fugue_spark#

fugue_spark.dataframe#

class fugue_spark.dataframe.SparkDataFrame(df=None, schema=None)[source]#

Bases: DataFrame

DataFrame that wraps Spark DataFrame. Please also read the DataFrame Tutorial to understand this Fugue concept

Parameters

df (Any) – pyspark.sql.DataFrame
schema (Any) – Schema like object or pyspark.sql.types.StructType, defaults to None.

Note

You should use fugue_spark.execution_engine.SparkExecutionEngine.to_df() instead of construction it by yourself.
If schema is set, then there will be type cast on the Spark DataFrame if the schema is different.

property alias: str#

alter_columns(columns)[source]#

Change column types

Parameters: columns (Any) – Schema like object, all columns should be contained by the dataframe schema
Returns: a new dataframe with altered columns, the order of the original schema will not change
Return type: DataFrame

as_array(columns=None, type_safe=False)[source]#

Convert to 2-dimensional native python array

Parameters

columns (Optional[List[str]]) – columns to extract, defaults to None
type_safe (bool) – whether to ensure output conforms with its schema, defaults to False

Returns

2-dimensional native python array

Return type

List[Any]

Note

If type_safe is False, then the returned values are ‘raw’ values.

as_array_iterable(columns=None, type_safe=False)[source]#

Convert to iterable of native python arrays

Parameters

columns (Optional[List[str]]) – columns to extract, defaults to None
type_safe (bool) – whether to ensure output conforms with its schema, defaults to False

Returns

iterable of native python arrays

Return type

Iterable[Any]

Note

If type_safe is False, then the returned values are ‘raw’ values.

as_local_bounded()[source]#

Convert this dataframe to a LocalBoundedDataFrame

Return type: LocalBoundedDataFrame

as_pandas()[source]#

Convert to pandas DataFrame

Return type: DataFrame

count()[source]#

Get number of rows of this dataframe

Return type: int

property empty: bool#: Whether this dataframe is empty

head(n, columns=None)[source]#

Get first n rows of the dataframe as a new local bounded dataframe

Parameters

n (int) – number of rows
columns (Optional[List[str]]) – selected columns, defaults to None (all columns)

Returns

a local bounded dataframe

Return type

LocalBoundedDataFrame

property is_bounded: bool#: Whether this dataframe is bounded

property is_local: bool#: Whether this dataframe is a local Dataset

property native: DataFrame#

The wrapped Spark DataFrame

Return type: pyspark.sql.DataFrame

native_as_df()[source]#

The dataframe form of the native object this Dataset class wraps. Dataframe form means the object contains schema information. For example the native an ArrayDataFrame is a python array, it doesn’t contain schema information, and its native_as_df should be either a pandas dataframe or an arrow dataframe.

Return type: DataFrame

property num_partitions: int#: Number of physical partitions of this dataframe. Please read the Partition Tutorial

peek_array()[source]#

Peek the first row of the dataframe as array

Raises: FugueDatasetEmptyError – if it is empty
Return type: List[Any]

rename(columns)[source]#

Rename the dataframe using a mapping dict

Parameters: columns (Dict[str, str]) – key: the original column name, value: the new name
Returns: a new dataframe with the new names
Return type: DataFrame

fugue_spark.execution_engine#

class fugue_spark.execution_engine.SparkExecutionEngine(spark_session=None, conf=None)[source]#

Bases: ExecutionEngine

The execution engine based on SparkSession.

Please read the ExecutionEngine Tutorial to understand this important Fugue concept

Parameters

spark_session (Optional[SparkSession]) – Spark session, defaults to None to get the Spark session by getOrCreate()
conf (Any) – Parameters like object defaults to None, read the Fugue Configuration Tutorial to learn Fugue specific options

broadcast(df)[source]#

Broadcast the dataframe to all workers for a distributed computing framework

Parameters: df (DataFrame) – the input dataframe
Returns: the broadcasted dataframe
Return type: SparkDataFrame

create_default_map_engine()[source]#

Default MapEngine if user doesn’t specify

Return type: MapEngine

create_default_sql_engine()[source]#

Default SQLEngine if user doesn’t specify

Return type: SQLEngine

distinct(df)[source]#

Equivalent to SELECT DISTINCT * FROM df

Parameters: df (DataFrame) – dataframe
Returns: [description]
Return type: DataFrame

dropna(df, how='any', thresh=None, subset=None)[source]#

Drop NA recods from dataframe

Parameters

df (DataFrame) – DataFrame
how (str) – ‘any’ or ‘all’. ‘any’ drops rows that contain any nulls. ‘all’ drops rows that contain all nulls.
thresh (Optional[int]) – int, drops rows that have less than thresh non-null values
subset (Optional[List[str]]) – list of columns to operate on

Returns

DataFrame with NA records dropped

Return type

DataFrame

fillna(df, value, subset=None)[source]#

Fill NULL, NAN, NAT values in a dataframe

Parameters

df (DataFrame) – DataFrame
value (Any) – if scalar, fills all columns with same value. if dictionary, fills NA using the keys as column names and the values as the replacement values.
subset (Optional[List[str]]) – list of columns to operate on. ignored if value is a dictionary

Returns

DataFrame with NA records filled

Return type

DataFrame

property fs: FileSystem#: File system of this engine instance

get_current_parallelism()[source]#

Get the current number of parallelism of this engine

Return type: int

intersect(df1, df2, distinct=True)[source]#

Intersect df1 and df2

Parameters

df1 (DataFrame) – the first dataframe
df2 (DataFrame) – the second dataframe
distinct (bool) – true for INTERSECT (== INTERSECT DISTINCT), false for INTERSECT ALL

Returns

the unioned dataframe

Return type

DataFrame

Note

Currently, the schema of df1 and df2 must be identical, or an exception will be thrown.

property is_distributed: bool#: Whether this engine is a distributed engine

property is_spark_connect: bool#

join(df1, df2, how, on=None)[source]#

Join two dataframes

Parameters

df1 (DataFrame) – the first dataframe
df2 (DataFrame) – the second dataframe
how (str) – can accept semi, left_semi, anti, left_anti, inner, left_outer, right_outer, full_outer, cross
on (Optional[List[str]]) – it can always be inferred, but if you provide, it will be validated against the inferred keys.

Returns

the joined dataframe

Return type

DataFrame

Note

Please read get_join_schemas()

load_df(path, format_hint=None, columns=None, **kwargs)[source]#

Load dataframe from persistent storage

Parameters

path (Union[str, List[str]]) – the path to the dataframe
format_hint (Optional[Any]) – can accept parquet, csv, json, defaults to None, meaning to infer
columns (Optional[Any]) – list of columns or a Schema like object, defaults to None
kwargs (Any) – parameters to pass to the underlying framework

Returns

an engine compatible dataframe

Return type

DataFrame

For more details and examples, read Zip & Comap.

property log: Logger#: Logger of this engine instance

persist(df, lazy=False, **kwargs)[source]#

Force materializing and caching the dataframe

Parameters

df (DataFrame) – the input dataframe
lazy (bool) – True: first usage of the output will trigger persisting to happen; False (eager): persist is forced to happend immediately. Default to False
kwargs (Any) – parameter to pass to the underlying persist implementation

Returns

the persisted dataframe

Return type

SparkDataFrame

Note

persist can only guarantee the persisted dataframe will be computed for only once. However this doesn’t mean the backend really breaks up the execution dependency at the persisting point. Commonly, it doesn’t cause any issue, but if your execution graph is long, it may cause expected problems for example, stack overflow.

register(df, name)[source]#

Returns the subclass, to allow usage as a class decorator.

Parameters

df (DataFrame) –
name (str) –

Return type

SparkDataFrame

repartition(df, partition_spec)[source]#

Partition the input dataframe using partition_spec.

Parameters

df (DataFrame) – input dataframe
partition_spec (PartitionSpec) – how you want to partition the dataframe

Returns

repartitioned dataframe

Return type

DataFrame

Note

Before implementing please read the Partition Tutorial

sample(df, n=None, frac=None, replace=False, seed=None)[source]#

Sample dataframe by number of rows or by fraction

Parameters

df (DataFrame) – DataFrame
n (Optional[int]) – number of rows to sample, one and only one of n and frac must be set
frac (Optional[float]) – fraction [0,1] to sample, one and only one of n and frac must be set
replace (bool) – whether replacement is allowed. With replacement, there may be duplicated rows in the result, defaults to False
seed (Optional[int]) – seed for randomness, defaults to None

Returns

sampled dataframe

Return type

DataFrame

save_df(df, path, format_hint=None, mode='overwrite', partition_spec=None, force_single=False, **kwargs)[source]#

Save dataframe to a persistent storage

Parameters

df (DataFrame) – input dataframe
path (str) – output path
format_hint (Optional[Any]) – can accept parquet, csv, json, defaults to None, meaning to infer
mode (str) – can accept overwrite, append, error, defaults to “overwrite”
partition_spec (Optional[PartitionSpec]) – how to partition the dataframe before saving, defaults to empty
force_single (bool) – force the output as a single file, defaults to False
kwargs (Any) – parameters to pass to the underlying framework

Return type

None

For more details and examples, read Load & Save.

property spark_session: SparkSession#

Returns: The wrapped spark session
Return type: pyspark.sql.SparkSession

subtract(df1, df2, distinct=True)[source]#

df1 - df2

Parameters

df1 (DataFrame) – the first dataframe
df2 (DataFrame) – the second dataframe
distinct (bool) – true for EXCEPT (== EXCEPT DISTINCT), false for EXCEPT ALL

Returns

the unioned dataframe

Return type

DataFrame

Note

Currently, the schema of df1 and df2 must be identical, or an exception will be thrown.

take(df, n, presort, na_position='last', partition_spec=None)[source]#

Get the first n rows of a DataFrame per partition. If a presort is defined, use the presort before applying take. presort overrides partition_spec.presort. The Fugue implementation of the presort follows Pandas convention of specifying NULLs first or NULLs last. This is different from the Spark and SQL convention of NULLs as the smallest value.

Parameters

df (DataFrame) – DataFrame
n (int) – number of rows to return
presort (str) – presort expression similar to partition presort
na_position (str) – position of null values during the presort. can accept first or last
partition_spec (Optional[PartitionSpec]) – PartitionSpec to apply the take operation

Returns

n rows of DataFrame per partition

Return type

DataFrame

to_df(df, schema=None)[source]#

Convert a data structure to SparkDataFrame

Parameters

data – DataFrame, pyspark.sql.DataFrame, pyspark.RDD, pandas DataFrame or list or iterable of arrays
schema (Optional[Any]) – Schema like object or pyspark.sql.types.StructType defaults to None.
df (Any) –

Returns

engine compatible dataframe

Return type

SparkDataFrame

Note

if the input is already SparkDataFrame, it should return itself
For RDD, list or iterable of arrays, schema must be specified
When schema is not None, a potential type cast may happen to ensure the dataframe’s schema.
all other methods in the engine can take arbitrary dataframes and call this method to convert before doing anything

union(df1, df2, distinct=True)[source]#

Join two dataframes

Parameters

df1 (DataFrame) – the first dataframe
df2 (DataFrame) – the second dataframe
distinct (bool) – true for UNION (== UNION DISTINCT), false for UNION ALL

Returns

the unioned dataframe

Return type

DataFrame

Note

Currently, the schema of df1 and df2 must be identical, or an exception will be thrown.

class fugue_spark.execution_engine.SparkMapEngine(execution_engine)[source]#

Bases: MapEngine

Parameters: execution_engine (ExecutionEngine) –

property is_distributed: bool#: Whether this engine is a distributed engine

property is_spark_connect: bool#: Whether the spark session is created by spark connect

map_dataframe(df, map_func, output_schema, partition_spec, on_init=None, map_func_format_hint=None)[source]#

Apply a function to each partition after you partition the dataframe in a specified way.

Parameters

df (DataFrame) – input dataframe
map_func (Callable[[PartitionCursor, LocalDataFrame], LocalDataFrame]) – the function to apply on every logical partition
output_schema (Any) – Schema like object that can’t be None. Please also understand why we need this
partition_spec (PartitionSpec) – partition specification
on_init (Optional[Callable[[int, DataFrame], Any]]) – callback function when the physical partition is initializaing, defaults to None
map_func_format_hint (Optional[str]) – the preferred data format for map_func, it can be pandas, pyarrow, etc, defaults to None. Certain engines can provide the most efficient map operations based on the hint.

Returns

the dataframe after the map operation

Return type

DataFrame

Note

Before implementing, you must read this to understand what map is used for and how it should work.

class fugue_spark.execution_engine.SparkSQLEngine(execution_engine)[source]#

Bases: SQLEngine

Spark SQL execution implementation.

Parameters: execution_engine (ExecutionEngine) – it must be SparkExecutionEngine
Raises: ValueError – if the engine is not SparkExecutionEngine

property dialect: Optional[str]#

property execution_engine_constraint: Type[ExecutionEngine]#

This defines the required ExecutionEngine type of this facet

Returns: a subtype of ExecutionEngine

property is_distributed: bool#: Whether this engine is a distributed engine

select(dfs, statement)[source]#

Execute select statement on the sql engine.

Parameters

dfs (DataFrames) – a collection of dataframes that must have keys
statement (StructuredRawSQL) – the SELECT statement using the dfs keys as tables.

Returns

result of the SELECT statement

Return type

DataFrame

Examples

dfs = DataFrames(a=df1, b=df2)
sql_engine.select(
    dfs,
    [(False, "SELECT * FROM "),
     (True,"a"),
     (False," UNION SELECT * FROM "),
     (True,"b")])

Note

There can be tables that is not in dfs. For example you want to select from hive without input DataFrames:

>>> sql_engine.select(DataFrames(), "SELECT * FROM hive.a.table")

fugue_spark.ibis_engine#

class fugue_spark.ibis_engine.SparkIbisEngine(execution_engine)[source]#

Bases: IbisEngine

Parameters: execution_engine (ExecutionEngine) –

select(dfs, ibis_func)[source]#

Execute the ibis select expression.

Parameters

dfs (DataFrames) – a collection of dataframes that must have keys
ibis_func (Callable[[BaseBackend], TableExpr]) – the ibis compute function

Returns

result of the ibis function

Return type

DataFrame

Note

This interface is experimental, so it is subjected to change.

fugue_spark#

fugue_spark.dataframe#

fugue_spark.execution_engine#

fugue_spark.ibis_engine#

fugue_spark.registry#