Since iris 3.6, iris supports delayed saves. This is a great new feature that allows:

• checking that a preprocessor run will succeed before doing any data crunching, saving (compute) time if there is something wrong with the data
• faster preprocessor runs because
  • intermediate results can be reused to write multiple output files (e.g. regrid only once for multimodel statistics)
  • I suspect this will allow to better overlap computations with writing/reading data, but I did not measure it so far.

See #1714 for an example implementation.