Streaming ROOT files over XRootD (uproot)

Read ROOT files — CAFs, ntuples, LArCV, … — straight from dCache over the XRootD protocol with uproot. uproot fetches only the branches and baskets you ask for, so you stream just the bytes you need: no local copy and no dependency on the /pnfs mount. Works anywhere with a network connection, including the Elastic Analysis Facility (EAF).

This is the ROOT counterpart of Streaming HDF5 over XRootD.

1. Install with pixi

# inside a pixi project (run `pixi init` first if you need one)
pixi add uproot fsspec-xrootd

fsspec-xrootd provides the root:// transport (via fsspec); uproot uses it automatically. Run Python with pixi run python ... (or pixi shell).

2. Get a bearer token

htgettoken -a htvaultprod.fnal.gov -i sbnd

The XRootD client picks it up automatically. Re-run it if you hit Auth failed: No protocols left to try.

3. Get a file’s XRootD URL

For a SAM-catalogued file:

samweb get-file-access-url <filename> --schema=root

or with ifdh (also works on a /pnfs path):

ifdh getUrl <filename> root

Either prints e.g. root://fndca1.fnal.gov:1094//pnfs/fnal.gov/usr/sbn/.../file.root.

4. Open and read (stream)

import uproot

url = "root://fndca1.fnal.gov:1094//pnfs/fnal.gov/usr/sbn/.../file.root"

with uproot.open(url) as f:
    print(f.keys(cycle=False))         # objects in the file (TTrees, ...)

    tree = f["opflash_tpc0_tree"]      # pick a TTree (use your own tree name)
    print(tree.num_entries)
    tree.show()                        # list branches and their types

    # read only the branches you name -> only those are streamed
    arr = tree.arrays(["_run", "_subrun", "_event"])
    print(arr["_event"])

    # a single branch; jagged (per-event) data comes back as an awkward Array
    times = tree["_flash_v._time"].array()

5. Stream large files in chunks

For files too big to hold in memory, uproot.iterate reads a few branches at a time, batch by batch:

import uproot

url = "root://fndca1.fnal.gov:1094//pnfs/fnal.gov/usr/sbn/.../file.root"

for batch in uproot.iterate({url: "opflash_tpc0_tree"},
                            ["_event", "_flash_v._time"],
                            step_size=1000):
    ...   # process `batch` (an awkward Array holding this chunk)

Note: custom C++ objects (LArCV / art)

uproot reads standard ROOT types and many custom classes, but it cannot deserialize objects written with ROOT memberwise streaming. In a LArCV file the std::vector<larcv::Particle> inside the particle_* / cluster3d_* trees raises NotImplementedError: memberwise serialization. Their event metadata (_run, _subrun, _event) and simpler products (opflash_*, crthit_*) read fine.

For the full SPINE particle/interaction content, read the SPINE HDF5 output instead — see Streaming HDF5 over XRootD.

Streaming ROOT files over XRootD with uproot (EAF)