Skimming with coffea and writing out RNTuples takes forever to read the back up

[ikrommyd]

Reproducer using the 40-event tests/samples/nano_dy.root file from the coffea repo:

import time
import awkward as ak
import uproot
from coffea.nanoevents import NanoEventsFactory, NanoAODSchema
NanoAODSchema.warn_missing_crossrefs = False


def is_rootcompat(a):
    """Is it a flat or 1-d jagged array?"""
    t = ak.type(a)
    if isinstance(t, ak.types.ArrayType):
        if isinstance(t.content, ak.types.NumpyType):
            return True
        if isinstance(t.content, ak.types.ListType) and isinstance(t.content.content, ak.types.NumpyType):
            return True
    return False


def uproot_writeable(events):
    """Restrict to columns that uproot can write compactly"""
    out = {}
    for bname in events.fields:
        if events[bname].fields:
            out[bname] = ak.zip(
                {
                    n: ak.to_packed(ak.without_parameters(events[bname][n]))
                    for n in events[bname].fields
                    if not n.endswith("IdxG")
                    if is_rootcompat(events[bname][n])
                }
            )
        else:
            out[bname] = ak.to_packed(ak.without_parameters(events[bname]))
    return out


filename = "tests/samples/nano_dy.root"
events = NanoEventsFactory.from_root({filename: "Events"}, mode="eager").events()

events = events[ak.sum(events.Electron.pt > 10, axis=1) > 0]
with uproot.recreate("skimmedevents.root") as fout:
    fout.mkrntuple("Events", uproot_writeable(events))

print("Finished skimming")
print("Reading the skimmed file with NanoEvents")
start = time.perf_counter()
events = NanoEventsFactory.from_root({"skimmedevents.root": "Events"}, mode="eager").events()
print("Elapsed time:", time.perf_counter() - start)
print("Done")

The above snippet takes 85 seconds for me in my machine to read the skimmed file skimmedevents.root.
If I swap mkrntuple to mktree it takes half a second.

This may not be an uproot issue directly but it may be an issue with how we read RNTuples in coffea.
Opening it for investigation.

[ikrommyd]

cc @ariostas