ACID correct distributed transactions create very, very painful performance problem because you have to lock across multiple physical machines separated by network connection(s).
Or y'know, you just don't do them correctly and accept a substantial rate of quietly corrupted data.
There isn't really a good option unless you have absurd amounts of money compared to your transaction volume.
They are possible but they are not possible to make per formant compared to the alternative.
Google falls under the "absurd amount of money" camp.
> Protocol Buffers have performance implications for query
processing. First, we always have to fetch entire Protocol Buffer columns from Spanner, even when we are only interested in a small subset of fields.
There are a number of performance costs documented in the paper. Simply because you can solve it via money doesn't mean it works for all situations.
Or y'know, you just don't do them correctly and accept a substantial rate of quietly corrupted data.
There isn't really a good option unless you have absurd amounts of money compared to your transaction volume.