The main advantage is that you can do time-consistent queries against any past database state across the entire database without any read locks. This frees resources and eliminates the necessity of timestamp tables.
Then there is Datomics caching model in combination with its local data based query engine. The vast majority of critical queries only read from memory. Data that is fetched doesn't block other consumers.
What RDBMS are you comparing to pretty much everything is MVCC so writers don't block readers and readers do not block writers mature RDBMS also are fairly highly optimized written in C with critical paths hand written in assembly so again would be interesting to learn for what workload Datomic is significantly faster. The biggest performance penalty in a modern RDBMS is GC of snapshots so if you need a design that is append only it will make modern RDBMS more performant.
I don't care so much about optimized tight loops written in assembly, rather about the ability to scale nowadays. Datomic uses databases like you describe e.g. Postgres as its storage. So it would be foolish to say that it beats their performance on a bare bones level. Instead, Datomics architecture and information model make it much easier and significantly reduce the overhead to design and implement applications that provide insane performance. I won't argue that you can hand rewrite every Datomic application in its underlying storage database and get more performance out of it if you do your caching and coordination right. It will take you much longer though (I'd guess a tenfold at least), likely have some very difficult to find bugs, and the result won't be as easy to extend. With Datomic, I get memory speed performance out of the box for the heavy hitters and so much more that it take would take some very uncommon requirements for me to choose something else nowadays.
Well I would buy the developer productivity argument but most applications have a mix of reporting requirements that are generally extremely hard to implement on anything "distributed". Also in a distributed system you are either running some consensus algorithm (paxos, RAFT etc) that will def. not be "insane performance" or you it will have issues with consistency.
Datomic uses distributed storage, writes are coordinated by a single instance (transactor). Reads don't block writers and immutability allows to query consistent snapshots. Does that address your concern?
