@taldcroft - I got to implementing the Time shape manipulation routines, using a new private version of replicate, which can not only copy but also do other tricks. As a side benefit, this also ensures __getitem__ works properly even in the presence of attributes that have different (but broadcastable) shape.

With this in place, it will become quite easy to support min, max, and sorting.

@mhvk - cool! I did a quick scan to see what's happening, and hope to have a more detailed look in the next few days.

@taldcroft - thanks! One question to ponder: In the present PR, I'm going somewhat out of my way to keep the shape of location, _delta_ut1_utc and _delta_tdb_tt close to how they were given. I think it may actually be a bit excessive. If you think it is better, I could also rewrite such that there is only a difference between scalar values and arrays, with the latter already broadcasted to the shape of the jd1, jd2 arrays when they are defined.

@taldcroft - one more thing: now that I'm adding more methods, I'm starting to wonder whether I put the documentation for reshape, etc., in the right place. Do you think it would be better to just start a new section at the very end?

@mhvk - Re: broadcasting of Time vals like location, I agree that keeping just a scalar version or the array version broadcasted to the jd1 shape would be fine, particularly if it cleans up the code logic.

@mhvk - finished my review with relatively minor comments, this is awesome work! This is a good example of writing a (rather complicated) numpy container class as opposed to subclassing. You can start having fun filling in all the methods! 😄

Slightly OT, I wonder if it would be possible to factor things out so we don't just have a monster core.py file.

@taldcroft - thanks for the comments. I'll try to get to them shortly.

In the meantime:

Slightly OT, I wonder if it would be possible to factor things out so we don't just have a monster core.py file.

Agreed. I think we could have the different time formats separately (formats.py?) and move the double-precision routines to an utils.py or so. That would leave just the top-level Time and TimeDelta classes, plus stuff related to time scales.

@mhvk - the refactor suggestion makes sense.

@mhvk - just want to check that you're not waiting on me for this. I just had a few minor comments. I was going to try writing a TimeColumnInfo subclass in #3731 that does the right thing for stats, but then realized that the min/max funcs are not in yet. I was going to review #3681 after this PR gets merged. In the mean time this one needs rebasing.

@taldcroft - the delay is solely because of other obligations. Hope to get to this and the column info class later this week.

@taldcroft - finally got around to updating the shape-changing methods for Time. As always, getting links to work in the documentation was the hardest part. Anyway, it should be all OK now.

@taldcroft - with appveyer now working again, this passes all tests. Could you have a last look? I did end up just broadcasting dimensions for location, etc. (unless they were single elements), making the code quite a bit simpler.