> I think it's in-memory.

In-memory support was added later, though I'm not sure if it made it into a release as I stopped paying attention around when they started developing that feature.

They originally ran exclusively on specialized appliances having battery-backed write caches. The implementation is all in C using Linux AIO w/epoll. It's a sophisticated, high-performance storage engine.

When everyone shifted to the cloud and stopped owning hardware, Clustrix was forced to add a file-based backend. This still uses the high-performance AIO+epoll engine.

I spent a lot of time at a previous job supporting an NDB cluster and I can attest first hand that it is awful.

Single node failures cause entire cluster shutdowns, the cluster then takes forever to recover and must be done in a specific order. In fact just thinking about it makes me anxious.

What version was this ?