Guidance on potential breaking change for test

[hmaarrfk]

We are trying to speed up certain operations in h5netcdf in h5netcdf/h5netcdf#197

One of them seems like it would break the changes introduced in:
#4879

I do not believe the issue is due to h5netcdf, rather to how h5py (and likely HDF5) handles files internally.

I adapted the problematic test, and ran it with h5py version 3.7.0

import shutil
import xarray
import os 
import h5py

example_1_path = "a.nc"
example_1_modified_path = "b.nc"

example_1 = xarray.open_dataset(
        os.path.join(os.path.dirname(xarray.__file__), "tests", "data", "example_1.nc"),
    )
example_1.to_netcdf(example_1_path, engine='h5netcdf')
example_1.rh.values += 100
example_1.to_netcdf(example_1_modified_path, engine='h5netcdf')
c = h5py.File(example_1_modified_path)
c_rh = c["/rh"][...]
assert np.array_equal(example_1.rh.values, c_rh)
c.close()

a = h5py.File(example_1_path, 'r')
a_rh = a["/rh"][...]
# a = xarray.open_dataset(example_1_path, engine='h5netcdf').load()
# a_rh = a.rh.values

# Simulate external process modifying example_1.nc while this script is running
shutil.copy(example_1_modified_path, example_1_path)

# Reopen example_1.nc (modified) as `b`; note that `a` has NOT been closed
# b = xarray.open_dataset(example_1_path, engine='h5netcdf').load()
# b_rh = b.rh.values
b = h5py.File(example_1_path, 'r')
b_rh = b["/rh"][...]

# Compare the old data
assert not np.array_equal(a_rh, b_rh)
assert a._id.fileno == b._id.fileno

# Compare newly loaded data
a_rh = a["/rh"][...]
b_rh = b["/rh"][...]
assert not np.array_equal(a_rh, b_rh)

While the first comparison passes the assertion, the second comparison fails because both values, a and b, are pointing to the new data (increased by 100).

It seems that h5py shares the same fileno.

Mostly, I find the test a little odd. Is there a flag we should be supplying to h5py to force it to open a new file handle?

[hmaarrfk]

Basically, this seems to be a quirk in h5py and I believe down to HDF5. Given the maturity of HDF5, I believe a flag likely needs to be provided to stop it from reusing the same file descriptor:

from h5py import h5f
a = h5f.open(b'a.nc', h5f.ACC_RDONLY)
b = h5f.open(b'a.nc', h5f.ACC_RDONLY)
assert a.fileno == b.fileno

[hmaarrfk]

import xarray as xr
d = xr.Dataset()
d.to_netcdf('a.nc')
from h5py import h5f
a = h5f.open(b'a.nc', h5f.ACC_RDONLY)
b = h5f.open(b'a.nc', h5f.ACC_RDONLY)
assert a.fileno == b.fileno

[hmaarrfk]

My current inclination, is that our proposed optimizations are attempting to keep references to the datasets alive and thus, HDF5 is trying to keep their cached values alive as a consequence.

https://docs.hdfgroup.org/hdf5/develop/group___f_a_p_l.html#ga034a5fc54d9b05296555544d8dd9fe89