Alternative storage finders

Built-in finders

The default graphite setup consists of:

  • A Whisper database
  • A carbon daemon writing data to the database
  • Graphite-web reading and graphing data from the database

It is possible to use an alternate storage layer than the default, Whisper, in order to accommodate specific needs. The setup above would become:

  • An alternative database
  • A carbon daemon or alternative daemon for writing to the database
  • A custom storage finder for reading the data in graphite-web

This section aims at documenting the last item: configuring graphite-web to read data from a custom storage layer.

This can be done via the STORAGE_FINDERS setting. This setting is a list of paths to finder implementations. Its default value is:


The default finder reads data from a Whisper database.

An alternative finder for the experimental Ceres database is available:


The setting supports multiple values, meaning you can read data from both a Whisper database and a Ceres database:


Custom finders

STORAGE_FINDERS being a list of arbitrary python paths, it is relatively easy to write a custom finder if you want to read data from other places than Whisper and Ceres. A finder is a python class with a find_nodes() method:

class CustomFinder(object):
    def find_nodes(self, query):
        # ...

query is a FindQuery object. find_nodes() is the entry point when browsing the metrics tree. It must yield leaf or branch nodes matching the query:

from graphite.node import LeafNode, BranchNode
from graphite.finders.utils import BaseFinder

class CustomFinder(BaseFinder):
    def find_nodes(self, query):
        # find some paths matching the query, then yield them
        for path in matches:
            if is_branch(path):
                yield BranchNode(path)
            if is_leaf(path):
                yield LeafNode(path, CustomReader(path))

LeafNode is created with a reader, which is the class responsible for fetching the datapoints for the given path. It is a simple class with 2 methods: fetch() and get_intervals():

from graphite.intervals import IntervalSet, Interval
from graphite.readers.utils import BaseReader

class CustomReader(BaseReader):
    __slots__ = ('path',)  # __slots__ is recommended to save memory on readers

    def __init__(self, path):
        self.path = path

    def fetch(self, start_time, end_time):
        # fetch data
        time_info = _from_, _to_, _step_
        return time_info, series

    def get_intervals(self):
        return IntervalSet([Interval(start, end)])

fetch() must return a list of 2 elements: the time info for the data and the datapoints themselves. The time info is a list of 3 items: the start time of the datapoints (in unix time), the end time and the time step (in seconds) between the datapoints.

The datapoints is a list of points found in the database for the required interval. There must be (end - start) / step points in the dataset even if the database has gaps: gaps can be filled with None values.

get_intervals() is a method that hints graphite-web about the time range available for this given metric in the database. It must return an IntervalSet of one or more Interval objects.

Advanced finders

Custom finders may also implement the following methods:


This class method is responsible for initializing and returning the finder object(s) as a list.

It may return a list of 1 or more instances of the finder, if multiple instances are returned they will be called concurrently in multiple threads. This is used by RemoteFinder to dispatch requests to multiple remote hosts in parallel.

If not defined, a single instance of the finder will be initialized with no parameters.

get_index(self, requestContext)

This method should return all node paths that the finder is aware of as a list of strings.

requestContext is a dict which may contain localOnly and forwardHeaders keys.

If not implemented, find_nodes() will be called with a query for ** and a list of the returned nodes’ paths will be returned.

find_multi(self, queries)

This method follows the same semantics as find_node() but accepts a list of queries.

If not implemented, find_nodes() will be called for each query specified.

fetch(self, patterns, start_time, end_time, now=None, requestContext=None)

This method is responsible for loading data for render requests.

It should return a list of result dicts, each of which contains:

  'pathExpression': '<the pattern that this path matched>',
  'path': 'the.metric.path',
  'name': 'the.metric.path',
  'time_info': (_from_, _to_, _step_),
  'values': [list of values],

If not implemented, find_multi() will be called with a list of queries and node.fetch() will be called on every result.

auto_complete_tags(self, exprs, tagPrefix=None, limit=None, requestContext=None)

This method is only used when tags = True is specified in the class definition.

If defined it should return an auto-complete list of tags for series that match the specified expressions.

auto_complete_values(self, exprs, tag, valuePrefix=None, limit=None, requestContext=None)

This method is only used when tags = True is specified in the class definition.

If defined it should return an auto-complete list of values for the specified tag on series that match the specified expressions.

Installing custom finders

In order for your custom finder to be importable, you need to package it under a namespace of your choice. Python packaging won’t be covered here but you can look at third-party finders to get some inspiration: