That makes sense, I will try out setting the value size knob in all processes of the server and backup_agent as well.
I think I might have seen at least one weird behaviour as a result of setting them only on the client - we do some parsing of backup files for some internal analysis, and noticed that log ranges had incomplete mutation blocks for certain versions. The size encoding at the beginning of each block mentioned (10000x + y) bytes for a particular version, but we found only 10000z bytes in that block, where x > z. Here, I am accounting for the fact that the mutation logs for each version can sometimes be spread across multiple parts and we have to concatenate those parts to get the final set of logs per version. This happened for a significant number of mutation blocks.