nanoevents/methods/physlite needs to distinguish between dak.Array and dak.Array._meta

[jpivarski]

This is following up from scikit-hep/awkward#3077.

In

coffea/src/coffea/nanoevents/methods/physlite.py

Lines 93 to 122 in 750b96d

	def _get_target_offsets(load_column, event_index):
	if isinstance(load_column, dask_awkward.Array) and isinstance(
	event_index, dask_awkward.Array
	):
	# wrap in map_partitions if dask arrays
	return dask_awkward.map_partitions(
	_get_target_offsets, load_column, event_index, label="get_target_offsets"
	)
	offsets = load_column.layout.offsets.data
	if isinstance(event_index, Number):
	return offsets[event_index]
	# let the necessary column optimization know that we need to load this
	# column to get the offsets
	if awkward.backend(load_column) == "typetracer":
	awkward.typetracer.touch_data(load_column)
	# necessary to stick it into the `NumpyArray` constructor
	# if typetracer is passed through
	offsets = awkward.typetracer.length_zero_if_typetracer(
	load_column.layout.offsets.data
	)
	def descend(layout, depth, **kwargs):
	if layout.purelist_depth == 1:
	return awkward.contents.NumpyArray(offsets)[layout]
	return awkward.transform(descend, event_index.layout)

load_column can have type dask_awkward.Array past the first if-guard if the event_index is not also dask_awkward.Array. Then some things are done with it that shouldn't be done with a dask_awkward.Array.

The first is ak.backend, a top-level Awkward function that hasn't had a dask-awkward overload (until PR dask-contrib/dask-awkward#498) because it would always return the same thing, "typetracer". But useless functions for the sake of conformity are worthwhile, so I'm in favor of adding dak.backend.

The next is ak.typetracer.touch_data, which is not a high-level function—not the sort of thing that we should be overriding with dask-awkward. It's a public function so that libraries that depend on Awkward (like Coffea) can use them, but it's not intended for data analysts, and that's where dask-awkward's coverage of the Awkward interface should stop. To avoid a confusing downstream error message, I've added PR scikit-hep/awkward#3079 to make Awkward complain about passing a dak.Array into ak.typetracer.touch_data, which means that the issue still has to be fixed.

The Coffea code should detect whether load_column is a dak.Array, and if so, pass load_column._meta into these hidden-but-public Awkward functions. Where, exactly, you want to do that depends on the logic of this function; I'm not going to suggest a change. I noticed that on line 115, you're again assuming that load_column is not a dak.Array (which doesn't have a layout), so someone knowledgeable about it, such as @nikoladze (the original author), should take a look.

[matthewfeickert]

I noticed that on line 115, you're again assuming that load_column is not a dak.Array (which doesn't have a layout), so someone knowledgeable about it, such as @nikoladze (the original author), should take a look.

Thanks @jpivarski! This is great. I think realistically though that @nikoladze is low on time to contribute patches, so I think this is going to be a joint effort from probably me, with @ekourlit, @alexander-held, and @kyungeonchoi giving advice (maybe @kratsg too). ATLAS still needs to find someone from the ATLAS Analysis Model Group (AMG) to be able to maintain the PHYSLITE code, but for the timelines that this needs to be dealt with, I'm going to take responsibility for this. Though I am not a good candidate to be that maintainer given my other project responsibilities.