issue_comments

@kynan I think this is fixed in #1128, which has a slightly more robust solution.

Thanks for your patience! This is a nice improvement.

I have an idea for a variation that might make for a cleaner (less dask specific) way to handle the remaining caching logic -- I'll add you a reviewer on that PR.

Awesome, thanks for your help!

On Sat, Nov 5, 2016 at 6:56 PM crusaderky notifications@github.com wrote:

roger that, getting to work :)

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/pull/1024#issuecomment-258647829, or mute the thread https://github.com/notifications/unsubscribe-auth/ABKS1mu6Gjv5ehzr-d_3gwKr8PPIgqarks5q7QmcgaJpZM4KLurN .

Anyway, I can open a new issue to discuss more about this if you think it's worth it.

Yes, please do!

@crusaderky I think we are OK going ahead here with Option D. If we do eventually extend xarray with out of core indexes, that will be done with a separate layer (not in IndexVariable).

I admit that currently xarray is perhaps not very suited for handling unstructured meshes, but IMO it has great potential (especially considering multi-index support) and I'd love to use it here.

Right now, xarray is not going to be great fit for such cases, because we already cache an index in memory for any labeled indexing operations. So at best, you could do something like ds.isel(mesh_edge=slice(int(1e6))). Maybe people already do this?

I doubt very many people are relying on this, though indeed, this would include some users of an array database we wrote at my former employer, which did not support different chunking schemes for different variables (which could make coordinate lookup very slow). CC @ToddSmall in case he has opinions here.

For out-of-core operations with labels on big unstructured meshes, you really need a generalization of the pandas.Index that doesn't need to live in memory (or maybe that lives in memory on some remote server).

I'm going to ping the mailing list for input, but I think it would be pretty safe.

On Tue, Oct 25, 2016 at 11:11 AM, crusaderky notifications@github.com wrote:

Hi Stephen, Thank you for your thinking. IMHO option D is the cleanest and safest. Could you come up with any example where it may be problematic?

On 21 Oct 2016 4:36 am, "Stephan Hoyer" notifications@github.com wrote:

I've been thinking about this... Maybe the simple, clean solution is to simply invoke compute() on all coords as soon as they are assigned to the DataArray / Dataset?

I'm nervous about eager loading, especially for non-index coordinates. They can have more than one dimension, and thus can contain a lot of data. So potentially eagerly loading non-index coordinates could break existing use cases.

On the other hand, non-index coordinates indeed checked for equality in most xarray operations (e.g., for the coordinate merge in align). So it is indeed useful not to have to recompute them all the time.

Even eagerly loading indexes is potentially problematic, if loading the index values is expensive.

So I'm conflicted: - I like the current caching behavior for coords and indexes - But I also want to avoid implicit conversions from dask to numpy, which is problematic for all the reasons you pointed out earlier

I'm going to start throwing out ideas for how to deal with this: Option A

Add two new (public?) methods, something like .load_coords() and .load_indexes(). We would then insert calls to these methods at the start of each function that uses coordinates: - .load_indexes(): reindex, reindex_like, align and sel - .load_coords(): merge and anything that calls the functions in core/merge.py (this indirectly includes Dataset.init and Dataset.setitem)

Hypothetically, we could even have options for turning this caching systematically on/off (e.g., with xarray.set_options(cache_coords=False, cache_indexes=True): ...).

Your proposal is basically an extreme version of this, where we call .load_coords() immediately after constructing every new object.

Advantages: - It's fairly predictable when caching happens (especially if we opt for calling .load_cords() immediately, as you propose). - Computing variables is all done at once, which is much more performant than what we currently do, e.g., loading variables as needed for .equals() checks in merge_variables one at a time.

Downsides: - Caching is more aggressive than necessary -- we cache indexes even if that coord isn't actually indexed.

Option B

Like Option A, but someone infer the full set of variables that need to be cached (e.g., in a .merge() operation) before it's actually done. This seems hard, but maybe is possible using a variation of merge_variables.

This solves the downside of A, but diminishes the predictability. We're basically back to how things work now. Option C

Cache dask.array in IndexVariable but not Variable. This preserves performance for repeated indexing, because the hash table behind the pandas.Index doesn't get thrown away.

Advantages: - Much simpler and easier to implement than the alternatives. - Implicit conversions are greatly diminished.

Downsides: - Non-index coordinates get thrown away after being evaluated once. If you're doing lots of operations of the form [ds + other for ds in datasets] where ds and other has conflicting coordinates this would probably make you unhappy.

Option D

Load the contents of an IndexVariable immediately and eagerly. They no longer cache data or use lazy loading.

This has the most predictable performance, but might cause trouble for

some edge use cases?

I need to think about this a little more, but right now I am leaning towards Option C or D.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/pull/1024#issuecomment-255286001, or mute the thread https://github.com/notifications/unsubscribe- auth/AF7OMLBh4eDuKRNv0x5HwRie_yaGh0Yzks5q2DMjgaJpZM4KLurN .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/pull/1024#issuecomment-256114879, or mute the thread https://github.com/notifications/unsubscribe-auth/ABKS1jUaNUCxHlCx86P4JjbhsLA99ZIqks5q3kZYgaJpZM4KLurN .

I've been thinking about this... Maybe the simple, clean solution is to simply invoke compute() on all coords as soon as they are assigned to the DataArray / Dataset?

I'm nervous about eager loading, especially for non-index coordinates. They can have more than one dimension, and thus can contain a lot of data. So potentially eagerly loading non-index coordinates could break existing use cases.

On the other hand, non-index coordinates indeed checked for equality in most xarray operations (e.g., for the coordinate merge in align). So it is indeed useful not to have to recompute them all the time.

Even eagerly loading indexes is potentially problematic, if loading the index values is expensive.

So I'm conflicted: - I like the current caching behavior for coords and indexes - But I also want to avoid implicit conversions from dask to numpy, which is problematic for all the reasons you pointed out earlier

I'm going to start throwing out ideas for how to deal with this:

Option A

Add two new (public?) methods, something like .load_coords() and .load_indexes(). We would then insert calls to these methods at the start of each function that uses coordinates: - .load_indexes(): reindex, reindex_like, align and sel - .load_coords(): merge and anything that calls the functions in core/merge.py (this indirectly includes Dataset.__init__ and Dataset.__setitem__)

Hypothetically, we could even have options for turning this caching systematically on/off (e.g., with xarray.set_options(cache_coords=False, cache_indexes=True): ...).

Your proposal is basically an extreme version of this, where we call .load_coords() immediately after constructing every new object.

Advantages: - It's fairly predictable when caching happens (especially if we opt for calling .load_cords() immediately, as you propose). - Computing variables is all done at once, which is much more performant than what we currently do, e.g., loading variables as needed for .equals() checks in merge_variables one at a time.

Downsides: - Caching is more aggressive than necessary -- we cache indexes even if that coord isn't actually indexed.

Option B

Like Option A, but someone infer the full set of variables that need to be cached (e.g., in a .merge() operation) before it's actually done. This seems hard, but maybe is possible using a variation of merge_variables.

This solves the downside of A, but diminishes the predictability. We're basically back to how things work now.

Option C

Cache dask.array in IndexVariable but not Variable. This preserves performance for repeated indexing, because the hash table behind the pandas.Index doesn't get thrown away.

Advantages: - Much simpler and easier to implement than the alternatives. - Implicit conversions are greatly diminished.

Downsides: - Non-index coordinates get thrown away after being evaluated once. If you're doing lots of operations of the form [ds + other for ds in datasets] where ds and other has conflicting coordinates this would probably make you unhappy.

Option D

Load the contents of an IndexVariable immediately and eagerly. They no longer cache data or use lazy loading.

This has the most predictable performance, but might cause trouble for some edge use cases?

I need to think about this a little more, but right now I am leaning towards Option C or D.