6.7. Conclusions

Tornado allows this kind of setup to be considered where it wasn't before. These capabilities should be easy to qualitatively accept, but do the hard numbers support it? After some custom data-collection was programmed into a large provider, the answer turned out to be a strong probably.

It was then deployed on that provider, and this is what was found:

A StormCellar with 3TB of cheap disk was placed on the network and added as a server to the front ends. A cascading feed was set up to it.

Over the next week, retention went from 7 days to 13 days in every read group, transparently and without any blips in service or completion.

The StormCellar ended up carrying between 15% and 20% (depending on time of day) of the load, measured in bytes downloaded. In other words, about half the retention was being hit only 1/5th the time, meaning the load-carrying capacity of the StormCellar needed to be only 1/5th that of the main servers, which it easily was, at well under 1/10th the cost.