Keep the characters after the length number in char_dim_name - #11757

Thanks for the fix! The suffix was definitely getting lost, but I think the new regex goes a bit too far. A number that isn't at the end of the name is no longer treated as the length at all, so the clash protection from #10395 doesn't apply to those names anymore.

1. Name with a number in the middle, matching length (fixed by this PR)

import numpy as np
import xarray as xr

ds = xr.Dataset({"a": ("x", np.array([b"ab", b"cdef"]))})
ds.a.encoding["char_dim_name"] = "n4_chars"
ds.to_netcdf("ex1.nc")
xr.open_dataset("ex1.nc", decode_cf=False).sizes
# main: {'x': 2, 'n4': 4}
# PR:   {'x': 2, 'n4_chars': 4}

2. Two variables share the name, one is shorter (regression)

ds = xr.Dataset(
    {
        "a": ("x", np.array([b"ab", b"cdef"])),
        "b": ("x", np.array([b"ab", b"cde"])),
    }
)
ds.a.encoding["char_dim_name"] = "n4_chars"
ds.b.encoding["char_dim_name"] = "n4_chars"
ds.to_netcdf("ex2.nc")
# main: works (with a warning), dims 'n4': 4 and 'n3': 3
# PR:   ValueError: shape mismatch: objects cannot be broadcast to a single shape.
#       Mismatch is between arg 0 with shape (2, 3) and arg 1 with shape (2, 4).

3. Round trip, then shorten one variable

ds = xr.Dataset(
    {
        "a": ("x", np.array([b"ab", b"cdef"])),
        "b": ("x", np.array([b"ab", b"cdef"])),
    }
)
ds.a.encoding["char_dim_name"] = "n4_chars"
ds.b.encoding["char_dim_name"] = "n4_chars"
ds.to_netcdf("ex3.nc")
ds2 = xr.open_dataset("ex3.nc").load()
ds2["b"] = ds2.b.str[:3].astype("S")
ds2.b.encoding = {"char_dim_name": "n4_chars", "original_shape": (2, 4)}
ds2.to_netcdf("ex3b.nc")
xr.open_dataset("ex3b.nc", decode_cf=False).sizes
# main: {'x': 2, 'n4': 4, 'n3': 3}
# PR:   {'x': 2, 'n4_chars': 4, 'n4_chars3': 3}

So the question is what the dimension name should be when the length changes and the number isn't at the end. Any thoughts on what the expected behavior would be here?

Thanks for the examples, you're right that the trailing-only regex dropped the clash protection.

I think the number should still be read as the length, wherever it is, and only the characters after it should be kept. So the last number gets replaced by the real length and the suffix stays. I pushed that in cacfc61:

1. n4_chars, length 4: {'x': 2, 'n4_chars': 4} (unchanged name, no warning)
2. a and b share n4_chars, b has length 3: {'x': 2, 'n4_chars': 4, 'n3_chars': 3}, with the naming-mismatch warning
3. round trip, then shorten b: {'x': 2, 'n4_chars': 4, 'n3_chars': 3}

Only the last number is swapped, so n2_4_chars with length 3 gives n2_3_chars. I added these cases to test_CharacterArrayCoder_dim_mismatch_warnings.

Congratulations on completing your first pull request! Welcome to Xarray! We are proud of you, and hope to see you again!