Well it's a distributed data store so I guess at times they took down certain nodes for upgrades or fixes etc then bought them back up afterwards where they became consistent again. This probably wouldn't have had any user impact.

These distributed data stores are so appealing from the perspective of uptime and stability.

Needing to upgrade the DB or take cold backups or run massive schema migrations is the reason for lots of scheduled downtime.

And removing the huge whopping single point of failure from your system that you would have with a traditional RDBMS can only do good things for your stability.