The easiest solution for us (a startup providing a product based on satellite data) is to rely on STAC + COG. It’s easy to handle and, from my perspective, covers almost all the use cases.
Retrieving data (by selecting the relevant pixels) can be done using odc.stac and rioxarray.
Visualization can be achieved by relying on titiler, as well as by directly adding the URL into QGIS (which works great for quick visualization), and large-scale processing can easily be handled using coiled (or HPC).
However, we haven’t succeeded in making it work for one use case:
→ Retrieving the time series of all Sentinel-2 data over more than 60,000 points.
We used xvec (a pretty awesome library), but it was still too slow…
I’m not sure if zarr would be a better candidate for this use case.
Regarding ML, I don’t know if batching COG (with xbatch) would have the same capabilities as zarr. I think not, but I’m not sure how many users will try it.