Mean cold starts measured in seconds? Just how much are people importing at startup? And if the snapshot covers everything that happens before the first interaction, I'd expect it could make a more significant improvement on frequently-used processes. "I implemented a cache for something that happens constantly, and performance improved by only 2-3x" seems kinda underwhelming.