consider removing default compressor / filters / serializer from config

[d-v-b]

Our config right now contains this logic for defining a default encoding scheme for a given data type:

zarr-python/src/zarr/core/config.py

Lines 85 to 107 in af55fcf

	"v2_default_filters": {
	"numeric": None,
	"string": [{"id": "vlen-utf8"}],
	"bytes": [{"id": "vlen-bytes"}],
	"raw": None,
	},
	"v3_default_filters": {"numeric": [], "string": [], "bytes": []},
	"v3_default_serializer": {
	"numeric": {"name": "bytes", "configuration": {"endian": "little"}},
	"string": {"name": "vlen-utf8"},
	"bytes": {"name": "vlen-bytes"},
	},
	"v3_default_compressors": {
	"numeric": [
	{"name": "zstd", "configuration": {"level": 0, "checksum": False}},
	],
	"string": [
	{"name": "zstd", "configuration": {"level": 0, "checksum": False}},
	],
	"bytes": [
	{"name": "zstd", "configuration": {"level": 0, "checksum": False}},
	],
	},

This approach is problematic because it requires dividing our data types into separate categories which are not very well defined -- is a fixed-length utf32 data type a "string" or "numeric" type?

Given the changes coming in #2874, I propose the following alteration to our approach here:

• Pull this stuff out of the config entirely.

• Confine all this logic to a single function for automatically picking a chunk encoding based on a data type + a requested chunk encoding. This function should also check for incompatibility between a data type and a requested chunk encoding. For example, if someone requests a variable-length string data type but does not specify vlen-utf8 as a serializer, then they should get a clear, early error.

These would be breaking changes, but our current approach is, IMO, unworkable.

[d-v-b]

note: any changes to the config API, e.g. those in #2874, should be preceded by deprecations wherever possible. This is supported by donfig: https://donfig.readthedocs.io/en/latest/configuration.html#deprecations.

xref: #2874 (comment)

[d-v-b]

closed via #3228