timholy
commented
3 months ago I've thought about this, but I think it's well-nigh impossible. In a certain sense, it's moved away from a serialization format and more towards direct usage of internal compiler structures. The linkages between these structures are enormously valuable and we'd have to duplicate everything to restore functionality. Given that I think https://github.com/julia-actions/julia-invalidations/issues/17 has a better solution (use SnoopCompileCore to collect the data, and then after all relevant invalidations are collected load SnoopCompile itself), I'll close.
We have greatly improved the dependency problem by letting SnoopCompileCore handle the recording phase. Also, I am aware that SnoopCompile is currently working on reducing hard dependencies.
However, I think that cross-environment analysis is not yet fully supported (due to the dependency of deserialization on the environment or world). An example of a specific problem is: https://github.com/julia-actions/julia-invalidations/issues/17
I cannot estimate how technically difficult this would be, nor do I personally think it is important. As the internal data structures of julia change, I believe that compatible serialization is also very costly to maintain. However, supporting this is not generally a bad approach in the long run, since it means that the analysis can always be done with the latest (or more stable than nightly) julia.