Thanks @axel-lauer for the review! The memory issue should be solved now and also the years could be also set now in the command line. Below you find my answers to your comments:

* Units are reported as `degC` but should be in `K` (fields also look like data are in `K`)

Cmor units are degC, the fields is now also converted.

* The variable attributes `host`, `reference`, `source`, `tier`, `user` and `version` should rather be global attributes

This happened due to a change in Iris. I updated the function set_global_atts in PR #3717

* Not sure about the variable attributes `mip`, `modeling_realm` and `project_id`. Other CMORrized datasets do not have these attributes. I would suggest to either change these attributes to global attributes or to remove them.

They are also global attributes now.

* I would remove the variable attribute `comment: Note that the variable tsStderr is an uncertainty not a standard error.` from variable `tos` as this only applies to `tosStderr`, which is written to separate files

The comment is now only attached to 'tosStderr'.

* the attribute `long_name` of variable `time` should be simply `time` (instead of `reference time of sst field`)

Thanks. Done.

* similarly, the attribute `long_name` of variables `lon` and `lat` should be `longitude` and `latitude` instead of `longitude coordinate` and `latitude coordinate`

Thanks. Done.

Thanks @LisaBock for addressing all my comments. Everything looks fine to me now. The only question mark left is whether the uncertainty data tsStderr should really be regridded. Strictly speaking, we would need to know the correlation lengths in order to be able to regrid the data, which we do not have at the moment. One option could be to simply not regrid tsStderr so we can look into this at a later stage.

Thanks again @axel-lauer !
I removed the regridding of the uncertainty field. As the files are now much huger we might not to cmorize the whole time period for tosStderr...