PySpark Datasource for ROOT

Apache Spark 4 Python Datasource for reading files in the ROOT data format used in High-Energy Physics (HEP).

Author and version
Luca.Canali@cern.ch · v0.1 (Sep 2025)

Highlights

• ✅ Allows to read ROOT data using Apache Spark using a custom Spark 4 Python DataSource.
• ✅ Works with local files, directories, and globs; optional XRootD (root://) support.
• ✅ Implements partitioning and optional schema inference.
• ✅ Powered by uproot, awkward, PyArrow and Spark's Python Datasource.

Related work & acknowledgments

• The ROOT format is part of the ROOT project
• Key dependencies from scikit-hep: uproot and awkward (thanks to Jim Pivarski)
• Spark Python Datasources: Python Data Source API, Spark Python datasources, Datasource for Huggingface datasets
• SPARK-48493 - Arrow batch support for improved performance (thanks to Allison Wang)
• Notes and example notebooks on Apache Spark for Physics and a note on reading ROOT files with Spark

Install

# From PyPI
pip install pyspark-root-datasource

# Or, local for development
pip install -e .

Quick start

from pyspark.sql import SparkSession
from pyspark_root_datasource import register

spark = (SparkSession.builder
         .appName("Read ROOT via PySpark + uproot")
         .getOrCreate())

# Register the datasource (short name = "root")
register(spark)

# Get the example ROOT file (2 GB)
# xrdcp root://eospublic.cern.ch//eos/opendata/cms/derived-data/AOD2NanoAODOutreachTool/Run2012BC_DoubleMuParked_Muons.root .
# if you don't have xrdcp installed, on Linux use wget or curl -O
wget https://sparkdltrigger.web.cern.ch/sparkdltrigger/Run2012BC_DoubleMuParked_Muons.root

# Best practice: provide a schema to prune branches early
schema = "nMuon int, Muon_pt array<float>, Muon_eta array<float>, Muon_phi array<float>, Muon_mass array<float>, Muon_charge array<int>"

df = (spark.read.format("root")
      .schema(schema)
      .option("path", "/data/Run2012BC_DoubleMuParked_Muons.root")
      .option("tree", "Events")
      .option("step_size", "1000000")
      .load())

df.show(5, truncate=False)
print("Count:", df.count())

# Use schema inference
df2 = (spark.read.format("root")
       .option("path", "/data/Run2012BC_DoubleMuParked_Muons.root")
       .option("tree", "Events")
       .option("sample_rows", "1000")   # default 1000
       .load())
df2.printSchema()

Examples and tests

• Read ROOT using PySpark: read_root_file.py
• Notebook computing data from ROOT files: Dimuon_mass_spectrum.ipynb
• Run tests with pytest

Options

• "path" (required) – file path, URL, comma-separated list, directory, or glob (e.g. "/data/*.root")
• "tree" (default: "Events") – TTree name
• "step_size" (default: "1000000") – entries per Spark partition (per file)
• "num_partitions" (optional, per file) – overrides step_size
• "entry_start", "entry_stop" (optional, per file) – index bounds
• "columns" – comma-separated branch names (if not providing a Spark schema)
• "list_to32" (default: "true") – Arrow list offset width
• "extensionarray" (default: "false") – Arrow extension array support
• "cast_unsigned" (default: "true") – cast uint* → signed (Spark lacks unsigned)
• "recursive" (default: "false") – expand directories recursively
• "ext" (default: "*.root") – filter pattern when path is a directory
• "sample_rows" (default: "1000") – rows for schema inference
• "arrow_max_chunksize" (default: "0") – if >0, limit rows per Arrow RecordBatch

Reading over XRootD (root://)

# fsspec plugins for xrootd
pip install fsspec fsspec-xrootd

# XRootD client libs + Python bindings
conda install -c conda-forge xrootd

Install the extras, then:

remote_file = "root://eospublic.cern.ch//eos/opendata/cms/derived-data/AOD2NanoAODOutreachTool/Run2012BC_DoubleMuParked_Muons.root"
df = (spark.read.format("root")
      .option("path", remote_file)
      .option("tree", "Events")
      .load())
df.show(3, truncate=False)

Reading folders, globs, recursion

# All .root files in a directory (non-recursive)
df = (spark.read.format("root")
      .option("path", "/data/myfolder")
      .load())

# Recursive directory expansion
df = (spark.read.format("root")
      .option("path", "/data/myfolder")
      .option("recursive", "true")
      .load())

# Custom extension used when 'path' is a directory
df = (spark.read.format("root")
      .option("path", "/data/myfolder")
      .option("ext", "*.parquet.root")
      .load())

# Glob
df = (spark.read.format("root")
      .option("path", "/data/*/atlas/*.root")
      .load())

Tips and troubleshooting

• Prefer explicit schemas to prune early and minimize I/O.
• Tune partitioning:
  • step_size = entries per Spark partition.
  • num_partitions (per file) overrides step_size.
• Large jagged arrays benefit from reasonable step_size (e.g., 100k–1M).
• If necessary, use arrow_max_chunksize to keep batch sizes moderate for downstream stages.
• cast_unsigned=true normalizes uint* to signed widths (Spark-friendly).
• Fixed-size lists are preserved as Arrow fixed_size_list (no silent downgrade).
• XRootD errors: install both fsspec and fsspec-xrootd, and the XRootD client libs. Conda is often the smoothest:

  pip install fsspec fsspec-xrootd
  conda install -c conda-forge xrootd
  

• Tree not found: double-check .option("tree", "..."); error messages list available keys.
• Different schemas across files: ensure compatible branch types or read by subsets, then reconcile in Spark.
• Driver vs executors env mismatch: set both spark.pyspark.python and spark.pyspark.driver.python to your Python.