We did experiment with a few ideas outside of this decomposition method. In the end what you talked about here in your comment is exactly correct. And we value boring solutions in the end: https://about.gitlab.com/handbook/values/#boring-solutions.
I do dream of the day when distributed databases are the "default" for new projects. Likely running just a single instance at the beginning. But then you have a seamless path forward when (and if) you need it. Even at medium size three small nodes will be easy to manage and give easy HA and allow you to do version updates safely and with no downtime.
I don't think we are there yet, but there are a few contenders in the running but Postgres is tough competition. Its years of stability and predictability give it huge points even if it has the downsides of a centralized system. But I think the turning point is steadily approaching.
However there are some new databases that maintain a relational model (often fully compatible with PostrgeSQL or other relational DBs). The only real downside to these right now is stability and performance. Performance isn't a big deal when you can scale horizontally so I think when this "NewSQL" generation matures it will be nearly all upsides.
I'm not sure we will be stuck with SQL forever but strong consistency, SQL and the database-backed constraints and indexes are invaluable and I wouldn't want to be stuck without them.
But there are also https://www.cockroachlabs.com/product/, https://github.com/pingcap/tidb and https://ydb.tech/ which each has promise.
There is definitely lots of progress to make in this space. From what I have seen they are all significantly slower on a single node and are not as battle-hardened. But I think with time we will have a few really nice options to pick from.
Only provides a quite barebones Key Value interface to play around with though, so not anywhere near a drop in replacement to your traditional SQL database.
> CloudKit allows you to store your data as CKRecord objects, and relationships between those objects as CKRecord.Reference associations.
https://developer.apple.com/documentation/cloudkit/designing...
I’m not a database guy so idk if that meets the technical definition of “relational” but it is a database with relationships :)
You don't always need to experiment with bleeding edge tech or very complicated multi-leader clusters (though those also have their own use cases at a certain scale), sometimes just splitting the whole thing, especially when using something as solid as PostgreSQL, is enough!