Second, while he does characterize it as a "huge mistake" for encoding/decoding performance reasons, there are a few caveats that are obvious here.

1) Half of his remarks here explain why varints are inappropriate for CapnProto, a separate project of his, inspired by his experience with protobuffs. CP has different design goals than PB, and in that context varints are inappropriate. Fine.

2) His argument here against varints in protobuffs (as contrasted with in capnproto) is only that they cost CPU time (and thus latency). That's surely true, but it's not at all clear to me that he's right that this is, in general, the wrong performance tradeoff.

3) In general, if you want to represent a long-tailed distribution of numbers, some variable-length encoding has merits, because the alternative (that protobuffs promote, IIUC) are fixed-width ints, which means you waste a lot of space to accommodate your long tail, and even so you might not have allocated enough space for the largest integer you ever need to handle.