data schema optimizations

[iksaif]

Looking at the benchmarks what we've done, it's way more efficient to:

• Write multiple points at once in a single partition
• Having a lot of cells doesn't seem to be very efficient (longer repairs, compactions)

We should also take into account that's common to read together the metrics of the same directory. This could be used to batch those writes together. Combined to UDF we could still have something easy to use.