A service implementing the Carbon protocol and storing time series data using kairos
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.



Version: 0.7.1
Keywords:python, redis, time, rrd, gevent, carbon, graphite, whisper, statsd, kairos

A suite of tools designed to replace Graphite and expand on its capabilities. Uses kairos to support storing and reading data from many different types of data stores, and focuses on providing programmatic tools for storing, retrieving and processing of streaming timeseries data.


Kairos, an RRD-inspired timeseries library, provides an improved storage engine and many more features than most other systems backing statsd. Compared to traditional disk stores such as RRD and Whisper, Torus adds:

  • simple runtime for ease in development and deployment
  • abstraction on top of kairos for histograms
  • compact storage for sparse data points
  • scaling with per-schema hosting and sharding
  • non-buffering semantics for aggregate processing
  • consistent hashing of timestamps for ease in interleaving and interpolation
  • programmatic interface to data processing

StatsD Quick Start

A configuration file that tracks hourly, daily and monthly data in SQLite is available in examples/statsd.py. The default will create a temp directory for the current user to store the databases (e.g. /tmp/torus.user on Unix). Change STORAGE_DIR at the top of the file to set a permanent location.

If you have installed torus in a virtual env, you can use foreman to start both karbon and torus. If you're running torus out of the repository, then you can use foreman start -f Procfile.dev.

The example configuration includes support for performance testing (see below).

Carbon Server

The karbon application runs the Carbon-compatible stat collection application. It is a drop-in replacement for the Carbon backend of statsd. It takes the following arguments:

usage: karbon [-h] [--tcp TCP] [--config CONFIG]

Karbon, a Carbon-replacement data collection server

optional arguments:
  -h, --help       show this help message and exit
  --tcp TCP        TCP binding, in the form of "host:port", ":port", or
                   "port". Defaults to "localhost:2003".
  --config CONFIG  Configuration file to load. Can be called multiple times
                   for multiple configuration files.

The configuration is documented below. To reload the configuration(s), send a SIGHUP to the karbon process.

Query Server

The torus application is a replacement for Graphite. It is not API compatible with Graphite though it does aim to be familiar to Graphite users and provides a graphite-compatible JSON format for ease in integrating with existing toolchains.

usage: torus [-h] [--tcp TCP] [--config CONFIG]

Torus, a web server for mining data out of kairos

optional arguments:
  -h, --help       show this help message and exit
  --tcp TCP        TCP binding, in the form of "host:port", ":port", or
                   "port". Defaults to "localhost:8080".
  --config CONFIG  Configuration file to load. Can be called multiple times
                   for multiple configuration files.

For most use cases it can share a configuration with karbon. However, one could use Chef, puppet or a similar tool to templatize the configuration, and replace strings such as the host definition, so as to target a specific set of resources at reading the data.

To reload the configuration(s), send a SIGHUP to the torus process.

torus will respond to http://$tcp/$command?$parameters for the following commands, where $parameters is a standard URL encoded parameter list.



DEPRECATED: formerly /data


Fetches data for one or more statistics and returns a list of objects for each statistic. Returns data from the first schema that matches a statistic.

  • stat

    The name of the statistic to fetch. Each instance of the stat parameter is interpreted as a separate statistic. The statistic can either be in the form of $stat_name or $func($stat_name), where $func can be one of:

    • avg - the average of each datapoints in each time slice.
    • min - the minimum value of datapoints in each time slice.
    • max - the maximum value of datapoints in each time slice.
    • sum - the sum of datapoints in each time slice.
    • count - the number of datapoints in each time slice.

    Additionally, $func can be either a transform or a macro defined in the configuration. The $func can be anything that matches the pattern [a-zA-Z0-9_].

  • format

    One of [graphite, json], where graphite is a Graphite-compatible json format and json offers more nuanced representation of kairos' data structures.

  • condense

    One of [true, false], if kairos resolutions are configured for a schema, determines whether resolutions are flattened or returned as-is. Forced to true for graphite format.

  • collapse

    One of [true, false], if true then all of the data for each time interval will be collapsed into a single value. This is useful for calculating aggregates across a range (e.g. "all hits in last 5 days").

  • schema

    In cases where multiple schemas match a stat name, force a particular schema to be used.

  • interval

    The interval to choose, one of the intervals available in whatever schema matches stat. Must apply to all stat arguments.

  • start

    An optional timestamp for the beginning of the return interval. Can be in the form of a unix timestamp, a strftime-formatted string, or a human-readable relative value such as "today", "5 days ago", "last week", etc.

  • end

    An optional timestamp for the end of the return interval. Can accept the same values as start. With no arguments, this is implicitely the time at which the query is made.

  • steps

    Given either a start or end timestamp, this parameter defines the number of intervals (inclusive) after or before (respectively) to return. So if start is "last week" and steps=7, the result data will end with yesterday's data. If no timestamps are given, this is the number of intervals before the current time (inclusive).


A json structure.

  'function': 'avg',
  'interval': 'hour',
  'schema': 'calls',
  'stat': 'avg(calls.system)',
  'stat_name' : 'calls.system',
  'target': 'calls.system',
  'datapoints': [[0.0391, 1362153600], [0, 1362157200]],


The stat field will be the full name of the corresponding parameter, including the function (if any). The stat_name field will be just the name of the statistic that was matched to the schema, and target will be a copy of the same for clients which are expecting data in graphite format.


All torus applications load one or more configurations, where a configuration is a python module that is loaded into the application. Torus looks for the constants documented below, but as the configuration is a full python module, extensions, plugins and additional runtime configuration can be included. For example, one can connect torus' use of the standard python logger to syslog, logstash or one of many error reporting tools, such as Sentry. For torus, log messages prioritize the header X-Forwarded-For and then use the remote IP address if that's not available. For this reason and general security, you should always use a proxy server in front of torus.

The configuration for torus includes a definition for schemas, aggregates, custom functions that can be used in queries, and debugging settings. The schema for torus is an extension of the kairos schema; each of the key-value pairs in a schema definition will be passed to the timeseries constructor. The configuration files can include 1 or more of the following.


If this is a callable, will be called the first time the configuration is loaded. Useful for one-time configuration such as Sentry logging handlers.


If this is a callable, will be called when the configuration module is reloaded.


A dictionary of unique names to the configuration for capturing and storing the statistics which match the regular expressions. A schema definition supports the following fields, many of which are passed directly to kairos.

  • type

    Required, defines the type of the timeseries. One of [series, histogram, count, set, gauge], depending on what the backend supports.

  • host

    Required, the URL connection string or an instance of a supported connection type. See Storage Engines

  • client_config

    Optional, is a dictionary of parameters to use in the connection constructor associated with the host URL. See Storage Engines

  • match

    A string, or a list of strings, which are regular expressions that define the stat names which should be stored and queried in this schema. In the case where a transform is defined, it is likely the one or more expressions will define the input stats, and another expression will define the stat which can be queried. See GitHub issue.

  • rolling

    Optional, defines how many intervals before (negative) or after (positive) that a copy of data should be written to whenever data is inserted. The extra storage size offsets much faster calculation of aggregates over pre-determined date range. For example, when storing daily values, a value of -30 will store a value as if it occurred any time in the last 30 days.

  • prefix

    Optional, is used to scope data in redis data stores. If supplied and it doesn't end with ":", it will be automatically appended.

  • transform

    Optional, allows one to replace the stat name and value with another. Takes two arguments and must return a tuple of two items (statistic, value). If the statistic is None, will skip writing the statistic. The value will be a string on input, and on output must be acceptable to any write_func defined. Example: transform: lambda s,v: (None,None) if 0>long_or_float(v)>3.14 else (s,v)

  • read_func

    Optional, is a function applied to all values read back from the database. Without it, values will be strings. Must accept a string value and can return anything. Defaults to long_or_float, which tries to cast to a long and failing that, cast to a float. long_or_float is available for all schemas to use.

  • write_func

    Optional, is a function applied to all values when writing. Can be used for histogram resolution, converting an object into an id, etc. Must accept whatever can be inserted into a timeseries and return an object which can be cast to a string. Defaults to long_or_float, which tries to cast to a long and failing that, cast to a float. Example: write_func: lambda v: '%0.3f'%(v)

  • intervals

    Required, defines the intervals in which data should be stored.

  • generator

    Optional, defines a function which can be used to generate load tests. Must return a tuple in the form (stat_name, value). Example: lambda: ('application.hits.%d'%(random.choice([200,404,500])), 1)



  'response_times' : {
    'type': 'histogram'
    'host': 'redis://localhost:6379/0'
    'match': [ 'application.*.response_time', 'application.response_time' ]
    'read_func': float
    'write_func': lambda v: '%0.3f'%(v)

    'intervals': {
      'minute': {
        'step': 60,
        'steps': 240,
      'daily' : {
        'step': 'daily',
        'steps': 30


Similar to Carbon aggregator but without the time buffer. Matching stats will be processed through any matching schemas. Is a list of tuples to support rolling up any number of dissimilar stats into a single one. At this time key names must be in the character set [a-zA-Z0-9_-]. Each aggregate is defined as a tuple in the form of (rollup_stat, source_stat). Captures can be defined in the form of <capture> and used in each rollup.


  ('application.response_time', 'application.*.response_time'),
  ('application.<status_code>', 'application.*.status.<status_code>'),


A named mapping of functions which can be used in queries. See kairos documentation for details.


  # Returns the number of elements
  'size' : lambda row, duration: len(row)


A named map of configuration options so that "foo(stat)" will result in a fixed set of options passed to kairos. This is especially useful for using the customized read feature of kairos. This example assumes a histogram stored in redis. A more complicated macro might use server-side scripting. All custom read functions exposed in kairos can be defined here. All fields of the query string, other than 'stat', can be set in the macro definition and will override those query parameters if they're provided. To use a transform in a macro, set the 'transform' field to either a string or a callable. Macros can make use of transforms defined in TRANSFORMS.


  'unique' : {
    'fetch' : lambda handle,key: handle.hlen(key)
    'condense' : lambda data: sum(data.values()),
    'process_row' : lambda data: data,
    'join_rows' : lambda rows: sum(rows),


A boolean or integer to define the amount of log output.

  • 0 or False

    Only errors are logged.

  • 1 or True

    Basic information is logged, should not generate substantial output.

  • 2

    Significant information is logged, particularly from the karbon process.


Debugging a schema or set of schemas can pose a challenge. Torus ships with schema_debug, a tool for testing any number of input strings against any number of schemas. It will output which rules match the input string, which database that match will be stored in, any aggregates that will be generated from the input rule, and then recursively any schemas and aggregates that match each aggregate.

usage: schema_debug [-h] [--config CONFIG] strings [strings ...]

Debugging tool for schemas

positional arguments:
  strings          One or more input strings to test against the scheams

optional arguments:
  -h, --help       show this help message and exit
  --config CONFIG  Configuration file to load. Can be called multiple times
                   for multiple configuration files.

Torus also supports the DEBUG flag which can be defined in any of the configuration files and which will cause karbon to print to stdout. If it is 0, or not defined, no output will be generated. If it is 1, karbon will log when it stores a raw value (STOR) or aggregate (AGRT), and statistics on the quantity and duration of processing (DONE). If DEBUG==2, karbon will also log every line it recieves (RECV) and lines that it skips (SKIP).

To use the debugging flag, you can change the value in one of the configuration files loaded by karbon, and then signal the process to reload with the command kill -SIGHUP `pidof karbon`.

Performance Testing

To test your schema for performance and regressions, torus includes schema_test. The tool looks for generator definitions in schemas, and continually calls them to emit data points that are processed through all the schemas and aggregates. Prints out some basic statistics.

usage: schema_test [-h] [--config CONFIG] [--clear] [--duration DURATION]

Tool for performance testing of schemas

optional arguments:
  -h, --help           show this help message and exit
  --config CONFIG      Configuration file to load. Can be called multiple
                       times for multiple configuration files.
  --clear              If true, clear all data before running the test.
                       Defaults to false.
  --duration DURATION  Duration of the test. Defaults to 60 seconds.


There will be times that you need to migrate data from one schema to another. Torus ships with migrate to facilitate that.

usage: migrate [-h] --config CONFIG --source SOURCE --destination DESTINATION
               --interval INTERVAL [--start START] [--end END]
               [--concurrency CONCURRENCY] [--stat STAT] [--match MATCH]
               [--dry-run] [--verbose]

A tool to migrate data from one schema to another

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       Configuration file to load. Can be called multiple
                        times for multiple configuration files.
  --source SOURCE       The name of the source schema [required]
  --destination DESTINATION
                        The name of the destination schema [required]
  --interval INTERVAL   The name of the interval from which to read data
  --start START         Only copy stats occurring on or after this date. Same
                        format as web parameter. [optional]
  --end END             Only copy stats occurring on or before this date. Same
                        format as web parameter. [optional]
  --concurrency CONCURRENCY
                        Set the concurrency on the schema target writing.
                        Defaults to 10.
  --stat STAT           The name of the stat to copy. Can be called multiple
                        times for a list of stats. If not provided, all stats
                        will be copied. [optional]
  --match MATCH         Pattern match to migrate a subset of the data.
  --dry-run             Print out status but do not save results in the
                        destination schema. [optional]
  --verbose             Print out even more information during the migration


Torus is available on pypi and can be installed using pip

pip install torus

If installing from source:

  • with development requirements (e.g. testing frameworks)

    pip install -r development.pip
  • without development requirements

    pip install -r requirements.pip


Torus installs SQLAlchemy to support SQL. To use your dialect of choice, you will likely have to install additional packages. Refer to the documentation for more details.


Use nose to run the test suite.

$ nosetests


  • Record metrics on karbon and torus usage
  • Improve parsing of stat parameter with pyparsing library
  • Add "dead letter" support for tracking stats that don't match any schema
  • Add stat delete endpoint to torus
  • Command line tools for querying data and optionally plotting using bashplotlib
  • Add tools for generating tasseo configurations (https://github.com/obfuscurity/tasseo)
  • Add ability to set transaction-commit intervals for Redis and SQLite backends
  • Investigate faster regular expression engines. pyre2 is currently in the running.
  • Expand supported stat naming (unicode, symbols, etc)
  • A relay host type for forwarding karbon data to another Carbon-compatible host
  • Schema migration tools
  • log and stdout for torus and karbon