There is no way to avoid cold starts althogether, there will always be tail latencies unless there is a generous amount of idle instances running that GCP charges for.
> under 1s. Hot request/responses are ~20ms.
Those numbers are great, I think Go plays a huge part in this. Google's Node.js firestore SDK is terrible... I've had 15 second cold starts, which is unacceptable for client-facing functions, there's a whole thread about it here [1]. GCP doesn't have a very wide range of language SDK support for those that don't want to use Go or Node.js...
> My hits are all US based
Edge compute, like Fly.io or Cloudflare Workers truly shines when you need to serve traffic close to the user around the world. Otherwise normal region-locked functions are fine. Vercel requires you to choose a single region, and it's locked to US-east for free-tier users. For us Europeans over here, SSR has to effectively cross the Atlantic ocean.