Right now, the CBOR encoding uses strings for builtins. I assume decoding performance would be better if there were integers assigned to these builtins; they’d be a few bytes and less worse to decode.
This would mean another go at the standard but it would fix a concrete problem.
I’m not sure if this has been discussed; I looked a bit in the issue tracker for the Dhall standard and here.
My Experiment
I had a look at the occurrences of builtins in the serialized kubernetes example; there are quite a lot!
dhall decode --file benchmark/examples/kubernetes.dhall.bin | rg 'Text' -c
So there might be a performance improvement to be had.