Lots of people are citing cutting-edge bells and whistles for advanced query plans and data types, but after working at $BIGCO and seeing what exists there (and has existed for years), I get frustrated by the lack of ergonomic and operational essentials in the public-cloud options.

1. Horizontal scaling. PG can do sharding or replication, but that's not the same thing. Manual resharding is something that, once you've experienced it, you don't want to do it ever again, especially when things like Spanner/Cockroach exist. Database-level liquid sharding is such a dramatically superior solution that it makes any system that depends on an enumerated set of master instances seem completely obsolete.

2. Strong commit timestamps, again a la Spanner/Cockroach. Globally-ordered commit times and read snapshots aren't something you'll need in every schema, but when you do, they're magical - suddenly, you can use your OLTP system for strong lease assignment (conditioning other transactions on lease validity), you can construct arbitrary application-level transaction semantics (non-transactional read followed by transactional compare-and-swap is pretty powerful in some scenarios), and all sorts of other things.

3. Interface abstraction. Databases that ship language-support drivers should also ship a fake in-memory implementation for testing purposes that supports the same interfaces and options. There's no reason why I should have to start a subprocess and emulate a network connection to tell whether a query functions properly or not, or whether I've set isolation properties correctly for a pipeline.

4. Declarative schemas, and a well-defined update process (that isn't human-written DDL), are essential at any sort of organizational scale. Every time I see an "ALTER TABLE" statement checked into a repository, I tear my hair out.

5. Queue support, where message publishing/consumption is transactional with other read/write operations. You can always build a queue with a time-keyed table, but it's kind of a pain in the ass to get all the edge cases right. It's much better as a first-class concept supporting automatic sharding, retries, batching, and bounded-out-of-order delivery with a global cursor.