It would be awesome to have easy clustering directly built-in. Similar to MongoDB, where you tell the primary instance to use a replica set, then simply connect two secondaries to primary, done.
It would be awesome to have easy clustering directly built-in. Similar to MongoDB, where you tell the primary instance to use a replica set, then simply connect two secondaries to primary, done.
This is very hard to fix and requires significant architectural changes (like Yugabyte or Neon have done).
Argument is that partitioning almost never happens, because of network reliably rerouting traffic.
Clearly, it's possible to reduce the risk to the point where many companies are willing to accept it, but it's still a problem and comes with high operational costs. And for some use cases (like finance), even a small risk of undefined database behavior or lost writes is unacceptable.
The future are distributed databases with consensus, and unfortunately, Postgres isn't there yet.
If your network is down, your clients can't connect DB too, and nothing works. State of network partitioning is harder to achieve, its more like if connection between two datacenters is down where DB replicas live, which is more rare because of all global connection redundancy.
> If your network is down, your clients can't connect DB too, and nothing works
This sounds simple, but can only be achieved with high certainty using distributed consensus.