[GiorgioG]

Unless you're operating at Google/FB/Twitter/etc scale, this is not a problem most of us face IRL. We need to stop pretending we have these problems at 99% of the places we work at. Most of us don't. We don't need 1,000 microservices, Kubernetes, Docker, and a terribly unreliable NoSQL database like MongoDb (there are better, more optimal solutions for those companies than MongoDb (Cassandra being one of them.))

[Thaxll]

The fact that someone has to manually promote a slave as master is a big operational cost. The people that actually use PG at scale have to write a lot of tools / plumbing to overcome those issues, the same "issues" that were solved by most NoSQL solution, horizontal scaling, HA, shards re-balancing ect ...

[pritambaral]

> The fact that someone has to manually promote a slave as master is a big operational cost.

It doesn't need to be manual, only in a layer above/outside PostgreSQL. There are solutions that automate leader elections for any secondary-promotable-to-primary system (not just PostgreSQL), just not built into PostgreSQL.

We looked for about two years, on and off, to setup smoother, less-manual HA for our PostgreSQL setup. We ended up writing some, as you say, "tools / plumbing to overcome those issues". By far, though, the most time we spent on it was been spent in reading, testing, and getting to understand how things work.

After everything, unfortunately, no other system fit our requirements. Many came close in their marketing copy, but on closer look, every system had caveats. Ultimately, PostgreSQL is the only system I could still trust to do the right thing. In fact, the solid foundation that Postgres provides is why I can trust third-party layers (similar to Stolon and Patron) to focus on the job of switchover correctly without losing or otherwise messing with my data. In fact, I can even switch which tool I use, without it affecting my data.

The little plumbing we've written over Pg's built-in replication allows us to horizontally scale using logical replication and assuredly have HA using physical replication. At the end of the day, any shop that does anything important with data at scale needs to know how their data is actually being treated, irrespective of what the marketing promises. Hell, even when the data-store does not blatantly lie to you about keeping your data safe or uncorrupted, you can lose data due to bad use of tooling or explicitly-set unsafe configuration. "I'll just use this; it'll solve all my problems!" is rarely a thing to blindly believe when dealing with important data.