ceph/doc/mgr/prometheus.rst

=================
Prometheus plugin
=================

Provides a Prometheus exporter to pass on Ceph performance counters
from the collection point in ceph-mgr.  Ceph-mgr receives MMgrReport
messages from all MgrClient processes (mons and OSDs, for instance)
with performance counter schema data and actual counter data, and keeps
a circular buffer of the last N samples.  This plugin creates an HTTP
endpoint (like all Prometheus exporters) and retrieves the latest sample
of every counter when polled (or "scraped" in Prometheus terminology).
The HTTP path and query parameters are ignored; all extant counters
for all reporting entities are returned in text exposition format.
(See the Prometheus `documentation <https://prometheus.io/docs/instrumenting/exposition_formats/#text-format-details>`_.)

Enabling prometheus output
==========================

The *prometheus* module is enabled with::

  ceph mgr module enable prometheus

Configuration
-------------

By default the module will accept HTTP requests on port ``9283`` on all
IPv4 and IPv6 addresses on the host.  The port and listen address are both
configurable with ``ceph config-key set``, with keys
``mgr/prometheus/server_addr`` and ``mgr/prometheus/server_port``.
This port is registered with Prometheus's `registry <https://github.com/prometheus/prometheus/wiki/Default-port-allocations>`_.

Statistic names and labels
==========================

The names of the stats are exactly as Ceph names them, with
illegal characters ``.``, ``-`` and ``::`` translated to ``_``, 
and ``ceph_`` prefixed to all names.


All *daemon* statistics have a ``ceph_daemon`` label such as "osd.123"
that identifies the type and ID of the daemon they come from.  Some
statistics can come from different types of daemon, so when querying
e.g. an OSD's RocksDB stats, you would probably want to filter
on ceph_daemon starting with "osd" to avoid mixing in the monitor
rocksdb stats.


The *cluster* statistics (i.e. those global to the Ceph cluster)
have labels appropriate to what they report on.  For example, 
metrics relating to pools have a ``pool_id`` label.


The long running averages that represent the histograms from core Ceph
are represented by a pair of ``<name>_sum`` and ``<name>_count`` metrics.
This is similar to how histograms are represented in `Prometheus <https://prometheus.io/docs/concepts/metric_types/#histogram>`_
and they can also be treated `similarly <https://prometheus.io/docs/practices/histograms/>`_.

Pool and OSD metadata series
----------------------------

Special series are output to enable displaying and querying on
certain metadata fields.

Pools have a ``ceph_pool_metadata`` field like this:

::

    ceph_pool_metadata{pool_id="2",name="cephfs_metadata_a"} 1.0

OSDs have a ``ceph_osd_metadata`` field like this:

::

    ceph_osd_metadata{cluster_addr="172.21.9.34:6802/19096",device_class="ssd",ceph_daemon="osd.0",public_addr="172.21.9.34:6801/19096",weight="1.0"} 1.0


Correlating drive statistics with node_exporter
-----------------------------------------------

The prometheus output from Ceph is designed to be used in conjunction
with the generic host monitoring from the Prometheus node_exporter.

To enable correlation of Ceph OSD statistics with node_exporter's 
drive statistics, special series are output like this:

::

    ceph_disk_occupation{ceph_daemon="osd.0",device="sdd", exported_instance="myhost"}

To use this to get disk statistics by OSD ID, use either the ``and`` operator or
the ``*`` operator in your prometheus query. All metadata metrics (like ``
ceph_disk_occupation`` have the value 1 so they act neutral with ``*``. Using ``*``
allows to use ``group_left`` and ``group_right`` grouping modifiers, so that
the resulting metric has additional labels from one side of the query.

See the
`prometheus documentation`__ for more information about constructing queries.

__ https://prometheus.io/docs/prometheus/latest/querying/basics

The goal is to run a query like

::

    rate(node_disk_bytes_written[30s]) and on (device,instance) ceph_disk_occupation{ceph_daemon="osd.0"}

Out of the box the above query will not return any metrics since the ``instance`` labels of
both metrics don't match. The ``instance`` label of ``ceph_disk_occupation``
will be the currently active MGR node.

 The following two section outline two approaches to remedy this.

Use label_replace
=================

The ``label_replace`` function (cp.
`label_replace documentation <https://prometheus.io/docs/prometheus/latest/querying/functions/#label_replace>`_)
can add a label to, or alter a label of, a metric within a query.

To correlate an OSD and its disks write rate, the following query can be used:

::

    label_replace(rate(node_disk_bytes_written[30s]), "exported_instance", "$1", "instance", "(.*):.*") and on (device,exported_instance) ceph_disk_occupation{ceph_daemon="osd.0"}

Configuring Prometheus server
=============================

honor_labels
------------

To enable Ceph to output properly-labeled data relating to any host,
use the ``honor_labels`` setting when adding the ceph-mgr endpoints
to your prometheus configuration.

This allows Ceph to export the proper ``instance`` label without prometheus
overwriting it. Without this setting, Prometheus applies an ``instance`` label
that includes the hostname and port of the endpoint that the series came from.
Because Ceph clusters have multiple manager daemons, this results in an
``instance`` label that changes spuriously when the active manager daemon
changes.

If this is undesirable a custom ``instance`` label can be set in the
Prometheus target configuration: you might wish to set it to the hostname
of your first mgr daemon, or something completely arbitrary like "ceph_cluster".

node_exporter hostname labels
-----------------------------

Set your ``instance`` labels to match what appears in Ceph's OSD metadata
in the ``instance`` field.  This is generally the short hostname of the node.

This is only necessary if you want to correlate Ceph stats with host stats,
but you may find it useful to do it in all cases in case you want to do
the correlation in the future.

Example configuration
---------------------

This example shows a single node configuration running ceph-mgr and
node_exporter on a server called ``senta04``. Note that this requires to add the
appropriate instance label to every ``node_exporter`` target individually.

This is just an example: there are other ways to configure prometheus
scrape targets and label rewrite rules.

prometheus.yml
~~~~~~~~~~~~~~

::

    global:
      scrape_interval:     15s
      evaluation_interval: 15s

    scrape_configs:
      - job_name: 'node'
        file_sd_configs:
          - files:
            - node_targets.yml
      - job_name: 'ceph'
        honor_labels: true
        file_sd_configs:
          - files:
            - ceph_targets.yml


ceph_targets.yml
~~~~~~~~~~~~~~~~


::

    [
        {
            "targets": [ "senta04.mydomain.com:9283" ],
            "labels": {}
        }
    ]


node_targets.yml
~~~~~~~~~~~~~~~~

::

    [
        {
            "targets": [ "senta04.mydomain.com:9100" ],
            "labels": {
                "instance": "senta04"
            }
        }
    ]


Notes
=====

Counters and gauges are exported; currently histograms and long-running 
averages are not.  It's possible that Ceph's 2-D histograms could be 
reduced to two separate 1-D histograms, and that long-running averages
could be exported as Prometheus' Summary type.

Timestamps, as with many Prometheus exporters, are established by
the server's scrape time (Prometheus expects that it is polling the
actual counter process synchronously).  It is possible to supply a
timestamp along with the stat report, but the Prometheus team strongly
advises against this.  This means that timestamps will be delayed by
an unpredictable amount; it's not clear if this will be problematic,
but it's worth knowing about.