supports integer precision for values up to 2<sup>53</sup>. Supporting
native 64 bit integers would (only) help if you need integer precision
above 2<sup>53</sup> but below 2<sup>63</sup>. In principle, support
for different sample value types (including some kind of big integer,
supporting even more than 64 bit) could be implemented, but it is not
a priority right now. Note that a counter, even if incremented
one million times per second, will only run into precision issues
after over 285 years.
### Why does Prometheus use a custom storage backend rather than [some other storage method]? Isn't the "one file per time series" approach killing performance?
Initially, Prometheus ran completely on LevelDB, but to achieve better
performance, we had to change the storage for bulk sample data. We
evaluated many storage backends that were available at the time,
without getting satisfactory results. So we implemented exactly the
parts we needed, while keeping LevelDB for indexes and making heavy
use of file system capabilities. Obviously, we could not evaluate
every single storage backend out there, and storage backends have
evolved meanwhile. However, the performance of the solution
implemented now is satisfactory for most use-cases. Our most important
requirements are an acceptable query speed for common queries and a
sustainable ingestion rate of many thousands of samples per
second. The latter depends on the compressibility of the sample data
and on the number of time series the samples belong to, but to give
you an idea, here are some results from benchmarks:
* On an older 8-core machine with Intel Core i7 CPUs and two spinning
disks (Samsung HD753LJ) in a RAID-1 setup, Prometheus sustained an
ingestion rate of 20k samples per second, belonging to 450k time
series, scraped from 1500 targets.
* On a modern server with SSD, Prometheus sustained an ingestion rate
of more than 100k samples per second, belonging to millions of time
series, scraped from thousands of targets.
In both cases, the bottleneck was identified as insufficiently
parallelized hash calculation, which happens before samples even hit
the storage backend.
Running out of inodes is highly unlikely in a usual set-up. There is a
possible downside: If you want to delete Prometheus's storage
directory, you will notice that some file systems are very slow when
