Similar to #18659 I decided to look at the overhead in Batch{Request,Response} for simple "kv" reads (i.e. a single Scan{Request,Response} in the batch).

  BatchRequest:               191
  BatchRequest.Header:        156
  BatchRequest.Header.Txn:    122
  BatchRequest.Requests[0]:   30
  BatchResponse:              189
  BatchResponse.Header:       138
  BatchResponse.Header.Txn:   122
  BatchResponse.Responses[0]: 46

The Scan{Request,Response} consume exactly 20% of the total bytes sent/received. The request and response txn consume 64% of the bytes. The response txn is nearly identical to the request txn with only Transaction.MaxTimestamp having changed.

There are some small gains to be had by converting Batch{Request,Response}, Transaction and TxnMeta to proto3 to avoid encoding zero value primitive fields. There is a large gain to be had by only returning the delta of the transaction in the response.

Cc @andreimatei who suggested last week that we develop a more principled way to update the transaction upon receiving a response.