The most important piece of advice is not proto specific: use pprof. It may sound obvious, but even though I have been using it for years, I find myself still not running it enough.

Proto decoding specific: if you're parsing a message with lots of fields, you can define a new message with just the fields you care about. I.e. if you're parsing

    message M {
      optional int32 F1 = 1;
      // ...
      optional int32 F100 = 100;
    }

but you only care about F1 and F2 in this message, define

    message SmallM {
      optional int32 F1 = 1;
      optional int32 F2 = 2;
    }

If you use SmallM to decode your messages, it will skip a lot of work filling out fields F3-F100, saving you a lot of CPU.