[metricspaces]

This is an interesting point but imo ignores the facilities afforded by an RDBMS and possibly even misunderstands the notion of fully hosting applications on a RDBMS.

Some sort of check pointing is inevitable for recoverable state-less systems (that naturally react to a stream of data/queries) and that ultimately is delegated to either a streaming/messaging (e.g. kafka) or a database (of some sort).

Components of an application hosted in a RDBMS are tables, indexes, views, and code. The code is stateless — they’re just functions. Application state is in tables. Replication is supposed to handle your more hairy operational concerns. What’s the problem?

[zokier]

> Replication is supposed to handle your more hairy operational concerns. What’s the problem?

Replication is slow, and requires each node to have lot of resources, and generally prefers nodes to be realtively homogenous. But for non-trivial applications replicating the whole application (or some shard of it) to all nodes is simply impractical, and it is useful to be able to scale and adjust different parts of system independently.

For example on AWS Lambda you can scale out up to 1000 new instances per 10 seconds (and scale in at similar rate). Can you imagine any DMBS replication working very effectively at such circumstances?