A columnar format is almost always what you want for analytical workloads, because their access patterns tend to iterate ranges of rows but select only a few co

A columnar format is almost always what you want for analytical workloads, because their access patterns tend to iterate ranges of rows but select only a few columns at random. About the only thing protocol buffers has in common is that it's a standardized binary format. The use case is largely non-overlapping, though. Protobuf is meant for transmitting monolithic datagrams, where the entire thing will be transmitted and then decoded as a monolithic blob. It's also, out of the box, not the best for efficiently transmitting highly repetitive data. Column-oriented formats cut down on some repetition of metadata, and also tend to be more compressible because similar data tends to get clumped together. Coincidentally, Arrow's format for transmitting data over a network, Arrow Flight, uses protocol buffers as its messaging format. Though the payload is still blocks of column-oriented data, for efficiency.