Please do also modify _find_subtraj_first. That should help get rid of some more UUID creation.

What you have here works, but I think there's a much easier way (as opposed to the exttraj iterator, etc.) Wouldn't this do the job:

# subtraj = traj[slice(start, stop)]
# replace with
subtraj = list(traj)[slice(start, stop)]

I think all we need is an iterable of snapshots, so this would create that with very little change to the code.

A general change that you will want to do everywhere, instead of list(trajectory). Discussed in more detail in a comment, but now that I think about it, it applies to the whole changeset:

You should always try to use one of the proxy methods from Trajectory.

• get_as_proxy(idx): get a single snapshot, allowing proxy snapshots
• as_proxies(): return a list of all snapshots, which may include proxy snapshots
• iter_proxies(): iterator over snaphots, allowing proxies

When you call list(trajectory), it will call Trajectory.__iter__(), which will fully reload the snapshot into memory. This will be slow and isn't needed for a lot of analysis. (I had forgotten about this when I suggested converting to list before; only remembered when I saw changes dropping the as_proxies.)

this might possibly break custom ensembles

I'm not sure where this risk would be. I suppose there might be something weird if you subclass SequentialEnsemble in some strange way. In general, we discourage subclassing SequentialEnsemble (unless you're changing the frame assignment algorithm). Instead, custom ensembles should be subclasses of WrappedEnsemble, wrapping the appropriate SequentialEnsemble. This is what we do for TISEnsemble, MinusInterfaceEnsemble, and VisitAllStatesEnsemble, as a few examples.

The only thing I was worried about with this change was handling proxy snapshots, which you are dealing with.

As ensembles are not fully covered under coverage, please do have a really good look at this PR.

Although ensemble.py doesn't have full coverage, the sections modified here (EnsembleCache and SequentialEnsemble) are pretty close, mainly missing coverage of logging stuff in debug guards and of a few error cases. Additionally, they're tested very thoroughly, involving what is probably thousands of individual asserts. SequentialEnsemble was some very complex code to write, and I didn't trust it until I had written every test I could think of at the time.

Looks pretty good. A small optional change in a comment, as well as the following thought:

It might be worth just having a function called _get_list_traj which does this getattr/list pair of operations. I suspect CodeClimate will complain about repeated code here, and since the reason for the need is obscure enough, it might be worth it to have a single location with a clearer docstring explaining what's going on. That's optional though.

In case you want to make those changes, I will leave this open for at least 24 hours, merging no earlier than Wed 24 Feb 16:00 GMT (17:00 local). If no comment/update by then, I will merge sometime after that.