FileDataSink

laktory.models.datasinks.FileDataSink

Bases: BaseDataSink

Data sink writing to disk file(s) as csv, parquet or Delta Table.

Examples:

Write polars DataFrame as CSV

import polars as pl

import laktory as lk

df = pl.DataFrame({"x": [0, 1]})

sink = lk.models.FileDataSink(
    path="./dataframe.csv", format="CSV", writer_kwargs={"separator": ";"}
)
sink.write(df)

Write Spark Streaming DataFrame as Delta ```python tag:skip-run from laktory import models

df = spark.readStream(...) # skip

sink = models.FileDataSink( path="./delta_table/", format="DELTA", mode="APPEND", checkpoint_path="./delta_table/checkpoint/", ) sink.write(df) ``` References

---

• Data Sources and Sinks

PARAMETER	DESCRIPTION
as_stream	If True output DataFrame is written as Streaming DataFrame. If None, write mode is derived fromDataFrame. TYPE: bool | None | VariableType DEFAULT: None
checkpoint_path_	Path to which the checkpoint file for which a streaming dataframe should be written. TYPE: str | Path | None | VariableType DEFAULT: None
custom_writer	Custom writer that fully replaces Laktory's built-in write logic. Laktory manages the streaming query lifecycle (foreachBatch, trigger, checkpoint, start/await). Can be set as a plain string (func_name only) or a full CustomWriter object with func_name, func_args, and func_kwargs. Mutually exclusive with mode and merge_cdc_options. TYPE: CustomWriter | None | VariableType DEFAULT: None
databricks_data_profiling_config	Databricks Data Quality Monitor data profiling configuration TYPE: Literal[None] | VariableType DEFAULT: None
format	Format of the data files. TYPE: str | VariableType
is_quarantine	Sink used to store quarantined results from a pipeline node expectations. TYPE: bool | VariableType DEFAULT: False
merge_cdc_options	Merge options to handle input DataFrames that are Change Data Capture (CDC). Only used when MERGE mode is selected. TYPE: DataSinkMergeCDCOptions | VariableType DEFAULT: None
metadata	Table and columns metadata. TYPE: Literal[None] | VariableType DEFAULT: None
mode	Write mode. Spark OVERWRITE: Overwrite existing data. APPEND: Append contents of this DataFrame to existing data. ERROR: Throw an exception if data already exists. IGNORE: Silently ignore this operation if data already exists. Spark Streaming APPEND: Only the new rows in the streaming DataFrame/Dataset will be written to the sink. COMPLETE: All the rows in the streaming DataFrame/Dataset will be written to the sink every time there are some updates. UPDATE: Only the rows that were updated in the streaming DataFrame/Dataset will be written to the sink every time there are some updates. Polars Delta OVERWRITE: Overwrite existing data. APPEND: Append contents of this DataFrame to existing data. ERROR: Throw an exception if data already exists. IGNORE: Silently ignore this operation if data already exists. Laktory MERGE: Append, update and optionally delete records. Only supported for DELTA format. Requires cdc specification. TYPE: Literal['MERGE', 'APPEND', 'OVERWRITE', 'ERROR', 'UPDATE', 'COMPLETE', 'ERRORIFEXISTS', 'IGNORE'] | None | VariableType DEFAULT: None
path	File path on a local disk, remote storage or Databricks volume. TYPE: str | VariableType
reset_mode	Strategy used to reset this sink's data on a full refresh or a reset run. DROP: Drop the table (or delete the file/data) entirely, then recreate it on next write. TRUNCATE: Remove all rows but keep the table/schema/location intact. Only supported by tables (not views) and DELTA file sinks, not with declarative orchestrators. Ignored by shared sinks (shared), which only delete the rows of their writer. To reset only part of a table, declare the rows owned by the sink with shared.where. Can be overridden for a single run (reset_mode run parameter). TYPE: Literal['DROP', 'TRUNCATE'] | VariableType DEFAULT: 'DROP'
schema_definition	Explicit table schema used when creating the table. If not set, schema is inferred from the transformer output DataFrame. TYPE: DataFrameSchema | VariableType DEFAULT: None
shared	Options for a sink written by multiple writers: each writer owns its rows, identified by a writer column or a SQL predicate, and a full refresh of a writer only deletes its own rows. Required on every sink of a target written by several nodes or pipelines. true for the default options. See DataSinkSharedOptions. TYPE: DataSinkSharedOptions | None | VariableType DEFAULT: None
type	Source Type TYPE: Literal['FILE'] | VariableType DEFAULT: 'FILE'
writer_kwargs	Keyword arguments passed directly to dataframe backend writer. Passed to .options() method when using PySpark. TYPE: dict[str | VariableType, Any | VariableType] | VariableType DEFAULT: {}
writer_methods	DataFrame backend writer methods. TYPE: list[ReaderWriterMethod | VariableType] | VariableType DEFAULT: []

METHOD	DESCRIPTION
as_source	Generate a file data source with the same path as the sink.
create	Creates an empty Delta table at self.path if the path does not already exist.
filter_writer_rows	Keep the rows written by the writer of a shared sink - matching its writer
is_streaming	Return True if the write should use Spark Structured Streaming.
purge	Delete sink data and checkpoints
read	Read dataframe from sink.
with_writer_column	Add the writer identifier column to a DataFrame written by a shared sink, as the
write	Write dataframe into sink.

ATTRIBUTE	DESCRIPTION
ldp_auto_cdc_flow_kwargs	Keyword arguments for dp.create_auto_cdc_flow function TYPE: dict[str, str]
sdp_pre_merge_view_name	SPD view applying node transformer prior to applying CDC changes.

ldp_auto_cdc_flow_kwargs property

Keyword arguments for dp.create_auto_cdc_flow function

sdp_pre_merge_view_name property

SPD view applying node transformer prior to applying CDC changes.

as_source(as_stream=None, reader_kwargs=None, reader_methods=None)

Generate a file data source with the same path as the sink.

PARAMETER	DESCRIPTION
as_stream	If True, sink will be read as stream. TYPE: bool DEFAULT: None
reader_kwargs	Keyword arguments passed to the dataframe backend reader. DEFAULT: None
reader_methods	DataFrame backend reader methods. DEFAULT: None

RETURNS	DESCRIPTION
FileDataSource	File Data Source

Source code in laktory/models/datasinks/filedatasink.py

def as_source(
    self, as_stream: bool = None, reader_kwargs=None, reader_methods=None
) -> FileDataSource:
    """
    Generate a file data source with the same path as the sink.

    Parameters
    ----------
    as_stream:
        If `True`, sink will be read as stream.
    reader_kwargs:
        Keyword arguments passed to the dataframe backend reader.
    reader_methods:
        DataFrame backend reader methods.

    Returns
    -------
    :
        File Data Source
    """

    source = FileDataSource(
        path=self.path, format=self.format, dataframe_backend=self.dataframe_backend
    )

    if as_stream:
        source.as_stream = as_stream
    if reader_kwargs:
        source.reader_kwargs.update(reader_kwargs)
    if reader_methods:
        source.reader_methods.extend(reader_methods)

    # if self.dataframe_backend:
    #     source.dataframe_backend = self.dataframe_backend
    source.parent = self.parent

    return source

create(df=None)

Creates an empty Delta table at self.path if the path does not already exist.

Returns True if the table was created, False otherwise. Schema is taken from schema_definition if set, otherwise inferred from df.

Source code in laktory/models/datasinks/filedatasink.py

def create(self, df=None) -> bool:
    """
    Creates an empty Delta table at `self.path` if the path does not already exist.

    Returns True if the table was created, False otherwise.
    Schema is taken from `schema_definition` if set, otherwise inferred from `df`.
    """

    # Create table only for formats that supports append/insert/merge/delete modes
    if self.format.lower() not in ["delta", "parquet", "iceberg"]:
        return False

    if self.exists():
        return False

    self._update_backend_from_df(df)

    # For merge CDC sinks, delegate to _init_target() which builds the correct
    # schema (including SCD2 extra columns). Using the raw df schema here would
    # produce a table without those columns, causing the first merge to fail.
    if self.merge_cdc_options is not None:
        if df is None:
            logger.info(
                f"Schema is empty and `df` is None. Skipping DataFrame '{self.path}' creation."
            )
            return False
        native_df = df.to_native()
        self.merge_cdc_options._source_schema = native_df.schema
        self.merge_cdc_options._init_target(native_df)
        return True

    schema = self._get_create_schema(df)

    if schema is None:
        logger.info(
            f"Schema is empty and `df` is None. Skipping DataFrame '{self.path}' creation."
        )
        return False

    # TODO: Add logging of schema
    logger.info(f"Creating empty DataFrame at '{self.path}'")

    if self.dataframe_backend == DataFrameBackends.PYSPARK:
        from laktory import get_spark_session

        spark = get_spark_session()

        df_empty = spark.createDataFrame(data=[], schema=schema)
        df_empty.write.format(self.format).mode("overwrite").save(self.path)

    elif self.dataframe_backend == DataFrameBackends.POLARS:
        import polars as pl

        df_empty = pl.DataFrame(schema=schema)

        if self.format.lower() == "delta":
            df_empty.write_delta(self.path, mode="overwrite")
        elif self.format.lower() == "parquet":
            df_empty.write_parquet(self.path)
        else:
            raise NotImplementedError(
                f"Format '{self.format}' is not support for {self.dataframe_backend} backend."
            )

    else:
        raise NotImplementedError(
            f"DataFrame creation is not implemented for '{self.dataframe_backend}' backend"
        )

    return True

filter_writer_rows(df)

Keep the rows written by the writer of a shared sink - matching its writer identifier or shared.where - without the writer column: the output of the writer read back from the shared target. Returns the DataFrame unchanged if the sink is not shared.

PARAMETER	DESCRIPTION
df	DataFrame read from the sink TYPE: AnyFrame

RETURNS	DESCRIPTION
AnyFrame	DataFrame with the rows of the writer

Source code in laktory/models/datasinks/basedatasink.py

def filter_writer_rows(self, df: AnyFrame) -> AnyFrame:
    """
    Keep the rows written by the writer of a shared sink - matching its writer
    identifier or `shared.where` - without the writer column: the output of the writer
    read back from the shared target. Returns the DataFrame unchanged if the sink is not
    shared.

    Parameters
    ----------
    df:
        DataFrame read from the sink

    Returns
    -------
    :
        DataFrame with the rows of the writer
    """
    if self.shared is None:
        return df

    is_nw = isinstance(df, (nw.DataFrame, nw.LazyFrame))
    _df = df if is_nw else nw.from_native(df)
    if self.shared.uses_writer_column:
        self._check_writer_id()
        column = self.shared.column
        if column in _df.columns:
            _df = _df.filter(nw.col(column) == self.shared.writer_id).drop(column)
    else:
        _df = nw.from_native(self._filter_where_native(_df))
    return _df if is_nw else _df.to_native()

is_streaming(df=None)

Return True if the write should use Spark Structured Streaming.

Resolution order: 1. If a Narwhals-wrapped PySpark DataFrame is provided, read its native isStreaming attribute. 2. Fall back to self.as_stream (explicit sink configuration). 3. Fall back to the parent node's source as_stream flag. 4. Default to False (static write).

If both the DataFrame state and the configuration are set and they disagree, a TypeError is raised to surface the misconfiguration early.

PARAMETER	DESCRIPTION
df	Optional Narwhals DataFrame or LazyFrame. Must be passed before calling .to_native() so that the Narwhals implementation attribute is still available. DEFAULT: None

Source code in laktory/models/datasinks/basedatasink.py

def is_streaming(self, df=None) -> bool:
    """
    Return `True` if the write should use Spark Structured Streaming.

    Resolution order:
    1. If a Narwhals-wrapped PySpark DataFrame is provided, read its native
       ``isStreaming`` attribute.
    2. Fall back to ``self.as_stream`` (explicit sink configuration).
    3. Fall back to the parent node's source ``as_stream`` flag.
    4. Default to ``False`` (static write).

    If both the DataFrame state and the configuration are set and they
    disagree, a ``TypeError`` is raised to surface the misconfiguration
    early.

    Parameters
    ----------
    df:
        Optional Narwhals DataFrame or LazyFrame. Must be passed before
        calling ``.to_native()`` so that the Narwhals ``implementation``
        attribute is still available.
    """
    # Check if DataFrame is streaming
    df_is_streaming = None
    if df is not None:
        df = nw.from_native(df)
        dataframe_backend = DataFrameBackends(df.implementation)
        if dataframe_backend == DataFrameBackends.PYSPARK:
            df_is_streaming = df.to_native().isStreaming

    # Check if configured as stream from writer or source
    configured_as_stream = self.as_stream
    if configured_as_stream is None:
        node = self.parent_pipeline_node
        if node is not None and node.sources:
            configured_as_stream = node.has_streaming_source

    # Resolve conflict
    if df_is_streaming is not None and configured_as_stream is not None:
        if df_is_streaming != configured_as_stream:
            if df_is_streaming:
                raise TypeError(
                    "Sink configured as static, but received dataframe is streaming."
                )
            else:
                raise TypeError(
                    "Sink configured as stream, but received dataframe is not streaming."
                )

    is_streaming = df_is_streaming or configured_as_stream or False

    return is_streaming

purge(mode=None)

Delete sink data and checkpoints

PARAMETER	DESCRIPTION
mode	Optional override for reset_mode (DROP or TRUNCATE), taking precedence over the resolved self.reset_mode value for this call only. TYPE: Literal['DROP', 'TRUNCATE'] | None DEFAULT: None

Source code in laktory/models/datasinks/filedatasink.py

def purge(self, mode: Literal["DROP", "TRUNCATE"] | None = None):
    """
    Delete sink data and checkpoints

    Parameters
    ----------
    mode:
        Optional override for `reset_mode` (`DROP` or `TRUNCATE`), taking precedence
        over the resolved `self.reset_mode` value for this call only.
    """
    deletes_writer_rows = self._deletes_writer_rows(mode)
    reset_mode = None if deletes_writer_rows else self._get_purge_mode(mode)

    # Remove Data
    if deletes_writer_rows:
        if self.exists():
            self._check_writer_column()
            self._purge_shared_data()
    elif reset_mode == "TRUNCATE":
        if self.exists():
            self._truncate()
    elif self.exists():
        is_dir = os.path.isdir(self.path)
        if is_dir:
            logger.info(f"Deleting data dir {self.path}")
            shutil.rmtree(self.path)
        else:
            logger.info(f"Deleting data file {self.path}")
            os.remove(self.path)

    # TODO: Add support for Databricks dbfs / workspace / Volume?

    # Remove Checkpoint
    self._purge_checkpoint()

read(as_stream=None, reader_kwargs=None, reader_methods=None)

Read dataframe from sink.

PARAMETER	DESCRIPTION
as_stream	If True, dataframe read as stream. DEFAULT: None
reader_kwargs	Keyword arguments passed to the dataframe backend reader. DEFAULT: None
reader_methods	DataFrame backend reader methods. DEFAULT: None

RETURNS	DESCRIPTION
AnyFrame	DataFrame

Source code in laktory/models/datasinks/basedatasink.py

def read(self, as_stream=None, reader_kwargs=None, reader_methods=None):
    """
    Read dataframe from sink.

    Parameters
    ----------
    as_stream:
        If `True`, dataframe read as stream.
    reader_kwargs:
        Keyword arguments passed to the dataframe backend reader.
    reader_methods:
        DataFrame backend reader methods.

    Returns
    -------
    AnyFrame
        DataFrame
    """
    return self.as_source(
        as_stream=as_stream,
        reader_kwargs=reader_kwargs,
        reader_methods=reader_methods,
    ).read()

with_writer_column(df)

Add the writer identifier column to a DataFrame written by a shared sink, as the first column so that its statistics are collected. Returns the DataFrame unchanged if the sink is not shared or identifies its rows with shared.where.

PARAMETER	DESCRIPTION
df	DataFrame to be written TYPE: AnyFrame

RETURNS	DESCRIPTION
AnyFrame	DataFrame with writer identifier column

Source code in laktory/models/datasinks/basedatasink.py

def with_writer_column(self, df: AnyFrame) -> AnyFrame:
    """
    Add the writer identifier column to a DataFrame written by a shared sink, as the
    first column so that its statistics are collected. Returns the DataFrame unchanged
    if the sink is not shared or identifies its rows with `shared.where`.

    Parameters
    ----------
    df:
        DataFrame to be written

    Returns
    -------
    :
        DataFrame with writer identifier column
    """
    if self.shared is None or not self.shared.uses_writer_column:
        return df

    self._check_writer_id()
    is_nw = isinstance(df, (nw.DataFrame, nw.LazyFrame))
    _df = df if is_nw else nw.from_native(df)
    column = self.shared.column
    columns = [c for c in _df.columns if c != column]
    _df = _df.with_columns(nw.lit(self.shared.writer_id).alias(column)).select(
        [column] + columns
    )
    return _df if is_nw else _df.to_native()

write(df=None, view_definition=None, mode=None)

Write dataframe into sink.

PARAMETER	DESCRIPTION
df	Input dataframe. TYPE: AnyFrame DEFAULT: None
mode	Write mode overwrite of the sink default mode. TYPE: str DEFAULT: None
view_definition	View definition for table data sinks of VIEW type TYPE: str DEFAULT: None

Source code in laktory/models/datasinks/basedatasink.py

def write(
    self,
    df: AnyFrame = None,
    view_definition: str = None,
    mode: str = None,
) -> None:
    """
    Write dataframe into sink.

    Parameters
    ----------
    df:
        Input dataframe.
    mode:
        Write mode overwrite of the sink default mode.
    view_definition:
        View definition for table data sinks of `VIEW` type
    """

    logger.info("Write initiated.")

    if getattr(self, "table_type", None) == "VIEW":
        if view_definition is None:
            raise ValueError(f"`view_definition` for '{self._id}' is `None`")

        from laktory.models.dataframe.dataframeexpr import DataFrameExpr

        if not isinstance(view_definition, DataFrameExpr):
            view_definition = DataFrameExpr(expr=view_definition)

        if self.dataframe_backend == DataFrameBackends.PYSPARK:
            self._write_spark_view(view_definition)
        elif self.dataframe_backend == DataFrameBackends.POLARS:
            self._write_polars_view(view_definition)
        else:
            raise ValueError(
                f"DataFrame backend '{self.dataframe_backend}' is not supported"
            )
        return

    if not isinstance(df, (nw.DataFrame, nw.LazyFrame)):
        df = nw.from_native(df)
    self._update_backend_from_df(df)
    if self.shared is not None:
        self._check_writer_column()
    df = self.with_writer_column(df)

    # Custom Writer
    if self.custom_writer:
        df_native = df.to_native()

        # Special Treatment for Spark Streaming
        if (
            self.dataframe_backend == DataFrameBackends.PYSPARK
            and self.is_streaming(df=df)
        ):
            if self.checkpoint_path is None:
                raise ValueError(
                    f"Checkpoint location not specified for sink '{self._id}'"
                )
            # Build context before the foreachBatch lambda so that _parent
            # references are captured while intact. Inside foreachBatch on
            # Databricks, the lambda closure is serialized via cloudpickle
            # and _parent attributes may not survive the round-trip.
            from laktory.models.laktorycontext import LaktoryContext

            _context = LaktoryContext(
                node=self.parent_pipeline_node,
                pipeline=self.parent_pipeline,
                sink=self,
            )
            query = (
                df_native.writeStream.foreachBatch(
                    lambda batch_df, _: self.custom_writer.execute(
                        batch_df, context=_context
                    )
                )
                .trigger(availableNow=True)
                .options(checkpointLocation=self.checkpoint_path)
                .start()
            )
            query.awaitTermination()

        else:
            self.custom_writer.execute(df)

        logger.info("Write completed.")
        return

    if mode is None:
        mode = self.mode

    if mode is None and self.is_quarantine and self.is_streaming(df=df):
        # Unlike a regular sink, a streaming quarantine sink has exactly
        # one mode that is both valid and correct, so defaulting it is
        # not a guess:
        # - COMPLETE requires a streaming aggregation; a quarantine
        #   DataFrame is a plain row filter, so Spark rejects it outright
        #   ([STREAMING_OUTPUT_MODE.UNSUPPORTED_OPERATION]).
        # - UPDATE is not supported by Delta as a streaming output mode
        #   at all ([DELTA_UNSUPPORTED_OUTPUT_MODE]).
        # - MERGE is semantically wrong here: rows are typically
        #   quarantined precisely because they violate the primary
        #   keys/constraints a merge would key off (e.g. a null id).
        # That leaves APPEND as the only mode that runs and means the
        # right thing. For a *static* quarantine sink there is no such
        # single answer - OVERWRITE vs APPEND depends on whether the
        # node fully recomputes each run or ingests incrementally - so
        # mode stays required there, same as any other sink.
        mode = "APPEND"

    self._validate_mode(mode, df)
    self._validate_format()

    if mode and mode.lower() == "merge":
        self.merge_cdc_options.execute(source=df)
        logger.info("Write completed.")
        return

    if self.dataframe_backend == DataFrameBackends.PYSPARK:
        self._write_spark(df=df, mode=mode)
    elif self.dataframe_backend == DataFrameBackends.POLARS:
        self._write_polars(df=df, mode=mode)
    else:
        raise ValueError(
            f"DataFrame backend '{self.dataframe_backend}' is not supported"
        )

    logger.info("Write completed.")