Difference Between Codecs and Compressors

[jmdelahanty]

I'm trying to understand what the difference between the codecs/compressors arguments in v3 and when I should use them in my code. Is there a general rule of thumb for when to specify a list of compressors vs. when to supply a list of codecs that I can reference?

[rabernat]

At the low level, filters, serializer, and compressors are all formally codecs in the V3 spec. The terminology of filters / serializer / compressors is used in the Zarr-Python API in order to:

• Provide a continuity with V2
• Be more explicit about the roles that these different filters play in the codec pipeline

Specifically, filters are "array-to_array" codecs

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

Lines 1064 to 1068 in e1990c0

	def filters(self) -> tuple[Numcodec, ...] | tuple[ArrayArrayCodec, ...]:
	"""
	Filters that are applied to each chunk of the array, in order, before serializing that
	chunk to bytes.
	"""

serializer (there can only be one) is an array-to-bytes codec

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

Lines 1080 to 1083 in e1990c0

	def serializer(self) -> ArrayBytesCodec | None:
	"""
	Array-to-bytes codec to use for serializing the chunks into bytes.
	"""

and compressors are bytes-to-bytes codecs

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

Lines 1106 to 1110 in e1990c0

	def compressors(self) -> tuple[Numcodec, ...] | tuple[BytesBytesCodec, ...]:
	"""
	Compressors that are applied to each chunk of the array. Compressors are applied in order, and after any
	filters are applied (if any are specified) and the data is serialized into bytes.
	"""

If I specify each of these, they all end up in the codecs metadata for the array, e.g.

import zarr
a = zarr.create_array(
    shape=10, dtype='f8',
    filters=[zarr.codecs.Delta()],
    serializer=zarr.codecs.BytesCodec(),
    compressors=[zarr.codecs.ZstdCodec()],
    store=zarr.storage.MemoryStore()
)
print(a.metadata.codecs)
# -> (Delta(codec_name='numcodecs.delta', codec_config={}), BytesCodec(endian=<Endian.little: 'little'>), ZstdCodec(level=0, checksum=False))

Hope this helps!

[jmdelahanty]

Thank you so much this does help!

So is it best practice then to specify each piece separately like in your final example? Or is how I've been trying to set things below sufficient? I'm trying to learn best practices for setting things up with zarr so I'm so grateful for your time!

# from my imports...
from zarr.codecs import BloscCodec

# later in the code...

# Compressor
    compressor = BloscCodec(typesize=1, cname='lz4', clevel=1, shuffle="bitshuffle")

    # Create output arrays
    roi_images = crop_group.create_array(
        'roi_images',
        shape=(total_detections, *roi_sz),
        chunks=(min(chunk_size * 4, total_detections), roi_sz[0], roi_sz[1]),
        dtype='uint8',
        overwrite=True,
        compressors=compressor
    )

O

[rabernat]

Bat you're doing is fine. There is no filter by default and the default serializer is fine.

[jmdelahanty]

Thank you for your help! Zarr is awesome.