Thank you pludov ! I checked and it looks like this is also what Gulinux did with AstrophotoPlus, he has a blob client specific class.
I decided to keep a "Big" client, responsible for every callbacks related to one server, but I changed the newBlob callback to use something less naive that was I was using before, with a proper process-based mechanism to enqueue the very next blob that comes in after an acquisition has been triggered. (Again, I took was Gulinux did in AstrophotoPlus, which is an amazing software).

I might change the big client approach in the future, depending on the performances.
Anyway, thanks to this queuing mechanism, it looks like I don't have this problem anymore, now, I don't really now if the problem was due to timing issues, or memory space / multi process synchronization (I don't remember if client callback are run in the same process or not), but now the problem is gone.

Read More...