including dict) — merge into overrides and construct fresh (override args win on key collision).
• :class:pa.DataType / :class:pa.Field / :class:pa.Schema / :class:Field / :class:Schema — treat as a target hint. Equivalent to check(target=options, **overrides).

source= / target= go straight into the dataclass slots (after :meth:Field.from_ normalization in __post_init__). Callers that want the peek-and-bind "only set if not already bound" semantic should chain :meth:check_source / :meth:check_target after the call.

columns= shortcut: a sequence of column names describing the desired output projection. When a target is already bound the names narrow it (target.select(columns)); otherwise they are promoted to a struct-shaped target field whose children default to :class:ObjectType — a "I want these columns, leave their types alone" placeholder that drives projection without casting. It lands on target (not source) so it never shadows the real source schema inferred at read time.

:raises TypeError: if options is a type the dispatch table doesn't cover.

field_names cached classmethod

field_names() -> frozenset[str]

Frozenset of this class's constructor-accepting field names.

Used by :meth:_build to filter **overrides down to keys the constructor will accept — callers funnel mixed kwargs through .check() (DataIO public methods often pass user kwargs straight through), and we don't want a stray filter= or columns= to crash construction.

Excludes init=False fields (the private memoization slots for merged / merged_schema); those are not valid __init__ keywords and a copy via dataclasses.replace would crash if it tried to forward them.

Cached per-class via :func:functools.cache so subclasses with extra fields get their own expanded set on first access.

copy

copy(**overrides: Any) -> T

Return a copy with overrides applied.

... values in overrides are ignored (keep existing). Pass source=/target= to swap either slot — :class:Field normalization runs in __post_init__ so any :class:Field-shaped input (pa.Schema, pa.DataType, dict, …) is accepted.

Implementation note: bypasses :func:dataclasses.replace, which rebuilds via cls(**all_fields) and pays a full __init__ + __post_init__ traversal even when the caller only tweaked a single bool. Cast pipelines call copy repeatedly per batched write (with_source / check_source / with_target all funnel through here), so the fast clone below — :func:object.__new__ + slot copy + targeted __post_init__ normalization for the overridden keys — is meaningfully cheaper.

check_source

check_source(obj: Any = None, *, copy: bool = True) -> T

Bind a :attr:source if one isn't already set.

Two ways to supply one:

1. source= on :meth:check / :meth:copy — explicit Field / Schema / pa type. Wins even if self.source is already set (explicit override).
2. obj= here — a peekable object. Only runs the peek when self.source is currently None — an already- bound field is never clobbered by a peek.

Returns self unchanged when neither is given. Used from :class:DataIO methods (collect_schema, read_arrow_dataset) that want to pin a source schema before running a batch walk.

checked_cast=True short-circuits — the caller guarantees the batch shape matches the target, so the peek (which would rebuild a yggdrasil :class:Field from the batch's :class:pa.Schema) is wasted work. Combined with the :meth:cast_arrow_tabular short-circuit, this collapses every per-batch cast pass to a single attribute read on the leaf write path — ~150 us / batch saved on a RESPONSE_SCHEMA-shaped write.

check_target

check_target(obj: Any = ..., *, copy: bool = True) -> T

Bind a :attr:target if one isn't already set.

Symmetry partner for :meth:check_source. See that method for the argument semantics — source/target behave identically.

with_source

with_source(source: 'Field', copy: bool = False) -> T

Return a copy with source as the new source field.

Accepts the same shapes :meth:Field.from_ does (pa schema, yggdrasil Field, dict, etc.) — normalized in __post_init__ via :func:dataclasses.replace. The frozen slot is updated through :func:object.__setattr__ in the post-init hook; we don't bypass it here because going through replace gets the normalization for free.

with_target

with_target(target: 'Field', copy: bool = True) -> T

Return a copy with target as the new target field.

with_checked_cast

with_checked_cast(value: bool = True, copy: bool = False) -> T

Return a copy (or in-place) with :attr:checked_cast set.

Mirror of :meth:with_source / :meth:with_target — keeps the per-call mutation behind a named method instead of having every writer-side caller reach for :func:dataclasses.replace / :func:object.__setattr__. Set when the caller knows every batch already matches the target (came from a :class:pa.Table, a :class:pa.RecordBatchReader, a polars / pandas frame, or another writer that just emitted the same schema); the leaf's :meth:check_source / :meth:cast_arrow_tabular then short-circuit straight to the write path.

need_cast

need_cast(
    source: Any | None = None,
    target: Any | None = None,
    check_names=False,
    check_dtypes=True,
    check_metadata=False,
    check_nullable: bool = False,
) -> bool

Return True if source and target fields differ enough to need casting.

When either field is unbound, returns False — there's nothing to compare against, so assume caller already sorted it.

Field equality semantics are the :meth:Field.equals rules: names, dtypes, metadata — each independently gateable. Metadata is off by default because it's commonly decorative (pandas preserves indices through metadata, arrow carries codec hints in field metadata) and comparing on it would demand a cast for cosmetic differences.

check_nullable is off by default because nullability rarely warrants a real value-level cast — primitives and lists pass through unchanged when only the flag differs. Tabular / struct casts pass check_nullable=True so the rebuild fires when child fields differ on nullability: Spark / Delta refuse to implicitly cast nullable→NOT NULL inside a struct (even when the data is in fact non-null), so the cast has to emit the target's field types verbatim to keep MERGE happy.

finalize

finalize(obj: Any, *, default_scalar: Any = None) -> Any

Finalize any object — delegates to :meth:Field.finalize.

finalize_spark_cast

finalize_spark_cast(obj: Any, *, default_scalar: Any = None) -> Any

Fill nulls and alias a Spark Column to the target name.

Direct parallel of :meth:finalize_polars_cast — Spark Columns, like polars Series/Expr, carry a name that can diverge from the target after a cast chain, so the alias step belongs in finalize rather than in each cast site.

finalize_arrow_cast

finalize_arrow_cast(obj: Any, *, default_scalar: Any = None) -> Any

Fill nulls on a pyarrow object to finish a cast chain.

No alias step: :class:pa.Array / :class:pa.ChunkedArray don't carry a name, and tabular rename (Table/RecordBatch) is a schema-level rebuild that :meth:cast_arrow_tabular already handles inline via the target schema. Finalize here just means "apply the default-scalar null fill."

finalize_pandas_cast

finalize_pandas_cast(obj: Any, *, default_scalar: Any = None) -> Any

Fill nulls on a pandas object to finish a cast chain.

No alias step exposed on :class:CastOptions for pandas — Series .name and DataFrame column labels get set by the cast methods directly. Finalize is fill-only, matching :meth:finalize_arrow_cast.

cast

cast(obj: Any) -> Any

Cast obj to :attr:target using its native engine.

Dispatches arrow types through :meth:cast_arrow, Tabular through :meth:cast_tabular. Everything else delegates to :meth:Field.cast.

cast_pyarrow

cast_pyarrow(obj: Any) -> Any

Cast any pyarrow object — delegates to :meth:Field.cast_arrow.

cast_arrow_array

cast_arrow_array(array: Any) -> Any

Cast a :class:pa.Array or :class:pa.ChunkedArray.

cast_arrow_batch

cast_arrow_batch(batch: 'pa.RecordBatch') -> 'pa.RecordBatch'

Filter + cast a :class:pa.RecordBatch.

cast_arrow_table

cast_arrow_table(table: 'pa.Table') -> 'pa.Table'

Filter + cast a :class:pa.Table.

cast_arrow_tabular

cast_arrow_tabular(data: 'ArrowTabular') -> 'ArrowTabular'

Filter + cast an :class:ArrowTabular (batch by batch).

cast_arrow

cast_arrow(data: Any) -> Any

Dispatch arrow types to the specific cast method.

cast_tabular

cast_tabular(data: Any) -> Any

Cast any Tabular-like object.

dedup_columns_on_read

dedup_columns_on_read() -> 'list[str]'

Return the column names that need client-side dedup at read time.

Sourced from :attr:unique_by — each Field's :attr:Field.name is the column the read pass must deduplicate on. Returns an empty list when :attr:unique_by is unset / empty.

dedup_arrow_batches

dedup_arrow_batches(
    batches: "Iterator[pa.RecordBatch]",
) -> "Iterator[pa.RecordBatch]"

Collapse duplicate rows on the columns flagged unique.

Resolves the dedup column set via :meth:dedup_columns_on_read, then delegates to :func:yggdrasil.arrow.ops.dedup_arrow_batches for the pure-Arrow group-by + take pass. Identity short-circuit when no column needs collapsing keeps the read path zero-cost on the common case (no target / no unique column / source already unique).

resample_on_read

resample_on_read() -> 'tuple[str, int, list[str], str] | None'

Return (time_column, sampling_seconds, partition_by, fill_strategy) to resample.

Picks the first entry of :attr:time_sample_by whose time_sampling metadata carries a positive ISO-8601 duration. The result drives :func:yggdrasil.arrow.ops.resample_arrow_table — a single (column, interval) is all that op consumes (you can only have one time axis per table to resample on at a time).

Each Field's sampling lives under its :attr:Field.metadata's non-prefixed b"time_sampling" key as an ISO-8601 duration string ("PT1H" / "P1D"). The non-prefixed key keeps the value off the schema-level tag registry (it's a per-call option, not a contract that rides with the data on disk).

partition_by is derived from the target schema's :attr:Field.primary_key set, minus the resample column itself if it's also primary. The rationale: on a per-entity time series (one symbol per row, partitioned by symbol), each entity's timeline should bucket independently — without partition_by the resample would collapse rows across instruments. Schemas with no primary keys (or where the only primary is the timestamp) fall back to a flat resample.

Returns None when :attr:time_sample_by is unset / empty or every listed Field's metadata fails to parse.

resample_arrow_batches

resample_arrow_batches(
    batches: "Iterator[pa.RecordBatch]",
) -> "Iterator[pa.RecordBatch]"

Snap rows to the target's time_sampling grid.

Resolves the resample column / interval / partition keys via :meth:resample_on_read, then delegates to :func:yggdrasil.arrow.ops.resample_arrow_batches. Identity short-circuit when no field is flagged keeps the read path zero-cost on the common case.

apply_post_read_table

apply_post_read_table(table: 'pa.Table') -> 'pa.Table'

Run column projection + resample + dedup on a materialised :class:pa.Table.

Same operations and same order as the streaming wraps — column projection first (trim I/O cost before any compute), resample second (its bucket collapse trims rows before the unique-tag walk), then dedup. Identity short-circuit when no pass is configured so the common case stays zero-cost.

Pyarrow / polars / pandas read paths that already produce a Table funnel through this method instead of the iterator wraps; the result is one Table.from_batches + one Table.take (per pass) instead of two Table.from_batches + a Table.to_batches rebatch sandwich.

apply_post_read_spark_frame

apply_post_read_spark_frame(df: Any) -> Any

Run resample + dedup directly on a Spark DataFrame.

Spark-side mirror of :meth:apply_post_read_table — same op order (resample first, dedup second), same identity short-circuit when neither is configured. Routes through :mod:yggdrasil.spark.ops so the heavy lifting stays on the executors (groupBy + applyInArrow for the partitioned resample, SQL window functions otherwise) instead of collecting the frame to the driver as Arrow.

Used by :meth:yggdrasil.spark.tabular.Dataset._read_spark_frame to apply the read-time passes before handing the frame back — saving a full df.toArrow → arrow.ops → createDataFrame round trip per configured op.

cast_arrow_batch_iterator

cast_arrow_batch_iterator(batches: Any) -> Any

Cast a stream of :class:pa.RecordBatch and rechunk by byte_size / row_size.

With a bound target: per-batch tabular cast + streamed rechunking via :meth:Field.cast_arrow_batch_iterator (which routes through the struct-side helper).

Without a target: rechunk-only when byte_size / row_size is set, otherwise passthrough. Lets callers that did an in-engine cast upstream still pick up the optimized rechunker.

fill_arrow_nulls

fill_arrow_nulls(obj: Any, *, default_scalar: Any = None) -> Any

Engine-level null fill — delegates to :meth:Field.fill_arrow.

fill_arrow_array_nulls

fill_arrow_array_nulls(array: Any, *, default_scalar: Any = None) -> Any

Narrow null fill for a :class:pa.Array / :class:pa.ChunkedArray.

cast_polars

cast_polars(obj: Any) -> Any

Cast any polars object — delegates to :meth:Field.cast_polars.

cast_polars_series

cast_polars_series(series: Any, *, default_scalar: Any = None) -> Any

Cast a :class:pl.Series.

cast_polars_expr

cast_polars_expr(expr: Any, *, default_scalar: Any = None) -> Any

Cast a :class:pl.Expr.

Wraps the expression tree with a cast operator — actual work fires when the containing LazyFrame is collected.

cast_polars_tabular

cast_polars_tabular(data: Any) -> Any

Cast a :class:pl.DataFrame or :class:pl.LazyFrame.

fill_polars_nulls

fill_polars_nulls(obj: Any, *, default_scalar: Any = None) -> Any

Engine-level polars null fill — delegates to :meth:Field.fill_polars.

polars_alias

polars_alias(obj: Any) -> Any

Rename a polars Series/Expr to the target name — no-op if matching.

Delegates to :meth:Field.polars_alias. When target is unbound there's no name to rename to, so we pass through.

cast_pandas

cast_pandas(obj: Any) -> Any

Cast any pandas object — delegates to :meth:Field.cast_pandas.

fill_pandas_nulls

fill_pandas_nulls(obj: Any, *, default_scalar: Any = None) -> Any

Engine-level pandas null fill — delegates to :meth:Field.fill_pandas.

cast_spark_frame

cast_spark_frame(df: Any) -> Any

Filter + cast a Spark DataFrame.

Applies :attr:predicate (when set) then delegates to :meth:Field.cast_spark_tabular for schema coercion.

cast_spark_tabular

cast_spark_tabular(data: Any) -> Any

Filter + cast a :class:Dataset (Spark Tabular wrapper).

cast_spark_column

cast_spark_column(obj: Any) -> Any

Cast a Spark Column.

cast_spark

cast_spark(obj: Any) -> Any

Dispatch spark types to the specific cast method.

fill_spark_nulls

fill_spark_nulls(obj: Any, *, default_scalar: Any = None) -> Any

Engine-level spark null fill — delegates to :meth:Field.fill_spark.

spark_alias

spark_alias(obj: Any) -> Any

Rename a Spark Column to the target name — delegates to :meth:Field.spark_alias.

ParquetFile

ParquetFile(
    data: Any = None,
    *,
    stat: IOStats | None = None,
    url: URL | None = None,
    binary: bytes | bytearray | memoryview | None = None,
    path: PathLike | None = None,
    holder: "IO | None" = None,
    owns_holder: bool = False,
    mode: ModeLike = "rb+",
    media_type: Any = None,
    temporary: bool = False,
    singleton_ttl: Any = ...,
    **kwargs
)

Bases: IO[bytes, ParquetOptions]

:class:Tabular leaf for Apache Parquet.

Initialize the IO.

Exactly one of url / binary / path / data / holder determines the seed; the rest are mutually exclusive (validated in :meth:__new__).

holder= (alias: parent=) borrows an existing IO as backing storage — every byte primitive then delegates through :meth:_active. owns_holder=True transfers close-ownership so closing this IO also closes the parent.

temporary=True marks the IO for self-cleanup on release: :meth:_release calls :meth:clear so the payload is dropped when the IO closes. Default False — clears only happen when the caller asks.

mode follows stdlib :func:open semantics, normalized to a :class:Mode enum. Side effects fire on :meth:_acquire, not here: cursor stays at byte 0 until then.

stat lets callers seed the metadata cache (size / mtime / media_type) when they already know it — saves a backend probe on the first :meth:stat call.

opened property

opened: bool

True iff :meth:_acquire has run and :meth:_release hasn't.

closed property

closed: bool

Stdlib IO[bytes] parity — False while the bound backing is reachable.

Stdlib semantics: closed means "file unusable for I/O." On a cursor the predicate flips only when teardown has dropped the parent reference; on a storage IO it always reads False (the storage owns its own bytes). Matters for pyarrow / pandas / polars / zipfile, which guard every op with an assert not closed.

url property writable

url: 'URL'

Canonical URL identifying this holder.

parent property

parent: 'IO | None'

The IO one level up — cursor parent first, else URL parent.

Resolution order:

1. The cursor parent (self._parent, set by :meth:IO.open and by format-leaf construction with parent= / holder=). When set, this IO is a cursor and the parent is its backing storage.
2. The URL parent — a sibling IO of the same concrete class at self.url.parent. Used by URL-shaped storage leaves (:class:Path / :class:LocalPath / remote paths) to walk up the filesystem.

Returns None when neither applies (top-level storage with no URL hierarchy — e.g., :class:Memory, which overrides :meth:_url_parent to skip the URL branch).

parents property

parents: 'Iterator[IO]'

Walk the parent chain outward, yielding one IO per step.

Each step follows :attr:parent — cursor parent first, then URL parent (when applicable), terminating when .parent returns None. Empty on top-level non-URL storage (:class:Memory).

size property

size: int

Current visible size in bytes.

Cursor / format-leaf IOs read the bound parent's size; storage subclasses override directly.

size_known property

size_known: bool

True when reading :attr:size won't trigger a backend probe.

Always true for in-memory IOs (size is a slot). Path IOs override to True only when their stat cache is warm — callers that want to short-circuit on an empty buffer (parquet / arrow IPC / CSV readers checking size == 0) can guard the check on this predicate so a cold remote path doesn't pay a HeadObject / get_status / get_metadata round trip just to discover the file is non-empty. Cursor / format-leaf IOs delegate to the parent.

holder_is_overwrite property

holder_is_overwrite: bool

True when the backing holder was opened in OVERWRITE mode.

Primitives use this to skip append checks: the holder was already truncated so there is no existing data to merge with.

mtime property

mtime: float

Last-modified time stamp.

media_type property writable

media_type

The holder's :class:MediaType, or None if unset.

Resolves lazily on first read: a fresh holder bound only by URL carries the sentinel ... in :attr:_media_type and runs :meth:URL.infer_media_type here once, caching the result back onto the slot. Subsequent reads (and pickling, IOStats snapshots, codec dispatch, …) hit the cached value.

Cursor IOs (those wrapping a :attr:parent storage) defer to the parent's stamped media type when their own slot is unset — the codec / format dispatch on a :class:JSONFile bound to a gzip-stamped :class:Memory parent needs to see the parent's media type, not its own (the cursor was constructed bare).

is_memory property

is_memory: bool

True when the IO lives entirely in process memory.

Cursor / format-leaf IOs delegate to the bound parent. Storage subclasses (:class:Memory) override directly.

is_local_path property

is_local_path: bool

True when the IO is a path on the local filesystem.

Cursor / format-leaf IOs delegate to the bound parent. Storage subclasses (:class:LocalPath) override directly.

is_remote_path property

is_remote_path: bool

True when the IO is a path on a non-local backend.

Cursor / format-leaf IOs delegate to the bound parent. Storage subclasses (remote paths) override directly.

is_streaming property

is_streaming: bool

True when :attr:size reflects only the bytes pulled so far.

Streaming holders (:class:MemoryStream over a live source) lazily pull bytes on read; their :attr:size grows as the cursor advances and may underreport the eventual total. Static holders (:class:Memory, :class:Path) know their full size up front so the default is False.

:class:IO.read checks this flag to decide whether to cap the requested byte count at :attr:size (static case — out-of-range reads would raise) or pass the request through unclamped (streaming case — the holder pulls until it has enough or EOF).

xxh3_64_digest property

xxh3_64_digest: bytes

8-byte big-endian payload digest — equivalent to xxh3_64().digest() but served from the cached :meth:xxh3_int64 so callers mixing the digest into a parent hash don't re-walk the payload.

holder property

holder: 'IO'

The bound parent IO (cursor case) or self (storage case).

Backwards-compatible alias preserved from the pre-merge IO.holder property — call sites that drilled through a cursor to reach its backing storage keep working.

owns_holder property

owns_holder: bool

Whether closing self also closes the bound parent.

mode property

mode: Mode

The typed :class:Mode enum this buffer was opened with.

pandas / pyarrow / zipfile inspect .mode for substrings like "b" to dispatch binary vs text reads; those sniffs still work because :class:Mode implements __contains__ against its :attr:~Mode.os_mode form ("b" in handle.mode → True). Reach for self.mode.os_mode when an actual POSIX string is required.

open

open(
    mode: ModeLike = "rb+",
    *,
    media_type: "MediaType | None" = None,
    owns_holder: bool = False,
    auto_open: bool = True,
    **kwargs: Any
) -> "IO"

Acquire the IO and return a fresh :class:IO cursor over it.

Dispatches to the format-specific :class:IO leaf via the IO's stamped media type (or media_type override), so LocalPath("data.parquet").open() lands on :class:ParquetFile, LocalPath("data.csv").open() on :class:CSVFile, and an unknown / no-media holder falls back to a plain :class:IO.

Pattern::

with LocalPath("/tmp/x.bin").open("wb") as bio:
    bio.write(b"hello")
# path released here.

with LocalPath("data.parquet").open() as bio:
    table = bio.read_arrow_table()  # Tabular surface
# path released here.

The default owns_holder=False returns a non-owning cursor — closing the cursor leaves the parent open, so the caller can mint multiple cursors against the same parent. Pass owns_holder=True to transfer close-ownership of the parent to the cursor (the cursor's close then also closes the parent).

commit

commit()

Commit current state

rollback

rollback()

Rollback current state

close

close(force: bool = False) -> None

Release the IO; on :attr:temporary, discard pending writes instead of committing them.

On a cursor with owns_holder=True the bound parent is closed too. Preserves the cursor position across the close — a reopen on the same instance lands at the byte the previous transaction left off.

mark_dirty

mark_dirty() -> None

Signal pending mutations — commit on next clean :meth:close.

for_scheme classmethod

for_scheme(scheme: Any) -> 'type[URLBased]'

Return the :class:URLBased subclass registered for scheme.

Lazy: if no subclass is registered yet, this routes through :meth:Scheme.path_class which imports the backend module on demand (firing :meth:__init_subclass__ as a side effect).

Raises :class:ValueError for an unknown scheme and :class:ImportError when the backend's optional dependencies aren't installed.

dispatch classmethod

dispatch(url: Any, **kwargs: Any) -> 'URLBased'

Build the right :class:URLBased subclass from url.

Looks up the subclass via :meth:for_scheme, then delegates to that subclass's :meth:from_url. Used as the cross-cutting entry point when the caller has a URL but doesn't know (or care) which concrete class owns its scheme.

URL.from_(url).scheme drives the lookup; an empty scheme falls back to the file:// handler so bare paths work.

from_url classmethod

from_url(url: URL, **kwargs) -> 'IO'

Create a new IO from a URL.

When cls is abstract (has subclasses but isn't itself constructible — e.g. :class:Path), the URL scheme is resolved through the :class:URLBased registry to a concrete subclass; an unknown scheme raises :class:ValueError instead of producing the obscure "Can't instantiate abstract class" :class:TypeError.

to_url

to_url() -> 'URL'

The canonical :class:URL that addresses this holder.

to_singleton

to_singleton(ttl: Any = ...) -> 'Singleton'

Promote this instance into the per-class _INSTANCES cache.

Hot listing paths (iterdir / _ls / glob) build children with singleton_ttl=False so the bounded cache doesn't fill up with thousands of short-lived entries. When a caller decides one of those children is worth keeping around (handing it to a long-running worker, returning it from an API), :meth:to_singleton registers self into the cache so the next constructor call with the same key collapses to the same instance.

ttl defaults to the subclass's _SINGLETON_TTL (... = no caching, None = process lifetime, or a seconds count). When a different instance is already cached under this key, that pre-existing one wins and is returned unchanged — the cache is the source of truth.

invalidate_singleton

invalidate_singleton(remove_global: bool = True) -> None

Pop self from the per-class _INSTANCES cache.

Mutating ops on a Singleton-cached object (writes, deletes, schema invalidations on a Databricks table, put_object on an :class:S3Path) want to make sure the next caller asking for the same key gets a fresh build rather than collapsing onto this stale handle — that's what remove_global=True (the default) does. The pop is :meth:identity-guarded: only an entry that still points at self is removed, so a concurrent re-construction that already raced past this thread is left alone.

remove_global=False is a no-op. The keyword exists so subclass invalidators (invalidate_singleton, _invalidate_entity_tag_cache, …) can offer the same switch without branching at the call site.

class_for_media_type classmethod

class_for_media_type(
    media_type: "MediaType | MimeType | str | Any", *, default: Any = ...
) -> "type"

Resolve a :class:MediaType (or coercible) to its format leaf.

Looks up :attr:MediaType.mime_type's name in :data:_HOLDER_FORMAT_REGISTRY. Codec is orthogonal — Parquet compressed with zstd or snappy still resolves to :class:ParquetFile; the codec layer is the holder's concern.

The returned class is a :class:Tabular subclass — typically a :class:Holder byte-backed leaf, occasionally a non-Holder leaf (:class:Folder, :class:DeltaFolder). Returns default on miss when supplied; otherwise raises :class:KeyError with the list of registered names.

matches_static

matches_static(
    predicate: "Predicate", *, free_cols: "tuple[str, ...] | None" = None
) -> bool

True iff predicate could match any row given :attr:static_values. Conservative on undecidables (column not in static values, predicate evaluator failure) so the caller still reads.

Builds a one-row pyarrow Table from the predicate's free columns that we have static values for, then evaluates the predicate against it — generalises the partition-only prune so any aggregator (folder read, future warehouse file skip) reuses the one helper.

free_cols lets a caller that's about to prune the same predicate against N children precompute the free-column tuple once and reuse it — :func:free_columns walks the AST every call, so on a 64-OR predicate (the cache batch lookup shape) the saving is N-1 full walks per iter_children loop. Default None keeps the call site short for one-off prune checks.

from_ classmethod

from_(
    obj: Any, *, url: URL | None = None, mode: ModeLike = "rb+", **kwargs
) -> "IO"

Auto-route obj to the right storage / cursor, return an owning IO.

Two shapes share the method:

• Storage subclasses (cls has a :attr:scheme — :class:IO itself, :class:Memory, :class:LocalPath, remote paths). The result is a storage IO that owns its bytes — IO.from_(b"x") → :class:Memory, IO.from_("file://...") → :class:LocalPath.
• Cursor / format-leaf subclasses (cls has no scheme — :class:IO, :class:ParquetFile, :class:CSVFile, …). The result is an owning cursor over a fresh storage parent built from obj.

Recognised input shapes:

• :class:IO of cls — pass through (idempotent).
• :class:IO of a different class — for storage cls, return the underlying parent; for cursor cls, borrow the same parent into a fresh cursor.
• bytes-like (bytes / bytearray / memoryview) — back with a fresh :class:Memory.
• path-like (str / pathlib.Path / URL) — back with the path-shaped storage class for the scheme.
• local file handle — back with :class:LocalPath; lazy read from disk (no drain).
• other file-like — drain into a fresh :class:MemoryStream.

check_options classmethod

check_options(
    options: "O | None" = None, overrides: "dict | None" = None, **kwargs: Any
) -> O

Validate and merge caller kwargs into a resolved options.

Canonical pattern: a public method passes overrides=locals() and the ...-defaulted entries are stripped, the rest merged.

cleanup

cleanup(wait: 'Any' = False) -> int

Garbage-collect stale state on this backend.

Default no-op (returns 0) — single-file leaves and warehouse-backed tables don't have a sweep concept the client owns. Folder-shaped subclasses override to unlink stale part-* files, throttled by TTL.

wait controls sync vs async dispatch on backends that support it: a truthy :class:yggdrasil.dataclasses.waiting.WaitingConfig (or True / a positive timeout) blocks until the sweep finishes; a falsy value (the default) hands the work off to a background thread. Backends without an async path treat both the same.

Returns the number of files / rows removed when known; 0 for fire-and-forget async dispatch or a no-op backend.

optimize

optimize(byte_size: 'int | None' = None, **kwargs: Any) -> int

Repartition / compact this Tabular's storage.

Default implementation is a no-op and returns 0 — single-file leaves (parquet, csv, arrow IPC, …) don't have a compaction concept. Aggregator subclasses (:class:Folder) override this to walk their child leaves and bin-pack small part files into bundles near byte_size. Files already close to the target size are left alone so a repeated call is cheap.

byte_size=None keeps the legacy "collapse every leaf with more than one part into a single file" behavior, which is what the local-cache compaction loop in :class:Session expects. Any extra keyword arguments are accepted and ignored so upstream callers can pass forward-compatible knobs without the base raising.

delete

delete(
    predicate: "PredicateLike" = None,
    *,
    wait: "WaitingConfigArg" = True,
    missing_ok: bool = False,
    delete_staging: bool = True,
    **kwargs: Any
) -> "Table"

Delete rows matching predicate; return this tabular.

predicate is a :class:Predicate from :mod:yggdrasil.execution.expr or a SQL string that parses into one ("id IN (1,2,3)", "price > 100 AND region = 'EU'"). None means "no filter" — every row is removed (DELETE FROM t with no WHERE).

wait / missing_ok / delete_staging are honoured by resource-backed subclasses (e.g. :class:yggdrasil.databricks.table.table.Table, which drops the table asset); the generic row-rewrite path ignores them. Any extra **kwargs (e.g. options=DeltaOptions(...)) flow through to :meth:_delete.

The default implementation reads every batch, drops rows the predicate accepts, and rewrites the leaf with the survivors. Aggregator subclasses (:class:yggdrasil.path.folder.Folder) override to walk children, prune subtrees whose partition bounds make the predicate trivially false, and only rewrite the leaves that actually hold matched rows.

collect_schema

collect_schema(options: 'O | None' = None, **kwargs: Any) -> Schema

Return this Tabular's :class:Schema, caching the first hit.

The cache slot is :attr:_schema_cache; on first call this method stamps the resolved schema into it so subsequent collect_schema calls short-circuit. Writers overwrite the slot via :meth:_persist_schema; lifecycle hooks clear it via :meth:_unpersist_schema.

count

count(options: 'O | None' = None, **kwargs: Any) -> int

Return the number of rows in this tabular.

scan_arrow_batches

scan_arrow_batches(
    options: "O | None" = None, **kwargs: Any
) -> Iterator[pa.RecordBatch]

Zero-copy scan — yield the source's :class:pa.RecordBatch views verbatim.

The lazy / zero-copy counterpart to :meth:read_arrow_batches, mirroring :meth:read_polars_frame vs :meth:scan_polars_frame. Where read_arrow_batches layers the full options pipeline on every batch — target cast, projection, resample, dedup, row-limit slicing, each of which can copy or re-encode — scan_arrow_batches hands back exactly what the leaf produced, untouched. For an in-memory source (:class:~yggdrasil.arrow.tabular.ArrowTabular) those batches are views over the held buffers (no copy); for a byte-backed leaf they're the freshly-decoded batches with none of the extra processing copies layered on. Use it when you want the raw Arrow stream and will project / filter downstream yourself.

scan_arrow_table

scan_arrow_table(options: 'O | None' = None, **kwargs: Any) -> pa.Table

Zero-copy scan into one chunked :class:pa.Table (no rechunk, no cast).

The zero-copy counterpart to :meth:read_arrow_table. Assembles the source batches with :func:pa.Table.from_batches, which references the batch buffers as table chunks rather than copying them — so no cast, no projection, no rechunk memcpy that read_arrow_table performs to coalesce + conform the result. An empty source yields an empty table carrying the bound schema.

The batches must share one schema (the zero-copy contract): read_arrow_table reconciles parts that drifted across writes, scan_arrow_table does not — reach for read_arrow_table when a source's parts are known to be heterogeneous.

scan_arrow_batch_reader

scan_arrow_batch_reader(
    options: "O | None" = None, **kwargs: Any
) -> "pa.RecordBatchReader"

Zero-copy scan as a streaming :class:pa.RecordBatchReader view.

The raw-reader counterpart to :meth:read_arrow_batch_reader: wraps the source batch stream in a reader without the per-batch conform / target-cast pass, so batches flow through as views over the source buffers. The reader's schema is the source's own — taken from the first batch, so it matches the raw views exactly (no collect_schema probe, which on a byte cursor would consume the stream out from under the read). Only the first batch is pulled up front to seed the schema; the rest stay lazy behind the reader.

read_table

read_table(options: 'O | None' = None, **kwargs: Any) -> 'Tabular | None'

Read into an in-memory :class:Tabular.

When options.spark_session is set, reads via :meth:_read_spark_frame and wraps in a :class:Dataset. Otherwise materializes Arrow batches into :class:ArrowTabular. Returns None when empty.

write_table

write_table(obj: Any, options: 'O | None' = None, **kwargs: Any) -> None

Dispatch obj to the best _write_* hook based on its runtime type.

Recognizes another :class:Tabular (drained as a pyarrow record-batch stream), pa.Table / pa.RecordBatch / pa.RecordBatchReader, polars DataFrame / LazyFrame, pandas DataFrame, pyspark DataFrame, list[dict], dict[str, list], and iterables of any of the above. Module-name sniffing keeps optional engine deps out of the import graph — we only touch a frame's API once we've confirmed it's an instance of one we know how to drain.

union

union(other: 'Any', *, mode: 'ModeLike | None' = None) -> 'Tabular'

Return a Tabular representing self UNION ALL other.

mode controls how mismatched schemas are reconciled:

• Mode.IGNORE (default) — keep self's schema; extra columns in other are dropped, missing ones are filled null.
• Mode.APPEND — widen to the superset schema (every field from both sides survives).

Concrete subclasses override :meth:_union for in-place mutation (Arrow batch append, Spark unionByName).

Accepts :class:Tabular, pa.RecordBatch, pa.Table, list[Response], or a Spark DataFrame. None returns self unchanged.

read_spark_dataset

read_spark_dataset(options: 'O | None' = None, **kwargs: Any) -> 'SparkDataset'

Read into a :class:Dataset holder.

Mirrors :meth:read_arrow_dataset for the Spark engine: the return type is a yggdrasil holder rather than the bare engine frame, so callers keep the Tabular surface (chained transforms, persist / insert / schema, …) without an extra wrap at the call site. :class:Dataset overrides :meth:_read_spark_dataset to return itself in place — no materialise round trip when the source already speaks Spark.

read_record_iterator

read_record_iterator(
    options: "O | None" = None, **kwargs: Any
) -> "Iterator[Mapping[str, Any]]"

Stream rows as plain dict. True streaming — the full table never materializes; batch.to_pylist() does the column→row rotation in pyarrow C++ once per batch.

read_records

read_records(options: 'O | None' = None, **kwargs: Any) -> 'Iterator[Any]'

Stream rows as :class:yggdrasil.data.record.Record. Lower per-row allocation than :meth:read_pylist for stable-schema sources — the underlying :class:Schema is materialized once and shared by reference across every record.

unique

unique(by: 'str | Any | Iterable[Any]') -> 'Tabular'

Drop duplicate rows on by; keep first occurrence per key tuple.

Parameters

by One or more column references — :class:str column names, :class:yggdrasil.data.Field instances (resolved via :attr:Field.name), or any iterable mixing the two. Empty / None is a no-op — returns self.

Returns

Tabular A new holder carrying the deduped rows. Spark-shaped inputs (anything whose :meth:_native_spark_frame exposes a :class:pyspark.sql.DataFrame) return a fresh :class:yggdrasil.spark.tabular.Dataset over the spark-side dedup; everything else collects through Arrow and returns an :class:yggdrasil.arrow.tabular.ArrowTabular.

resample

resample(
    on: "str | Any",
    sampling: "int | float | Any",
    *,
    partition_by: "str | Any | Iterable[Any] | None" = None,
    fill_strategy: "str | None" = "ffill"
) -> "Tabular"

Align rows to a fixed time grid on on; one row per bucket.

Parameters

on The time column to resample on — column name (:class:str) or :class:yggdrasil.data.Field. sampling Bucket size. Accepted shapes:

* :class:`int` / :class:`float` — seconds (floats are
  rounded to the nearest integer second).
* :class:`datetime.timedelta` — total seconds.
* :class:`str` — ISO-8601 duration (``"PT1H"``,
  ``"P1D"``, ``"PT15M"``) parsed via
  :func:`yggdrasil.data.types.primitive.temporal._parse_iso_duration`.

``sampling <= 0`` is a short-circuit — returns ``self``.

partition_by Entity columns the resample is independent on. None / empty → flat global timeline. Same coercion as :meth:unique's by. fill_strategy How to fill nulls left by the bucket's "first" aggregation. "ffill" (default), "bfill", or "none" / None to disable. See :func:yggdrasil.arrow.ops.fill_arrow_table for the full semantics.

Returns

Tabular Spark-shaped holders return a :class:Dataset over the spark-side resample; everything else returns an :class:ArrowTabular over the arrow-side resample.

select

select(*columns: 'str | Any') -> 'Tabular'

Project to columns and return a new Tabular.

Each entry is a column reference — :class:str, a :class:yggdrasil.data.Field (resolved via :attr:Field.name), or an iterable mixing both. The result preserves the caller's order, which matches both :meth:pyarrow.Table.select and :meth:pyspark.sql.DataFrame.select semantics.

Raises :class:ValueError on an empty selection — a zero- column projection is almost always a caller mistake; pass :class:Schema.empty projections through the cast surface instead.

drop

drop(*columns: 'str | Any') -> 'Tabular'

Return a new Tabular with the named columns removed.

Columns missing from the source are silently ignored — matches Spark's :meth:DataFrame.drop and pyarrow's :meth:Table.drop_columns (when filtered to existing names). An empty argument list is a no-op that returns self.

filter

filter(predicate: 'PredicateLike') -> 'Tabular'

Drop rows where predicate is false.

predicate accepts every shape :meth:yggdrasil.execution.expr.Expression.from_ recognises:

• a SQL predicate string ("x > 0 AND y IS NOT NULL"), parsed by the in-tree SQL parser;
• a yggdrasil :class:Predicate node (col("x") > 0, :func:is_in, :func:between, …);
• a native engine expression — :class:pyarrow.compute.Expression, :class:polars.Expr, or :class:pyspark.sql.Column — lifted via the matching backend.

The predicate is parsed once and dispatched to the typed :meth:_filter hook; the engine-side filter then runs in its native kernel (Arrow C++, Spark Catalyst) so the row scan stays vectorised.

cast

cast(options: 'O | None' = None, **kwargs) -> 'Tabular'

Cast rows, returning a new :class:Tabular.

Accepts a :class:Schema or :class:CastOptions. When options is given, reads to arrow and casts each batch through :meth:CastOptions.cast_arrow_batch.

display

display(n: int = 10, *, max_width: int = 32) -> str

Render the first n rows as an aligned, typed text table.

Columns and their types come from this Tabular's own :meth:collect_schema — the header is two rows: the column names, then their type tags (the project :class:~yggdrasil.data.Field's :meth:Field.short → :meth:DataType.short, recursive for nested types — i64 / str / list<str> / struct<name:str, age:i64>). Columns are separated by │ with a ─┼─ rule; numbers/booleans right-align; nested cell values are compacted to one line. Long values and headers are clipped (cells to max_width, type/name tags to a slightly larger cap) so one long string or column name can't balloon the table. The n rows are pushed down as a row_limit so no more than that is ever read.

print(dbc.sql.execute("SELECT * FROM t").display())
print(IO.from_("data.parquet").display(5))

lazy

lazy() -> 'LazyTabular'

Return a :class:LazyTabular wrapping this source.

Transformations on the returned object (select, filter, join, …) accumulate in an :class:ExecutionPlan without touching data. Any read_* call materialises the plan.

joinpath

joinpath(*segments: Any) -> 'IO'

Build a sibling IO at self.url joined with segments.

URL-shaped IOs (:class:LocalPath, remote paths) use this to mint a child path; :class:Memory and other non-URL leaves raise :class:ValueError.

from_bytes classmethod

from_bytes(data: bytes, **kwargs) -> 'IO'

Create a new IO from bytes.

from_holder classmethod

from_holder(
    holder: "IO",
    *,
    owns_holder: bool = False,
    mode: ModeLike = "rb+",
    media_type: Any = None,
    auto_open: bool = True,
    **kwargs: Any
) -> "IO"

Construct a cursor over holder, dispatching to the format leaf.

Resolves the format-specific :class:IO leaf via media_type (when given) or the holder's stamped stat().media_type, and returns an instance of that leaf bound to holder. When no leaf can be resolved, falls back to cls itself.

With auto_open=True (the default) the returned cursor is already acquired, so the caller can immediately read/write without entering a with block. Set auto_open=False to defer the acquire to the caller's with / :meth:acquire.

owns_holder=True hands close-ownership of holder to the returned cursor — closing the cursor closes the holder. The default False keeps the holder's lifetime in the caller's hands; the returned cursor is a non-owning borrow.

for_holder classmethod

for_holder(
    holder: "IO",
    *,
    media_type: "MediaType | MimeType | str | None" = None,
    default: Any = ...,
    **kwargs: Any
) -> "Tabular"

Build the right format leaf for holder.

Resolution order for the format discriminator:

1. The explicit media_type kwarg, when supplied.
2. holder.stat().media_type — set by the holder from its URL extension, magic-byte sniff, or content-type header.

The resolved class is instantiated as Cls(holder=holder, **kwargs). On lookup miss, falls back to default when supplied; otherwise raises :class:KeyError.

registered_classes classmethod

registered_classes() -> 'dict[str, type]'

Snapshot of the registry — debugging / introspection only.

read_mv

read_mv(size: int = -1, offset: int = 0, *, cursor: bool = False) -> memoryview

Slice size bytes from offset as a :class:memoryview.

cursor=True ignores the explicit offset and reads from the holder's internal cursor (:attr:tell), advancing it past the bytes returned. cursor=False (default) keeps the cursor-less positional contract — the cursor is untouched.

Cursor IOs (those wrapping a :attr:parent storage) delegate the whole call through :meth:_active so the parent's bounds-check uses its own size — avoids a redundant stat probe on remote backings when the cursor has no local size cache, and routes through any subclass _active override (lazy materialization on :class:ZipEntryFile, …).

write_mv

write_mv(
    data: memoryview,
    offset: int = 0,
    *,
    size: int = -1,
    overwrite: bool = False,
    update_stat: bool = True,
    cursor: bool = False
) -> int

Splice data at offset, pre-growing the holder as needed.

size caps the byte count written — size=-1 (default) writes all of data; size>=0 writes min(len(data), size) bytes. Caps via a slice of data (zero-copy on memoryview / bytes), so downstream pipelines that only need the first N bytes of a larger buffer skip the trailing tail.

overwrite declares that this write replaces the holder's tail past offset + size — after the splice, :attr:size is set to offset + size. Callers that currently do truncate(0) followed by write_bytes(...) collapse to a single write_bytes(..., overwrite=True), which on whole-blob remote backends saves a SDK round trip (the atomic upload at offset == 0 already replaces the object — no preceding truncate needed).

Pipeline:

1. Slice data to size if capped.
2. Normalize offset (-1 → append, -N → self.size - N).
3. Pre-grow visible :attr:size to cover the splice via :meth:resize.
4. Hand the normalized (data, offset) to :meth:_write_mv.
5. Truncate tail past offset + n when overwrite.
6. Mark dirty + bump cached mtime if anything was written.

update_stat=False skips the post-write :meth:_touch_stat and :meth:mark_dirty calls. Use it for bulk loops that want a single stat refresh at the end (one :func:time.time call instead of one per write); the caller is then responsible for calling :meth:_touch_stat (or re-statting via the path-side _stat for filesystem backends) once the loop finishes.

Cursor IOs (those wrapping a :attr:parent storage) delegate the whole call through :meth:_active so the parent's resize / bounds-check / dirty-marking fires once, on the backing storage — the cursor only advances its own _pos.

reserve

reserve(n: int) -> None

Pre-grow capacity to at least n bytes.

Capacity-only — does NOT change :attr:size. Idempotent when capacity ≥ n already. Subclasses with no growable capacity layer may treat this as a no-op. Cursor / format-leaf IOs delegate to the bound parent.

resize

resize(n: int) -> int

Grow visible :attr:size to at least n bytes (one-way).

Sister of :meth:truncate, but never shrinks. Used by :meth:write_mv to pre-allocate a known target before the splice so :meth:_write_mv doesn't have to manage size.

• n <= size → no-op, returns current :attr:size.
• n > size → extends with zero-padding via :meth:truncate, returns n.

Subclasses with a native grow-only primitive (capacity hint to a remote upload session, posix_fallocate on local fd) override for the cheaper path; the default works on every backend.

truncate

truncate(size: 'int | None' = None) -> int

Set the visible :attr:size to exactly size bytes.

Shrinks drop the tail; extends zero-pad. Returns the new size.

On a cursor (self._parent is not None), size=None truncates at the current cursor position and the cursor is clamped if it would exceed the post-truncate size. On a storage IO size=None is invalid — pass an explicit byte count.

clear

clear() -> None

Drop the IO's payload entirely.

:class:Memory resets the underlying bytearray to zero bytes (capacity drops too). :class:yggdrasil.io.path.Path unlinks the backing file with missing_ok=True so the operation is idempotent. After :meth:clear, :attr:size reads 0 and the IO is still usable — subsequent writes grow it from scratch.

stat

stat() -> IOStats

Snapshot the holder's metadata into a fresh :class:IOStats.

Delegates to :meth:_stat for the backend-specific fields (kind and the live size for path-bound holders); mutating the returned instance does NOT round-trip onto the holder. Use the holder's own setters / :meth:_touch_stat when you need to update metadata.

touch_mtime

touch_mtime(when: float | None = None) -> None

Stamp the holder's mtime with the current time.

Bulk-write helper — call once after a write loop instead of letting every :meth:write_mv call sample the clock. when accepts an explicit timestamp (e.g. an upstream "Last-Modified" header); None defaults to :func:time.time.

acquire

acquire() -> 'IO'

Bring the IO's backing into the acquired state.

Lifecycle primitive — idempotent. Returns self. :meth:__enter__ calls this; so does :meth:open before constructing its cursor IO.

flush

flush() -> None

Push buffered writes to the durable backing.

Cursor IOs forward the flush to their bound parent; storage IOs go through :meth:Disposable.commit (default no-op unless a subclass overrides).

pread

pread(n: int, pos: int, *, cursor: bool = False) -> bytes

Positional read. Returns at most n bytes at pos.

cursor=True reads from the internal cursor instead of pos and advances it past the bytes returned.

pwrite

pwrite(
    data: Union[bytes, bytearray, memoryview],
    pos: int,
    *,
    update_stat: bool = True,
    cursor: bool = False
) -> int

Positionally write. Returns bytes actually written.

update_stat=False defers the post-write stat refresh to the caller — see :meth:write_mv for the bulk-write rationale. cursor=True writes at the internal cursor instead of pos and advances it by the bytes written.

memoryview

memoryview() -> memoryview

View over the holder's visible bytes.

iter_mv

iter_mv(
    chunk_size: int = 256 * 1024,
    *,
    start: int = 0,
    length: Optional[int] = None
) -> Iterator[memoryview]

Yield [start, start+length) in bounded, zero-copy memoryview chunks (default: the whole holder from start).

Each chunk is a :meth:read_mv slice — a view straight into the live in-memory window, or a bounded read for spilled / file-backed storage — so a consumer like http.client can sock.sendall it without a copy, and never more than chunk_size is resident at once. Reads are positional (the cursor is untouched), so the holder can be iterated again — e.g. a connection retry re-sending the same body — by calling this afresh.

read_bytes

read_bytes(size: int = -1, offset: int = 0, *, cursor: bool = False) -> bytes

Read size bytes starting at offset as :class:bytes.

size=-1 reads to EOF; offset accepts negative indices via :func:_resolve_pos (-1 → size, -N → self.size - N). cursor=True reads from the internal cursor and advances it past the bytes returned.

write_bytes

write_bytes(
    data: Any,
    offset: int = 0,
    *,
    size: int = -1,
    overwrite: "bool | None" = None,
    cursor: bool = False
) -> int

Splice data at offset. Returns bytes written.

overwrite defaults to None → resolved: a whole-content write from the start (offset == 0, size == -1, no cursor) replaces the object (pathlib write_bytes truncate semantics), so a whole-blob remote backend does it in a single PUT instead of a stat + read-page + upload read-modify-write. A positional / cursor / size-capped write is a splice that preserves the rest, so it resolves to False. Pass an explicit True / False to force either.

size caps the byte count written — size=-1 (default) writes the entire source; size>=0 writes at most size bytes. The cap is forwarded into each type-directed branch so a stream source stops reading after size bytes (no over-pull) and a bytes-like source slices its tail off before dispatching.

overwrite declares that this write replaces every byte from offset onward. The holder ends at offset + bytes_written regardless of its prior size, and whole-blob remote backends collapse the implied truncate(...) + write(...) pair into one SDK call.

Type-directed dispatch — bytes-like payloads (:class:bytes, :class:bytearray, :class:memoryview, and str after UTF-8 encoding) splice through :meth:write_mv; other :class:Holder instances route through :meth:write_holder (size-aware: small payloads write inline, large ones stream); file-like sources (anything exposing .read) drain through :meth:write_stream. Subclasses override :meth:_write_mv, :meth:_write_stream, and / or :meth:_write_holder rather than this dispatch.

read_text

read_text(
    encoding: str = "utf-8",
    errors: str = "strict",
    *,
    size: int = -1,
    offset: int = 0,
    cursor: bool = False
) -> str

Decode size bytes at offset as text.

cursor=True reads from the internal cursor and advances it.

write_text

write_text(
    text: str,
    encoding: str = "utf-8",
    errors: str = "strict",
    *,
    offset: int = 0,
    cursor: bool = False
) -> int

Encode text and splice at offset. Returns bytes written.

cursor=True writes at the internal cursor and advances it.

head

head(size: int, *, offset: int = 0) -> bytes

Peek the first size bytes from offset (default 0).

A bounded positional read off the front of the object that leaves the internal cursor (:meth:tell) untouched — head composes with cursor reads without disturbing them. size is clamped to what's available, so a short object (or one shorter than offset + size) returns fewer bytes rather than raising; size < 0 reads from offset to EOF.

tail

tail(size: int) -> bytes

Peek the last size bytes, leaving the cursor untouched.

The end-anchored companion to :meth:head — a bounded positional read off the back of the object. size is clamped to the object's length, so requesting more than exists (or size < 0) returns the whole object. The internal cursor (:meth:tell) is not moved.

readinto

readinto(buffer: Any, *, offset: int = 0, cursor: bool = False) -> int

Fill buffer with bytes starting at offset.

Returns the number of bytes written into buffer — min(len(buffer), self.size - offset). Matches the stdlib :meth:io.RawIOBase.readinto shape. cursor=True reads from the internal cursor and advances it.

On a cursor IO (_parent is not None) the default flips to cursor-anchored — stdlib readinto(buf) then matches the BinaryIO contract.

readline

readline(limit: int = -1, *, offset: int = 0, cursor: bool = False) -> bytes

Read up to the next newline starting at offset.

Returns the line including the trailing \n (or short when EOF lands first). limit >= 0 caps the byte count. cursor=True reads from the internal cursor and advances it past the returned line. On a cursor IO the default flips to cursor-anchored.

readlines

readlines(
    hint: int = -1, *, offset: int = 0, cursor: bool = False
) -> list[bytes]

Read every line from offset to EOF (or until hint bytes).

cursor=True reads from the internal cursor and advances it past the bytes consumed. On a cursor IO the default flips to cursor-anchored.

tell

tell() -> int

Current cursor position.

seek

seek(offset: int, whence: int = 0) -> int

Seek the internal cursor to offset relative to whence.

Mirrors :meth:io.IOBase.seek with two ergonomic deviations:

• seek(-1, SEEK_SET) is a "go to end" sentinel — pairs with read(-1) / "read all". Any other negative SEEK_SET offset raises :class:ValueError.
• SEEK_CUR / SEEK_END with a negative offset that would land before byte 0 clamps to 0 instead of raising.

write_local_path

write_local_path(
    path: PathLike,
    *,
    pos: int = 0,
    n: int = -1,
    chunk_size: int = _COPY_CHUNK,
    cursor: bool = False
) -> int

Load path's bytes into this holder at pos.

n < 0 reads the whole file; n >= 0 caps the source bytes pulled at n. Streams in chunk_size slices so a large file doesn't materialize into memory.

Pre-allocates the holder via :meth:resize when the source size is known up front (n >= 0 or local stat available), so the inner loop only writes — no per-chunk grow.

write_stream

write_stream(
    src: Any,
    *,
    offset: int = 0,
    size: int = -1,
    overwrite: bool = False,
    batch_size: int = _COPY_CHUNK,
    cursor: bool = False
) -> int

Drain a binary source into this holder at offset.

Public entry point: accepts a yggdrasil :class:IO[bytes], a stdlib :class:typing.BinaryIO (io.BytesIO, open(..., "rb"), urllib3 responses, …), or any file-like carrying a .read. Non-:class:IO sources are coerced via :meth:IO.from_ so subclass-side :meth:_write_stream always receives a real :class:IO[bytes].

size caps the byte count drained from src — size=-1 (default) reads to EOF; size>=0 stops at size bytes (no over-pull from the source).

overwrite truncates the holder's tail past offset + bytes_written; whole-blob remote backends get a single atomic PUT instead of an explicit truncate followed by a write.

batch_size is the read/write chunk size for the default streaming path (:data:_COPY_CHUNK, 1 MiB). Tune up for high-throughput remote sinks where the per-call overhead dominates, or down to bound peak memory on a slow consumer.

write_holder

write_holder(
    src: "Holder",
    *,
    offset: int = 0,
    size: int = -1,
    overwrite: bool = False,
    batch_size: int = _COPY_CHUNK,
    cursor: bool = False
) -> int

Splice another :class:Holder's bytes into this one at offset.

Public entry point: validates the inputs, then dispatches to :meth:_write_holder. size caps the byte count pulled from src — size=-1 (default) writes the whole source; size>=0 writes the first size bytes. overwrite truncates the tail past offset + bytes_written (collapses truncate(...) + write_holder(...) into one operation for whole-blob remote backends). batch_size is forwarded to the streaming path for above-threshold payloads.

Subclasses override the private hook to swap in a backend-aware fast path (Workspace / Volumes / S3 can hand the source straight to their atomic-upload SDK call without ever materialising the bytes in Python).

upload

upload(src: Any, *, size: int = -1, offset: int = 0) -> 'Holder'

Upload src's bytes into this holder.

Symmetric to :meth:download but indexed from the destination side — dst.upload(src) makes the destination's content equal to the source's.

src accepts any of:

• :class:Holder (incl. any :class:Path subclass) — its bytes are pulled starting at offset.
• :class:IO cursor — offset (if non-zero) seeks before read(); otherwise the cursor's current position is honoured.
• str / :class:os.PathLike — coerced via Path.from_(src) and treated as a holder.

size and offset slice the source: size=-1 (default) reads to EOF, size>=0 caps the byte count, offset is the starting offset. Slicing forces the whole-payload fast path in :meth:_transfer_to to defer to a bytes copy (the backend-specific shortcuts — shutil.copyfile, write_local_path — don't expose a window).

When self is a :class:Path whose URL ends in a trailing / (directory shape), the source's filename (src.url.name or "download" for nameless holders) is joined onto it. No remote stat is issued — the trailing slash is a purely local, cp-style hint.

Returns the resolved destination so chains like dst.upload(src).read_bytes() work.

Subclasses with a faster move (e.g. local→local via sendfile, local→remote chunked stream) override :meth:_transfer_to, not this method.

download

download(to: Any = None, *, size: int = -1, offset: int = 0) -> 'Holder | IO'

Copy this holder's bytes to a local target.

When to is :data:None, bytes land in the user's ~/Downloads folder under :attr:url.name (or "download" for nameless holders), with browser-style (1) / (2) / … suffixes appended on name conflict. Otherwise to accepts the same shapes as :meth:upload (:class:Holder, :class:IO, str / :class:os.PathLike). size and offset slice this holder: size=-1 (default) reads to EOF, size>=0 caps the byte count, offset is the starting offset. Returns the resolved target.

to_bytes

to_bytes() -> bytes

Full payload as :class:bytes — alias for read_bytes().

getvalue

getvalue() -> bytes

Stdlib :class:io.BytesIO parity — alias for :meth:to_bytes.

decode

decode(encoding: str = 'utf-8', errors: str = 'replace') -> str

Decode the whole payload as text. Cursorless — does not seek.

to_base64

to_base64(urlsafe: bool = True) -> str

Return the payload base64-encoded as an ASCII str.

urlsafe=True (default) uses :func:base64.urlsafe_b64encode — - / _ in place of + / / so the result drops cleanly into a URL or filename. urlsafe=False falls back to the standard alphabet.

xxh3_64

xxh3_64()

Return an :class:xxhash.xxh3_64 instance over the payload.

Always rebuilds an updatable :class:xxhash.xxh3_64 so callers can keep mixing more bytes in if they want. The expensive part — walking the payload — is short-circuited via the cached digest; we just seed a fresh hasher with the cached value's bytes when available.

xxh3_int64

xxh3_int64() -> int

64-bit xxh3 hash of the payload as a signed int64.

xxh3_64 produces an unsigned 64-bit value; downstream Arrow schemas pin the field as int64, so the digest is wrapped into signed range [-2**63, 2**63). Memoized against (_size, _mtime) — which every write path bumps via :meth:_touch_stat — so repeated reads pay the walk once.

remaining_bytes

remaining_bytes() -> int

Bytes from the cursor to EOF on the active payload.

arrow_input_stream

arrow_input_stream() -> '_ArrowInputStreamContext'

Context manager yielding the cheapest :class:pa.NativeFile over the payload.

Local-path holder + no codec → :func:pyarrow.memory_map (zero-copy). Codec-tagged holder → decompress, then wrap in a :class:pa.BufferReader. Anything else → snapshot and wrap. The yielded stream is always a real :class:pa.NativeFile, so the caller hands it directly to pyarrow readers.

arrow_output_stream

arrow_output_stream(*, append: bool = False) -> '_ArrowOutputStreamContext'

Context manager yielding a :class:pa.BufferOutputStream writer.

with bio.arrow_output_stream() as sink: writer(sink). The yielded sink accepts the format encoder's writes against a pure-Arrow in-memory buffer. On a clean exit the encoded bytes are committed to self via :meth:_commit_format_payload, which handles codec compression and the overwrite-vs-append disposition.

appendable

appendable() -> bool

True when writes append at EOF — :data:Mode.APPEND only.

with_media_type

with_media_type(media_type: Any, *, copy: bool = False) -> 'IO'

Stamp media_type onto the bound IO's metadata.

With copy=False (the default), mutates self and returns it. copy=True allocates a fresh holder over the same bytes and returns a new IO over it.

as_media

as_media(media_type: Any = None) -> 'IO'

Return a typed Tabular leaf bound to this buffer's holder.

.. deprecated:: Use :meth:open with a media_type instead — holder.open(media_type=...) dispatches to the same format leaf and returns an acquired cursor. as_media is retained for callers that haven't migrated.

Resolution: explicit media_type wins; otherwise the buffer's stamped media type is used. The leaf borrows the same backing storage so durable bytes are shared without a copy. When self is already an instance of the resolved leaf class, returns self unchanged.

Raises :class:KeyError when no media type can be resolved or the resolved type has no registered Tabular leaf.

fileno

fileno() -> int

Underlying fd if the holder exposes one. Raises otherwise.

read

read(size: int = -1) -> bytes

Read up to size bytes from the cursor, advancing past them.

Stdlib :meth:io.RawIOBase.read semantic: size < 0 / None reads to EOF; otherwise reads up to size bytes, returning fewer at EOF.

Static IOs (:class:Memory, :class:Path) know their full size up front; cap the request at self.size - self._pos before dispatching so the storage's strict read_bytes doesn't trip on an out-of-range window. Streaming IOs (:class:MemoryStream — is_streaming) lazily pull bytes; forward the request unclamped so the storage pulls until it has enough or signals EOF.

readall

readall() -> bytes

Read from cursor to EOF, advancing the cursor.

write

write(b: Any, *, update_stat: bool = True) -> int

Write b at the cursor, advancing it.

Accepts bytes-like, str (UTF-8), io.BytesIO, or any file-like with .read. File-like sources route through :meth:write_stream so backends with an atomic whole-object upload push a single request. The buffer-protocol fallback catches things like :class:pyarrow.Buffer that aren't bytes/bytearray/memoryview but ARE memoryview-able.

read_bytes_u32

read_bytes_u32() -> bytes

Length-prefixed (uint32 LE) bytes blob.

read_str_u32

read_str_u32(encoding: str = 'utf-8') -> str

Length-prefixed UTF-8 string.

json_load

json_load(*, media_type: Any = None, orient: Any = None) -> Any

Parse the buffer, auto-detecting media type and compression.

Resolution order for the media type:

1. Explicit media_type kwarg.
2. Cached :attr:media_type on the IO.
3. Magic-byte sniff via :meth:MediaType.from_io — when this fires and the IO had no cached media type, the sniffed value is stamped onto the IO so future callers (codec handling, tabular dispatch) see it without re-sniffing.

If the resolved type carries a codec the buffer is decompressed first and the inner mime is stamped onto the decompressed buffer. JSON / NDJSON / opaque-bytes payloads go through json.loads (or pandas.read_json when orient is set); every other registered format dispatches to its :class:Tabular leaf and returns read_pylist().

decompress

decompress(*, codec: Any = None, copy: bool = True) -> 'IO'

Return a new IO over the decompressed payload.

codec may be a :class:Codec, a codec name ("gzip", "zstd", …), or a :class:MediaType-shaped object whose codec attribute is read. Returns the original buffer when no codec is set / supplied.