Fix hashing

[ElliottKasoar]

• Fixes the hasher definition, as discussed in the issue.

  • I've removed the second creation of a hasher object, which means the overall hash is similar to the erroneous behaviour before, in that the structure hash contributes to the final hash.

• Also updates the deprecated utcnow, and tidies the code slightly.

Edit: hash_strucutre seems fixed in test_write_and_read, but actually I'm not sure the example in the issue quite works yet, so I'll leave this as a draft for now.

[ElliottKasoar]

Ah, so @stenczelt, while this does fix the hash_structure example currently in the tests, the example in the linked issue still fails, which must be because the (float) positions read in by ASE differ to the original random numbers:

  comparison failed. Mismatched elements: 8 / 9:
  Max absolute difference: 4.723046176380308e-09
  Max relative difference: 2.2429495787383598e-08
  Index  | Obtained   | Expected                     
  (0, 0) | 0.77395605 | 0.7739560485559633 ± 1.0e-10 
  (0, 1) | 0.43887844 | 0.4388784397520523 ± 1.0e-10 
  (1, 0) | 0.69736803 | 0.6973680290593639 ± 1.0e-10 
  (1, 1) | 0.09417735 | 0.09417734788764953 ± 1.0e-10
  (1, 2) | 0.97562235 | 0.9756223516367559 ± 1.0e-10 
  (2, 0) | 0.7611397  | 0.761139701990353 ± 1.0e-10  
  (2, 1) | 0.78606431 | 0.7860643052769538 ± 1.0e-10 
  (2, 2) | 0.12811363 | 0.12811363267554587 ± 1.0e-10

Since test_write_and_read only looks at integer positions, this doesn't get picked up.

Do we want to try truncating the precision used in the hash to try to fix this?

It feels slightly dangerous, but I'm not sure what the alternative is, unless we accept that this is a limitation.

[stenczelt]

I agree it would be dangerous to truncate positions, especially since then it's not really a hash or checksum of the object but turning into some similarity metric (albeit a poor one).
As the ASE write & read ops are apparently destructive to some of the structure, we may not be testing what we are meaning to test, so can also change the test. Or perhaps if we have control over the float formatting of ASE then we could force it to not introduce any difference?

[ElliottKasoar]

I agree it would be dangerous to truncate positions, especially since then it's not really a hash or checksum of the object but turning into some similarity metric (albeit a poor one). As the ASE write & read ops are apparently destructive to some of the structure, we may not be testing what we are meaning to test, so can also change the test. Or perhaps if we have control over the float formatting of ASE then we could force it to not introduce any difference?

For some file formats e.g. castep-cell there's a precision option, but I'm not sure we can control it for (ext)xyz files, or more generally.

Perhaps we stick with this for now as a starting point for fixing hashing, since this fixes a source of similar errors?

#118 isn't resolved, but I think the existing tests do all work in the way that you would expect, including the read/write example - I don't think it's wrong to test that the hash is preserved in that example, as long as we're aware that we can't guarantee it'll always be preserved for now.

[stenczelt]

Happy with the code as per discussed over the last couple weeks, see a short ask regarding the tests.