a65e7e9b59
Currently we are using `instance` label to query about host metrics to prometheus. This label is assigned to the url of each endpoint being scrapped. While this work fine in one-exporter-per-compute cases as the driver is mapping the fqdn_label value to the `instance` label value, it fails when there are more that one target with the same value for the fqdn label. This is a valid case, to be able to query by fqdn and do not care about what exporter in the host is providing the metric. This patch is changing the queries we use for hosts to be based on the fqdn_label instead of the instance one. To implement it, we are also simplifying the way we check the metric exist for the host by converting prometheus_fqdn_instance_map into a prometheus_fqdn_labels set which stores the list of fqdn found in prometheus. Closes-Bug: #2103451 Change-Id: I3bcc317441b73da5c876e53edd4622370c6d575e
141 lines
6.0 KiB
ReStructuredText
141 lines
6.0 KiB
ReStructuredText
=====================
|
|
Prometheus datasource
|
|
=====================
|
|
|
|
Synopsis
|
|
--------
|
|
The Prometheus datasource allows Watcher to use a Prometheus server as the
|
|
source for collected metrics used by the Watcher decision engine. At minimum
|
|
deployers must configure the ``host`` and ``port`` at which the Prometheus
|
|
server is listening.
|
|
|
|
Requirements
|
|
-------------
|
|
It is required that Prometheus metrics contain a label to identify the hostname
|
|
of the exporter from which the metric was collected. This is used to match
|
|
against the Watcher cluster model ``ComputeNode.hostname``. The default for
|
|
this label is ``fqdn`` and in the prometheus scrape configs would look like:
|
|
|
|
.. code-block::
|
|
|
|
scrape_configs:
|
|
- job_name: node
|
|
static_configs:
|
|
- targets: ['10.1.2.3:9100']
|
|
labels:
|
|
fqdn: "testbox.controlplane.domain"
|
|
|
|
This default can be overridden when a deployer uses a different label to
|
|
identify the exporter host (for example ``hostname`` or ``host``, or any other
|
|
label, as long as it identifies the host).
|
|
|
|
Internally this label is used in creating ``fqdn_instance_labels``, containing
|
|
the list of values assigned to the the label in the Prometheus targets.
|
|
The elements of the resulting fqdn_instance_labels are expected to match the
|
|
``ComputeNode.hostname`` used in the Watcher decision engine cluster model.
|
|
An example ``fqdn_instance_labels`` is the following:
|
|
|
|
.. code-block::
|
|
|
|
[
|
|
'ena.controlplane.domain',
|
|
'dio.controlplane.domain',
|
|
'tria.controlplane.domain',
|
|
]
|
|
|
|
For instance metrics, it is required that Prometheus contains a label
|
|
with the uuid of the OpenStack instance in each relevant metric. By default,
|
|
the datasource will look for the label ``resource``. The
|
|
``instance_uuid_label`` config option in watcher.conf allows deployers to
|
|
override this default to any other label name that stores the ``uuid``.
|
|
|
|
Limitations
|
|
-----------
|
|
The current implementation doesn't support the ``statistic_series`` function of
|
|
the Watcher ``class DataSourceBase``. It is expected that the
|
|
``statistic_aggregation`` function (which is implemented) is sufficient in
|
|
providing the **current** state of the managed resources in the cluster.
|
|
The ``statistic_aggregation`` function defaults to querying back 300 seconds,
|
|
starting from the present time (the time period is a function parameter and
|
|
can be set to a value as required). Implementing the ``statistic_series`` can
|
|
always be re-visited if the requisite interest and work cycles are volunteered
|
|
by the interested parties.
|
|
|
|
One further note about a limitation in the implemented
|
|
``statistic_aggregation`` function. This function is defined with a
|
|
``granularity`` parameter, to be used when querying whichever of the Watcher
|
|
``DataSourceBase`` metrics providers. In the case of Prometheus, we do not
|
|
fetch and then process individual metrics across the specified time period.
|
|
Instead we use the PromQL querying operators and functions, so that the
|
|
server itself will process the request across the specified parameters and
|
|
then return the result. So ``granularity`` parameter is redundant and remains
|
|
unused for the Prometheus implementation of ``statistic_aggregation``. The
|
|
granularity of the data fetched by Prometheus server is specified in
|
|
configuration as the server ``scrape_interval`` (current default 15 seconds).
|
|
|
|
Configuration
|
|
-------------
|
|
A deployer must set the ``datasources`` parameter to include ``prometheus``
|
|
under the watcher_datasources section of watcher.conf (or add ``prometheus`` in
|
|
datasources for a specific strategy if preferred eg. under the
|
|
``[watcher_strategies.workload_stabilization]`` section).
|
|
|
|
The watcher.conf configuration file is also used to set the parameter values
|
|
required by the Watcher Prometheus data source. The configuration can be
|
|
added under the ``[prometheus_client]`` section and the available options are
|
|
duplicated below from the code as they are self documenting:
|
|
|
|
.. code-block::
|
|
|
|
cfg.StrOpt('host',
|
|
help="The hostname or IP address for the prometheus server."),
|
|
cfg.StrOpt('port',
|
|
help="The port number used by the prometheus server."),
|
|
cfg.StrOpt('fqdn_label',
|
|
default="fqdn",
|
|
help="The label that Prometheus uses to store the fqdn of "
|
|
"exporters. Defaults to 'fqdn'."),
|
|
cfg.StrOpt('instance_uuid_label',
|
|
default="resource",
|
|
help="The label that Prometheus uses to store the uuid of "
|
|
"OpenStack instances. Defaults to 'resource'."),
|
|
cfg.StrOpt('username',
|
|
help="The basic_auth username to use to authenticate with the "
|
|
"Prometheus server."),
|
|
cfg.StrOpt('password',
|
|
secret=True,
|
|
help="The basic_auth password to use to authenticate with the "
|
|
"Prometheus server."),
|
|
cfg.StrOpt('cafile',
|
|
help="Path to the CA certificate for establishing a TLS "
|
|
"connection with the Prometheus server."),
|
|
cfg.StrOpt('certfile',
|
|
help="Path to the client certificate for establishing a TLS "
|
|
"connection with the Prometheus server."),
|
|
cfg.StrOpt('keyfile',
|
|
help="Path to the client key for establishing a TLS "
|
|
"connection with the Prometheus server."),
|
|
|
|
The ``host`` and ``port`` are **required** configuration options which have
|
|
no set default. These specify the hostname (or IP) and port for at which
|
|
the Prometheus server is listening. The ``fqdn_label`` allows deployers to
|
|
override the required metric label used to match Prometheus node exporters
|
|
against the Watcher ComputeNodes in the Watcher decision engine cluster data
|
|
model. The default is ``fqdn`` and deployers can specify any other value
|
|
(e.g. if they have an equivalent but different label such as ``host``).
|
|
|
|
So a sample watcher.conf configured to use the Prometheus server at
|
|
``10.2.3.4:9090`` would look like the following:
|
|
|
|
.. code-block::
|
|
|
|
[watcher_datasources]
|
|
|
|
datasources = prometheus
|
|
|
|
[prometheus_client]
|
|
|
|
host = 10.2.3.4
|
|
port = 9090
|
|
fqdn_label = fqdn
|