Jan Hellar sorry I didn't see these pings until now, for some reason my old (apparently not deleted) account was the one you pinged.

Looking at what you said above I agree with Wojciech Trocki that either case introduces a lot of edge cases that we will have to test. I think a happy medium could be adding some form of retry link that is specifically interested in network errors and will retry mutations accordingly.

This guy has an interesting approach to this: https://github.com/carvajalconsultants/apollo-link-offline where he enqueues and retries network error'd mutations. The differences with his approach to yours are that he syncs (parallel retries everything).