Warning: This document is for the development version of Gnocchi.

Project Architecture

Gnocchi consists of several services: a HTTP REST API (see REST API Usage), an optional statsd-compatible daemon (see Statsd Daemon Usage), and an asynchronous processing daemon. Data is received via the HTTP REST API and statsd daemon. The asynchronous processing daemon, called gnocchi-metricd, performs operations (statistics computing, metric cleanup, etc...) on the received data in the background.

Both the HTTP REST API and the asynchronous processing daemon are stateless and are scalable. Additional workers can be added depending on load.


Gnocchi uses two different back-end for storing data: one for storing the time series (the storage driver) and one for indexing the data (the index driver).

The storage is responsible for storing measures of created metrics. It receives timestamps and values, and pre-computes aggregations according to the defined archive policies.

The indexer is responsible for storing the index of all resources, along with their types and properties. Gnocchi only knows about resource types from the OpenStack project, but also provides a generic type so you can create basic resources and handle the resource properties yourself. The indexer is also responsible for linking resources with metrics.

How to choose back-ends

Gnocchi currently offers different storage drivers:

  • File
  • Swift
  • Ceph (preferred)

The drivers are based on an intermediate library, named Carbonara, which handles the time series manipulation, since none of these storage technologies handle time series natively.

The three Carbonara based drivers are working well and are as scalable as their back-end technology permits. Ceph and Swift are inherently more scalable than the file driver.

Depending on the size of your architecture, using the file driver and storing your data on a disk might be enough. If you need to scale the number of server with the file driver, you can export and share the data via NFS among all Gnocchi processes. In any case, it is obvious that Ceph and Swift drivers are largely more scalable. Ceph also offers better consistency, and hence is the recommended driver.

How to plan for Gnocchi’s storage

Gnocchi uses a custom file format based on its library Carbonara. In Gnocchi, a time series is a collection of points, where a point is a given measure, or sample, in the lifespan of a time series. The storage format is compressed using various techniques, therefore the computing of a time series’ size can be estimated based on its worst case scenario with the following formula:

number of points × 8 bytes = size in bytes

The number of points you want to keep is usually determined by the following formula:

number of points = timespan ÷ granularity

For example, if you want to keep a year of data with a one minute resolution:

number of points = (365 days × 24 hours × 60 minutes) ÷ 1 minute
number of points = 525 600


size in bytes = 525 600 × 8 = 4 204 800 bytes = 4 106 KiB

This is just for a single aggregated time series. If your archive policy uses the 8 default aggregation methods (mean, min, max, sum, std, median, count, 95pct) with the same “one year, one minute aggregations” resolution, the space used will go up to a maximum of 8 × 4.1 MiB = 32.8 MiB.

How to define archive policies

In Gnocchi, the archive policy definitions are expressed in number of points. If your archive policy defines a policy of 10 points with a granularity of 1 second, the time series archive will keep up to 10 seconds, each representing an aggregation over 1 second. This means the time series will at maximum retain 10 seconds of data (sometimes a bit more) between the more recent point and the oldest point. That does not mean it will be 10 consecutive seconds: there might be a gap if data is fed irregularly.

There is no expiry of data relative to the current timestamp. Also, you cannot delete old data points (at least for now).

Therefore, both the archive policy and the granularity entirely depends on your use case. Depending on the usage of your data, you can define several archiving policies. A typical low grained use case could be:

3600 points with a granularity of 1 second = 1 hour
1440 points with a granularity of 1 minute = 24 hours
720 points with a granularity of 1 hour = 30 days
365 points with a granularity of 1 day = 1 year

This would represent 6125 points × 9 = 54 KiB per aggregation method. If you use the 8 standard aggregation method, your metric will take up to 8 × 54 KiB = 432 KiB of disk space.

Be aware that the more definitions you set in an archive policy, the more CPU it will consume. Therefore, creating an archive policy with 2 definitons (e.g. 1 second granularity for 1 day and 1 minute granularity for 1 month) will consume twice CPU than just one definition (e.g. just 1 second granularity for 1 day).

Default archive policies

By default, 3 archive policies are created using the default archive policy list (listed in default_aggregation_methods, i.e. mean, min, max, sum, std, median, count, 95pct):

  • low (maximum estimated size per metric: 406 MiB)
    • 5 minutes granularity over 30 days
  • medium (maximum estimated size per metric: 887 KiB)
    • 1 minute granularity over 7 days
    • 1 hour granularity over 365 days
  • high (maximum estimated size per metric: 1 057 KiB)
    • 1 second granularity over 1 hour
    • 1 minute granularity over 1 week
    • 1 hour granularity over 1 year