"""``AbstractDataSet`` implementation to access Spark dataframes using
``pyspark`` on Apache Hive.
"""
import pickle
from copy import deepcopy
from typing import Any, Dict, List
from pyspark.sql import DataFrame, SparkSession, Window
from pyspark.sql.functions import col, lit, row_number
from kedro.io.core import AbstractDataSet, DataSetError
# pylint:disable=too-many-instance-attributes
[docs]class SparkHiveDataSet(AbstractDataSet):
"""``SparkHiveDataSet`` loads and saves Spark dataframes stored on Hive.
This data set also handles some incompatible file types such as using partitioned parquet on
hive which will not normally allow upserts to existing data without a complete replacement
of the existing file/partition.
This DataSet has some key assumptions:
- Schemas do not change during the pipeline run (defined PKs must be present for the
duration of the pipeline)
- Tables are not being externally modified during upserts. The upsert method is NOT ATOMIC
to external changes to the target table while executing.
Upsert methodology works by leveraging Spark DataFrame execution plan checkpointing.
Example adding a catalog entry with
`YAML API <https://kedro.readthedocs.io/en/stable/data/\
data_catalog.html#using-the-data-catalog-with-the-yaml-api>`_:
.. code-block:: yaml
>>> hive_dataset:
>>> type: spark.SparkHiveDataSet
>>> database: hive_database
>>> table: table_name
>>> write_mode: overwrite
Example using Python API:
::
>>> from pyspark.sql import SparkSession
>>> from pyspark.sql.types import (StructField, StringType,
>>> IntegerType, StructType)
>>>
>>> from kedro.extras.datasets.spark import SparkHiveDataSet
>>>
>>> schema = StructType([StructField("name", StringType(), True),
>>> StructField("age", IntegerType(), True)])
>>>
>>> data = [('Alex', 31), ('Bob', 12), ('Clarke', 65), ('Dave', 29)]
>>>
>>> spark_df = SparkSession.builder.getOrCreate().createDataFrame(data, schema)
>>>
>>> data_set = SparkHiveDataSet(database="test_database", table="test_table",
>>> write_mode="overwrite")
>>> data_set.save(spark_df)
>>> reloaded = data_set.load()
>>>
>>> reloaded.take(4)
"""
DEFAULT_SAVE_ARGS = {} # type: Dict[str, Any]
# pylint:disable=too-many-arguments
[docs] def __init__(
self,
database: str,
table: str,
write_mode: str = "errorifexists",
table_pk: List[str] = None,
save_args: Dict[str, Any] = None,
) -> None:
"""Creates a new instance of ``SparkHiveDataSet``.
Args:
database: The name of the hive database.
table: The name of the table within the database.
write_mode: ``insert``, ``upsert`` or ``overwrite`` are supported.
table_pk: If performing an upsert, this identifies the primary key columns used to
resolve preexisting data. Is required for ``write_mode="upsert"``.
save_args: Optional mapping of any options,
passed to the `DataFrameWriter.saveAsTable` as kwargs.
Key example of this is `partitionBy` which allows data partitioning
on a list of column names.
Other `HiveOptions` can be found here:
https://spark.apache.org/docs/latest/sql-data-sources-hive-tables.html#specifying-storage-format-for-hive-tables
Note:
For users leveraging the `upsert` functionality,
a `checkpoint` directory must be set, e.g. using
`spark.sparkContext.setCheckpointDir("/path/to/dir")`
or directly in the Spark conf folder.
Raises:
DataSetError: Invalid configuration supplied
"""
_write_modes = ["append", "error", "errorifexists", "upsert", "overwrite"]
if write_mode not in _write_modes:
valid_modes = ", ".join(_write_modes)
raise DataSetError(
f"Invalid `write_mode` provided: {write_mode}. "
f"`write_mode` must be one of: {valid_modes}"
)
if write_mode == "upsert" and not table_pk:
raise DataSetError("`table_pk` must be set to utilise `upsert` read mode")
self._write_mode = write_mode
self._table_pk = table_pk or []
self._database = database
self._table = table
self._full_table_address = f"{database}.{table}"
self._save_args = deepcopy(self.DEFAULT_SAVE_ARGS)
if save_args is not None:
self._save_args.update(save_args)
self._format = self._save_args.get("format") or "hive"
self._eager_checkpoint = self._save_args.pop("eager_checkpoint", None) or True
def _describe(self) -> Dict[str, Any]:
return dict(
database=self._database,
table=self._table,
write_mode=self._write_mode,
table_pk=self._table_pk,
partition_by=self._save_args.get("partitionBy"),
format=self._format,
)
@staticmethod
def _get_spark() -> SparkSession:
"""
This method should only be used to get an existing SparkSession
with valid Hive configuration.
Configuration for Hive is read from hive-site.xml on the classpath.
It supports running both SQL and HiveQL commands.
Additionally, if users are leveraging the `upsert` functionality,
then a `checkpoint` directory must be set, e.g. using
`spark.sparkContext.setCheckpointDir("/path/to/dir")`
"""
_spark = SparkSession.builder.getOrCreate()
return _spark
def _create_hive_table(self, data: DataFrame, mode: str = None):
_mode: str = mode or self._write_mode
data.write.saveAsTable(
self._full_table_address,
mode=_mode,
format=self._format,
**self._save_args,
)
def _load(self) -> DataFrame:
return self._get_spark().read.table(self._full_table_address)
def _save(self, data: DataFrame) -> None:
self._validate_save(data)
if self._write_mode == "upsert":
# check if _table_pk is a subset of df columns
if not set(self._table_pk) <= set(self._load().columns):
raise DataSetError(
f"Columns {str(self._table_pk)} selected as primary key(s) not found in "
f"table {self._full_table_address}"
)
self._upsert_save(data=data)
else:
self._create_hive_table(data=data)
def _upsert_save(self, data: DataFrame) -> None:
if not self._exists() or self._load().rdd.isEmpty():
self._create_hive_table(data=data, mode="overwrite")
else:
_tmp_colname = "tmp_colname"
_tmp_row = "tmp_row"
_w = Window.partitionBy(*self._table_pk).orderBy(col(_tmp_colname).desc())
df_old = self._load().select("*", lit(1).alias(_tmp_colname))
df_new = data.select("*", lit(2).alias(_tmp_colname))
df_stacked = df_new.unionByName(df_old).select(
"*", row_number().over(_w).alias(_tmp_row)
)
df_filtered = (
df_stacked.filter(col(_tmp_row) == 1)
.drop(_tmp_colname, _tmp_row)
.checkpoint(eager=self._eager_checkpoint)
)
self._create_hive_table(data=df_filtered, mode="overwrite")
def _validate_save(self, data: DataFrame):
# do not validate when the table doesn't exist
# or if the `write_mode` is set to overwrite
if (not self._exists()) or self._write_mode == "overwrite":
return
hive_dtypes = set(self._load().dtypes)
data_dtypes = set(data.dtypes)
if data_dtypes != hive_dtypes:
new_cols = data_dtypes - hive_dtypes
missing_cols = hive_dtypes - data_dtypes
raise DataSetError(
f"Dataset does not match hive table schema.\n"
f"Present on insert only: {sorted(new_cols)}\n"
f"Present on schema only: {sorted(missing_cols)}"
)
def _exists(self) -> bool:
# noqa # pylint:disable=protected-access
return (
self._get_spark()
._jsparkSession.catalog()
.tableExists(self._database, self._table)
)
def __getstate__(self) -> None:
raise pickle.PicklingError(
"PySpark datasets objects cannot be pickled "
"or serialised as Python objects."
)