Spark#

Tumult Analytics uses Spark as its underlying data processing framework. This topic guide covers relevant information about Spark for Tumult Analytics users.

Configuring Spark sessions#

Tumult Analytics uses Spark sessions to do all data processing operations. As long as Spark has an active session, any calls that “create” new Spark sessions use that active session.

If you want Tumult Analytics to use Spark sessions with any specific properties, then before running Tumult Analytics code, you should create that Spark session:

from pyspark.sql import SparkSession
spark = SparkSession.builder.<your options here>.getOrCreate()

As long as this session is active, Tumult Analytics will use it.

Connecting to Hive#

If you want to connect Spark to an existing Hive database, you should use the following options when creating a Spark session:

from pyspark.sql import SparkSession
spark = SparkSession.builder.<your options here>
    .config('spark.sql.warehouse.dir', '<Hive warehouse directory>')
    .enableHiveSupport()
    .getOrCreate()

To see where Hive’s warehouse directory is, you can use the Hive CLI (or its replacement, Beehive) to view the relevant configuration parameter:

> set hive.metastore.warehouse.dir;
hive.metastore.warehouse.dir=/hive/warehouse

Materialization and data cleanup#

Tumult Analytics is built on top of Tumult Core, which implements all of the differential privacy primitives that Tumult Analytics uses. Tumult Core uses a Spark database (named “tumult_temp_<time>_<uuid>”) to materialize dataframes after noise has been added. This ensures that repeated queries on a dataframe of results do not re-evaluate the query with fresh randomness.

This has a few consequences for users:

  • Queries are eagerly-evaluated, instead of lazily-evaluated.

  • Operations create a temporary database in Spark.

  • Tumult Analytics does not support multi-threaded operations, because the materialization step changes the active Spark database. (The active database is changed back to the original database at the end of the materialization step.)

Automatically cleaning up temporary data#

Tumult Core registers a cleanup function with atexit (see Python’s atexit documentation). If a Spark session is still active when the program exits normally, this cleanup function will automatically delete the materialization database.

If you wish to call spark.stop() before program exit, you should call cleanup() first. This will delete the materialization database. This function requires an active Spark session, but is otherwise safe to call at any time in a single-threaded program. (If cleanup() is called before a materialization step, Core will create a new materialization database.)

Finding and removing leftover temporary data#

The materialization database is stored as a folder in your Spark warehouse directory. If your program exits unexpectedly (for example, because it was terminated with Ctrl-C), or if the cleanup function is called without an active Spark session, this temporary database (and its associated folder) may not be deleted.

Tumult Analytics has a function to delete any of these folders in the current Spark warehouse: remove_all_temp_tables(). As long as your program is single-threaded, it is safe to call this function at any time.

You can also manually delete this database by deleting its directory from your Spark warehouse directory. (If you did not explicitly configure a Spark warehouse directory, look for a directory called spark-warehouse.) Spark represents databases as folders; the databases used for materialization will be folders named “tumult_temp_<time>_<uuid>”. Deleting the folder will delete the database.

These folders are safe to manually delete any time that your program is not running.

Performance and profiling#

All queries made with Tumult Analytics are executed by Spark. If you are having performance problems, you will probably want to look at Spark performance-tuning options.

RAM usage#

By default, Spark allocates itself 1GB of RAM (see Spark’s configuration documentation). Tumult Analytics programs often need more RAM than this. Usually, a program needs more RAM because:

  • the input data is large (10M rows or more)

  • a keyset used for a grouping operation is large (10k rows or more)

  • the output data is large (10M rows or more)

You can adjust the amount of memory available to Spark when creating your Spark session. For example, to configure Spark with 8 gigabytes of RAM, you can run this code:

spark = SparkSession.builder.config('spark.driver.memory', '8g').getOrCreate()

This only applies when running Spark on a single, local node; see Spark’s documentation for how to configure Spark to use more RAM across a cluster.

Saving results (to CSV or other formats)#

Converting large Spark dataframes (10M rows or more) to Pandas dataframes can be very resource-intensive. We recommend using pyspark.sql.DataFrame.write() to save results to file, instead of using pyspark.sql.DataFrame.toPandas() and then saving the Pandas dataframe.

For example, to save a dataframe as CSV, you can do this:

import os
import tempfile

df.write.csv(os.path.join(tempfile.mkdtemp(), 'data'))