HTML profiling reports from Apache Spark DataFrames
For each column the following statistics - if relevant for the column type - are presented in an interactive HTML report:
- Essentials: type, unique values, missing values
- Quantile statistics like minimum value, Q1, median, Q3, maximum, range, interquartile range
- Descriptive statistics like mean, mode, standard deviation, sum, median absolute deviation, coefficient of variation, kurtosis, skewness
- Most frequent values
All operations are done efficiently, which means that no Python UDFs or
.map() transformations are used at all; only Spark SQL's Catalyst (and the Tungsten execution engine) is used for the retrieval of all statistics.
If you are using Anaconda, you already have all the needed dependencies. So you just have to
pip install the package without dependencies (just in case pip tries to overwrite your current dependencies):
pip install --no-deps spark-df-profiling-optimus
If you don't have pandas and/or matplotlib installed:
pip install spark-df-profiling-optimus
The profile report is written in HTML5 and CSS3, which means that you may require a modern browser.
Keep in mind that you need a working Spark cluster (or a local Spark installation). The report must be created from
pyspark. To point pyspark driver to your Python environment, you must set the environment variable
PYSPARK_DRIVER_PYTHON to your python environment where spark-df-profiling is installed. For example, for Anaconda:
And then you can execute
/path/to/your/bin/pyspark to enter pyspark's CLI.
Jupyter Notebook (formerly IPython)
We recommend generating reports interactively by using the Jupyter notebook.
To use pyspark with Jupyter, you must also set
IPYTHON_OPTS is removed: the environment variable you want to set is
Now you can create a new notebook, which will run pyspark.
To use spark-df-profiling, start by loading in your Spark DataFrame, e.g. by using
# sqlContext is probably already created for you. # To load a parquet file as a Spark Dataframe, you can: df = sqlContext.read.parquet("/path/to/your/file.parquet") # And you probably want to cache it, since a lot of # operations will be done while the report is being generated: df_spark = df.cache()
To display the report in a Jupyter notebook, run:
import spark_df_profiling spark_df_profiling.ProfileReport(df_spark)
If you want to generate a HTML report file, save the ProfileReport to an object and use the
profile = spark_df_profiling.ProfileReport(df_spark) profile.to_file(outputfile="/tmp/myoutputfile.html")
- Python (
- Apache Spark (who would imagine!) -> requires Spark
- An internet connection. spark-df-profiling requires an internet connection to download the Bootstrap and JQuery libraries. You can choose to embed them in the HTML template code, should you desire.
- jinja2 (
>=2.8) -> needed for template rendering. Only needed in the Spark driver.
- matplotlib (
>=2.0) -> needed for histogram creation. Only needed in the Spark driver.
- pandas (
>=0.17.0) -> needed for internal data arrangement. Only needed in the Spark driver.
- six (
>=1.9.0) -> needed for py2/3 compatibility. Only needed in the Spark driver.