Soda Spark

Data testing, monitoring, and profiling for Spark Dataframes.

Soda Spark is an extension of Soda SQL that allows you to run Soda SQL functionality programmatically on a Spark data frame.

Soda SQL is an open-source command-line tool. It utilizes user-defined input to prepare SQL queries that run tests on tables in a data warehouse to find invalid, missing, or unexpected data. When tests fail, they surface "bad" data that you can fix to ensure that downstream analysts are using "good" data to make decisions.

Requirements

Soda Spark has the same requirements as soda-sql-spark.

Install

From your shell, execute the following command.

$ pip install soda-spark

Use

From your Python prompt, execute the following commands.

>>> from pyspark.sql import DataFrame, SparkSession
>>> from sodaspark import scan
>>>
>>> spark_session = SparkSession.builder.getOrCreate()
>>>
>>> id = "a76824f0-50c0-11eb-8be8-88e9fe6293fd"
>>> df = spark_session.createDataFrame([
...	   {"id": id, "name": "Paula Landry", "size": 3006},
...	   {"id": id, "name": "Kevin Crawford", "size": 7243}
... ])
>>>
>>> scan_definition = ("""
... table_name: demodata
... metrics:
... - row_count
... - max
... - min_length
... tests:
... - row_count > 0
... columns:
...   id:
...     valid_format: uuid
...     tests:
...     - invalid_percentage == 0
... sql_metrics:
... - sql: |
...     SELECT sum(size) as total_size_us
...     FROM demodata
...     WHERE country = 'US'
...   tests:
...   - total_size_us > 5000
... """)
>>> scan_result = scan.execute(scan_definition, df)
>>>
>>> scan_result.measurements  # doctest: +ELLIPSIS
[Measurement(metric='schema', ...), Measurement(metric='row_count', ...), ...]
>>> scan_result.test_results  # doctest: +ELLIPSIS
[TestResult(test=Test(..., expression='row_count > 0', ...), passed=True, skipped=False, ...)]
>>>

Or, use a scan YAML file

>>> scan_yml = "static/demodata.yml"
>>> scan_result = scan.execute(scan_yml, df)
>>>
>>> scan_result.measurements  # doctest: +ELLIPSIS
[Measurement(metric='schema', ...), Measurement(metric='row_count', ...), ...]
>>>

See the scan result object for all attributes and methods.

Or, return Spark data frames:

>>> measurements, test_results, errors = scan.execute(scan_yml, df, as_frames=True)
>>>
>>> measurements  # doctest: +ELLIPSIS
DataFrame[metric: string, column_name: string, value: string, ...]
>>> test_results  # doctest: +ELLIPSIS
DataFrame[test: struct<...>, passed: boolean, skipped: boolean, values: map<string,string>, ...]
>>>

See the _to_data_frame functions in the scan.py to see how the conversion is done.

Send results to Soda cloud

Send the scan result to Soda cloud.

>>> import os
>>> from sodasql.soda_server_client.soda_server_client import SodaServerClient
>>>
>>> soda_server_client = SodaServerClient(
...     host="cloud.soda.io",
...     api_key_id=os.getenv("API_PUBLIC"),
...     api_key_secret=os.getenv("API_PRIVATE"),
... )
>>> scan_result = scan.execute(scan_yml, df, soda_server_client=soda_server_client)
>>>

Understand

Under the hood soda-spark does the following.

1. Setup the scan
   • Use the Spark dialect
   • Use Spark session as warehouse connection
2. Create (or replace) global temporary view for the Spark data frame
3. Execute the scan on the temporary view