issue_comments

Just noticed this issue; people needing to do this sort of thing might want to look at stackstac (especially playing with the chunks= parameter) or odc-stac for loading the data. The graph will be cleaner than what you'd get from xr.concat([xr.open_rasterio(...) for ...]).

still appears to "over-eagerly" load more than just what is being worked on

FYI, this is basically expected behavior for distributed, see: * https://github.com/dask/distributed/issues/5223 * https://github.com/dask/distributed/issues/5555

Hi @shaprann, I haven't re-visited this exact workflow recently, but one really good option (if you can manage the intermediate storage cost) would be to try to use new tools like http://github.com/pangeo-data/rechunker to pre-process and prepare your data archive prior to analysis.

This particular use case is extremely common when working with spatio-temporal data. Can anyone suggest a good workaround for this?

Has there been any progress on this issue? I am bumping into the same problem.

Thanks for all the suggestions!

An update from when I originally posted this. Aligning with @shoyer, @darothen and @scottyhq 's comments, we've tested the code using cloud-optimized geoTIF files and regular geoTIFs, and it does perform better with the cloud-optimized form, though still appears to "over-eagerly" load more than just what is being worked on. With the cloud-optimized form, performance is much better when we specify the chunking strategy on the initial open_rasterio and it aligns with the chunk sizes. e.g. rasterlist = [xr.open_rasterio(x, chunks={'x': 256, 'y': 256}) for x in filelist] vs. rasterlist = [xr.open_rasterio(x, chunks={'x': None, 'y': None}) for x in filelist]

The result is a larger task graph (and much more time spent developing the task graph) but more cases where we don't run into memory problems. There still appears to be a lot more memory used than I expect, but am actively working on exploring options.

We've also noticed better performance using a k8s Dask cluster distributed across multiple "independent" workers as opposed to using a LocalCluster on a single large machine. As in, with the distributed cluster the "myfunction" (fit) operation starts happening on chunks well before the entire dataset is loaded, whereas in the LocalCluster it still tends not to begin until all chunks have been loaded in. Not exactly sure what would cause that...

I'm intrigued by @shoyer 's last suggestion of an "intermediate" chunking step. Will test that and potentially the manual iteration over the tiles. Thanks for all the suggestions and thoughts!

Can you provide a gdalinfo of one of the GeoTiffs? I'm still working on some documentation for use-cases with cloud-optimized GeoTiffs to supplement @scottyhq's fantastic example notebook. One of the wrinkles I'm tracking down and trying to document is when exactly the GDAL->rasterio->dask->xarray pipeline eagerly load the entire file versus when it defers reading or reads subsets of files. So far, it seems that if the GeoTiff is appropriately chunked ahead of time (when it's written to disk), things basically work "automagically."