PoC, not optimized for performance.
Store data as JSON
Use one file per timeseries
Metadata is a dict of dimension/value pairs.
We can find each file quickly by using a hash of the metadata as the
filename
Can we find all timeseries which match only some of the dimensions
(e.g. response times of a particular service across all nodes)?
Opening each file is going to be slow pretty fast. So we need an index.
We don't expect new timeseries to spring into existence all that often,
so I guess for now we can just rewrite the whole index when a new
timeseries is added. Structure is pretty simple: Two levels of
dict (dimension, value) and then a list of matching timeseries.

Using an RDBMS doesn't seem like a good idea,

I'll design an efficient binary format later.