Yeah that makes sense. But lets assume that your data access follows a power law. One piece of data will need to be accessed much more than everything else. You’ll still end up with contention on a single piece of data. In either scenario the benchmark will be the same. Have multiple concurrent readers try to access data in a process and in ETS and compare their tail latencies (the worst 5% is typical).
If you assume that data access is uniform then using multiple processes will help to spread out the load. But most data access doesn’t follow a uniform distribution. Even if it did, accessing an ETS table is going to be faster. I threw together this gist to demonstrate. You can call the time/0 function on both of those modules and compare results.
