The thing that kills you is IOPs per query.

At 30 TBs the hot data set is huge an there is a long tail of queries not loaded into memory.

Plus they were saying 10k+ connections per server. So we are probably looking st 500k+ connections.

There is also a regular influx on unoptimized / badly written queries as people add features/games etc. Plus you can end up with random spikes.

This setup is likely optimized for high read availability. The slides mention the mental overhead of failovers. This is a thing. Try to explain and help each team understand how to do those safely. Downtime are also not an option.

So there is quite some information and context (likely) missing. I used to manage a similar scale zoo of MySQL (similar peak QPS, less total data iirc). I am not sure I would subscribe so much on the Vitess as solution side though. Galera and Group Replication can give you hassle free failovers, add some read replicas and you have a really nice DB setup that can scale horizontally amd vertically until you hit the write limitations per server.