How is it useful for the community to have these datasets on object storage? Why / how is it useful for you in practice?
And are the transformed cloud-optimised formats - COG, Zarr, Parquet, GeoParquet, etc. - better for your workflows in general? Are you using any particular standards or data formats more often with polar data?
For any datasets I just often (bedmap, REMA, etc) I will just download and use them locally, but datasets I less frequently, use just for making figures, or that are really large (ice velocity, imagery etc.) its nice to not have to have them locally and be able to access them remotely.
I think a single point of access for many datasets at once would improve research - partly because it’s easier for me to combine different datasets, and also because it could more easily highlight data gaps which could encourage research groups to create new datasets. I also think it’s also useful for a quick check to see if there is any data at all. My area of interest is the AP, which frequently is covered by poor or missing data. Because of the datacube quick access I’ve been able to check that datasets like the grounding lines don’t cover much of my region, so it’s not worth downloading it.
I second Matt’s thoughts that I would generally download datasets I need to touch frequently (my DEM subset dataset is about 40Tb when not compressed - inefficient and costly to access many times in the cloud). Being able to access infrequently-required datasets on demand without having to download the whole thing would be very useful. I do think cloud-optimised file formats still provide improvements to performance when stored locally, and frequently reduce required storage space in my use case, so I almost always convert my data to Zarr or similar.
I think the key thing which controls the usefulness of the data for workflows is the chunking relative to the most frequent access pattern. For example, I’d say that 90% of the time when looking at ice velocity data I’m performing operations along the time dimension. The chunking the Zarr stores I create is reflective of that, with a chunk covering the full time series but small spatial footprint. Some times this isn’t ideal, such as if you want to plot something spatially and then you could either provide precomputed mean fields (mean, median, standard deviation for each variable, instead of computing every time you open the dataset) or provide a two versions of each datacube, one chunked in time and one chunked in space, and open them as an xarray datatree.