This optimizes our read performance by using NDBuffer.empty instead of NDBuffer.create with a fill value. All chunks are filled, either with the values from the Store if that chunk was initialized, or the fill value if the Store returns None (the
chunk wasn't initialized). By not memsetting values that will be overwritten anyway, things will be a bit faster.

I didn't see any documentation about CodecPipeline in the docs. Technically, users might see a behavior change, if the user's codec pipeline happened to rely on receiving an array filled with the fill_value. Do we know of anyone writing / using custom codec pipelines?

Here's a mini benchmark:

import asyncio
import statistics
import time

import zarr.api.asynchronous
import zarr.storage


async def main():
    store = zarr.storage.MemoryStore()
    array = await zarr.api.asynchronous.create_array(
        store, shape=(125_000_000,), dtype="float64", chunks=(125_000,), fill_value=1.0
    )
    times = []
    for _ in range(10):
        t0 = time.monotonic()
        await array.getitem(slice(None))
        t1 = time.monotonic()
        times.append(t1 - t0)

    print("Unitialized", statistics.median(times))

    await array.setitem(slice(None), 2.0)
    times = []
    for _ in range(10):
        t0 = time.monotonic()
        await array.getitem(slice(None))
        t1 = time.monotonic()
        times.append(t1 - t0)
    print("Initialized", statistics.median(times))


if __name__ == "__main__":
    asyncio.run(main())

main:

❯ python bench.py
Unitialized 0.6652364450274035
Initialized 0.9079880800272804

HEAD:

❯ python bench.py
Unitialized 0.5070544500194956
Initialized 0.8395943949872162

Closes #3368

TODO:

• Add unit tests and/or doctests in docstrings
• Add docstrings and API docs for any new/modified user-facing classes and functions
• New/modified features documented in docs/user-guide/*.rst
• Changes documented as a new file in changes/
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)