On Thu, Jan 16, 2025 at 2:02 PM Laurenz Albe <laurenz.albe@xxxxxxxxxxx> wrote:
On Thu, 2025-01-16 at 15:25 +0300, anlex N wrote:
> I have searched all archives, but have no results. Please help me.
That's probably because the question doesn't make much sense.
A single PostgreSQL instance (=cluster) enforces consistency,
and it doesn't matter which data type you are using.
So I guess the answer should be "strong consistency".
I guess what the original author was trying to ask is how many databases have to acknowledge a transaction before it is considered committed, stuff that's called in mongodb write_concern and read_concern.
@anlex, PG at its heart is not a distributed system. When you talk about consistency in PG you normally mean foreign key consistency or similar, properties where data in one table is somehow related to data in another table. A slightly different meaning of consistency comes when you talk about transaction isolation levels.
All of this is different from Mongodb's read/write consistency. In Mongodb a write can go to one machine while the subsequent read can fetch data from another. So, you worry about linearizability, causal consistency and such things. The only slightly similar question in PG is "Can I see data somehow, on the master or a replica, that has been written to the database but where the master has not yet acknowledged the commit to the client?"
The short answer is, yes. As soon as the commit record has been written to the WAL, it is being transferred and replayed on streaming replicas. After that the master might still have to do things like waiting for a synchronous replica to acknowledge the transaction before the transaction becomes visible on the master. On a fast enough replica, the transaction can, hence, become visible before it's visible on the master.
You can control that behavior with the `synchronous_commit` setting in combination with `synchronous_standby_names`. You can tell the DB, for instance, a transaction should be considered as committed only when 3 out of 5 replicas have acknowledged it.
If you set synchronous_commit=remote_apply and configure N out of N replicas in synchronous_standby_names, that would likely give you a distributed database with strong consistency where you can direct writes to the master and read-only load to the replicas. But I guess nobody in his right mind would do that unless your database is rarely written to.
