[ROADMAP] Dask roadmap

[remi-braun]

Using Dask arrays doesn't currently break EOReader. However it's support is very little.
Major improvements have to be done in this direction:

• About reprojection: Abstract reprojection sertit-utils#41

• About any other sertit-utils methods: Daskify 'rasters' module: roadmap sertit-utils#27

  • Waiting for Add Xarray accessor mirroring Raster class GlacioHack/geoutils#446
  • and Add dem Xarray accessor mirroring DEM class GlacioHack/xdem#656

• Ensure EOReader works lazily

  • Beware of ifs conditions, large graphs, etc!

• ⚠️ Major limitation: SAR data is handled through SNAP, so remove this softyware to begin to correctly use dask for SAR ENH: [SAR] Don't use SNAP #2

• Correlated to OPTIM: Speeding up EOReader #58

[remi-braun]

See https://github.com/pytroll/satpy/pull/3082/files and https://dask.discourse.group/t/optimal-graph-optimization-when-mixing-dask-objects/3867/4 for delayed optimisations

[remi-braun]

https://distributed.dask.org/en/latest/_modules/distributed/diagnostics/memory_sampler.html may be of use

[guillaumeeb]

A quick note following a private discussion: two problems detected in temporal series processing of Sentinel 2 images:

• EOReader "loading" product is not lazy: 2m30 for opening 4 bands using Dask. Some process is probably occuring. It takes only about 3 seconds using rioxarray only.
• There is a useless band dimension in the Variable.

Some examples:

datalake-rioxarray-dask (5).ipynb

datalake-eoreader-dask.ipynb