BUG: (GH4708) A zero length series written to HDF cannot be read back.

[prossahl]

closes #4708 by spotting when the sentinel single value has been written. This was not being done for the index which resulted in a values/index length miss-match.

[jreback]

can you add a release notes entry under bug fixes, sub-heading HDFStore (which is already there), otherwise looks good

[jreback]

@prossahl you need to handle the empty M8[ns] case; it fails because that case is not caught

[jreback]

@prossahl want to finish this up?

[jreback]

@prossahl ping

[prossahl]

Still working on this. Failing Travis builds can not be reproduced locally.

[jreback]

Your test should be this; we don't try to serialize numpy arrays per se
and it DOES fail for me, the datetime64[ns] is not being caught (as empty)

    def test_empty_series(self):
        for dtype in [np.int64, np.float64, np.object, 'M8[ns]']:
            s = Series(dtype=dtype)
            self._check_roundtrip(s, tm.assert_series_equal)

[prossahl]

Travis build is now green with libhdf5 versions 1.8.4 as well as later ones.

[jreback]

• can you add m8[ns] to the tests as well (you convert it but don't test it).
• pls add a one-liner release note (doc/source/release.rst) (ref the original issue 4708), pls put with the other HDFStore bug fixes
• pls squash down commits to 1

and we can get this in

thanks!

[jreback]

thanks for the PR!