cassandra-exporter CircleCI

cassandra-exporter is a Java agent (with optional standalone mode) that exports Cassandra metrics to Prometheus.

Project Status: beta


cassandra-exporter enables high performance collection of Cassandra metrics and follows the Prometheus best practices for metrics naming and labeling.

Benchamrk Results

cassandra-exporter is fast. In a worst-case benchmark, where the Cassandra schema contains 1000+ tables (resulting in ~174 thousand metrics), cassandra-exporter completes exposition in ~140ms. Compared to the next-best, jmx_exporter, which completes exposition in ~8 seconds. Other solutions can take tens of seconds, during which CPU time is consumed querying JMX and serialising values.

See the Exported Metrics wiki page for a list of available metrics.

All but a few select metrics exposed by cassandra-exporter are live with no caching involved. The few that are cached are done so for performance reasons.

cassandra-exporter exports metric families, where the names, labels, metric types (gauge, counter, summary, etc), and value scales have been hand-tuned to produce easy-to-query output.

For example, the following PromQL query will return an estimate of the number of pending compactions per-keyspace, per-node.

sum(cassandra_table_estimated_pending_compactions) by (cassandra_node, keyspace)


cassandra-exporter is has been tested with:

Component Version
Apache Cassandra 3.0.17 (experimental), 3.11.2, 3.11.3
Prometheus 2.0 and later

Other Cassandra and Prometheus versions will be tested for compatibility in the future.



Download the latest release and copy cassandra-exporter-agent-<version>.jar to $CASSANDRA_HOME/lib (typically /usr/share/cassandra/lib in most package installs).

Then edit $CASSANDRA_CONF/ (typically /etc/cassandra/ and append the following:

JVM_OPTS="$JVM_OPTS -javaagent:$CASSANDRA_HOME/lib/cassandra-exporter-agent-<version>.jar"

Then (re-)start Cassandra.

Prometheus metrics will now be available at http://localhost:9500/metrics.


While it is preferable to run cassandra-exporter as a Java agent for performance, it can instead be run as an external application if required.

Download the latest release and copy cassandra-exporter-standalone-<version>.jar to a location of your choosing.

The exporter can be started via java -jar /path/to/cassandra-exporter-standalone-<version>.jar.

Prometheus metrics will now be available at http://localhost:9500/metrics.

In this mode metrics will be queried via JMX which will incur a performance overhead. The standalone mode was originally designed to assist with benchmarking and development of the exporter.

The set of metrics available is close to that of the agent -- Gossiper related metrics are unavailable as these aren't readily available over JMX.

Currently some additional metadata labels, such as the table type (table, index, view, etc) attached to the cassandra_table_* metrics, are not available (this feature has yet to be written).

Prometheus Configuration

Configure Prometheus to scrape the endpoint by adding the following to prometheus.yml:


  - job_name: 'cassandra'
      - targets: ['<cassandra node IP>:9500', '<cassandra node IP>:9500', '<cassandra node IP>:9500', ...]

See the Prometheus documentation for more details on configuring scrape targets.

To view the raw, plain text metrics (in the Prometheus text exposition format), request the endpoint (by default, http://localhost:9500/metrics) with a HTTP client such as a browser or cURL.

Experimental JSON output is also provided if the Accept: application/json header or ?x-accept=application/json URL parameter is specified. The format/structure of the JSON output is subject to change.


The available command line options may be seen by passing -h/--help:

Usage: cassandra-exporter [-hV] [--enable-collector-timing]
                          [--exclude-system-tables] [--no-fast-float]
                          [--no-global-labels] [--no-table-labels] [-v]...
                          [--cql-password=PASSWORD] [--cql-user=NAME]
                          [--family-help=VALUE] [--jmx-password=PASSWORD]
                          [--jmx-service-url=URL] [--jmx-user=NAME]
                          [--keyspace-metrics=FILTER] [--node-metrics=FILTER]
                          [--exclude-keyspaces=<excludedKeyspaces>]... [-g=LABEL
                          [,LABEL...]]... [-l=[ADDRESS][:PORT]]... [-t=LABEL[,
                          LABEL...]]... [-e=EXCLUSION...]...
  -g, --global-labels=LABEL[,LABEL...]
                            Select which global labels to include on all exported
                              metrics. Valid options are: 'CLUSTER' (cluster name),
                              'NODE' (node endpoint IP address), 'DATACENTER' (DC
                              name), 'RACK' (rack name). The default is to include
                              all global labels except HOST_ID. To disable all
                              global labels use --no-global-labels.
  -t, --table-labels=LABEL[,LABEL...]
                            Select which labels to include on table-level metrics.
                              Valid options are: 'TABLE_TYPE' (table, view or
                              index), 'INDEX_TYPE' (for indexes -- keys, composites
                              or custom), 'INDEX_CLASS' (the index class name for
                              custom indexes),  'COMPACTION_STRATEGY_CLASS' (for
                              tables & views, compaction-related metrics only). The
                              default is to include all table labels. To disable all
                              table labels use --no-table-labels.
                            Select which table-level metrics to expose. Valid
                              options are: 'ALL' (all metrics), 'HISTOGRAMS' (only
                              histograms & summaries), 'NONE' (no metrics). The
                              default is 'ALL'.
                            Select which keyspace-level aggregate metrics to expose.
                              Valid options are: 'ALL' (all metrics), 'HISTOGRAMS'
                              (only histograms & summaries), 'NONE' (no metrics).
                              The default is 'HISTOGRAMS'.
      --node-metrics=FILTER Select which node-level aggregate metrics to expose.
                              Valid options are: 'ALL' (all metrics), 'HISTOGRAMS'
                              (only histograms & summaries), 'NONE' (no metrics).
                              The default is 'HISTOGRAMS'.
                            Collect per-thread CPU times, where each thread gets its
                              own time-series. (EXPERIMENTAL)
                            Record the cumulative time taken to run each collector
                              and export the results.

  -e, --exclude=EXCLUSION...
                            Exclude a metric family or MBean from exposition.
                              EXCLUSION may be the full name of a metric family
                              (wildcards or patterns not allowed) or the ObjectName
                              of a MBean or a ObjectName pattern that matches
                              multiple MBeans. ObjectNames always contain a colon
                              (':'). See the ObjectName JavaDoc for details. If
                              EXCLUSION is prefixed with an '@', it is interpreted
                              (sans @ character) as a path to a file containing
                              multiple EXCLUSION values, one per line. Lines
                              prefixed with '#' are considered comments and are
                              ignored. This option may be specified more than once
                              to define multiple exclusions.
      --no-global-labels    Disable all global labels.
      --no-table-labels     Disable all table labels.
      --no-fast-float       Disable the use of fast float -> ascii conversion.
                            Exclude system table/keyspace metrics.
  -l, --listen=[ADDRESS][:PORT]
                            Listen address (and optional port). ADDRESS may be a
                              hostname, IPv4 dotted or decimal address, or IPv6
                              address. When ADDRESS is omitted, (wildcard)
                              is substituted. PORT, when specified, must be a valid
                              port number. The default port 9500 will be substituted
                              if omitted. If ADDRESS is omitted but PORT is
                              specified, PORT must be prefixed with a colon (':'),
                              or PORT will be interpreted as a decimal IPv4 address.
                              This option may be specified more than once to listen
                              on multiple addresses. Defaults to ''
      --family-help=VALUE   Include or exclude metric family help in the exposition
                              format. AUTOMATIC excludes help strings when the user
                              agent is Prometheus and includes them for all other
                              clients (cURL, browsers, etc). Currently Prometheus
                              discards help strings. Excluding help strings saves
                              bandwidth. Can be overridden with the "?
                              help=true|false" URI query parameter. Valid values:
                              INCLUDE, EXCLUDE, AUTOMATIC. Defaults to AUTOMATIC.
      --jmx-service-url=URL JMX service URL of the Cassandra instance to connect to
                              and collect metrics. Defaults to 'service:jmx:rmi:
      --jmx-user=NAME       JMX authentication user name.
                            JMX authentication password.
                            Address/hostname and optional port for the CQL metadata
                              connection. Defaults to 'localhost:9042'
      --cql-user=NAME       CQL authentication user name.
                            CQL authentication password.
  -v, --verbose             Enable verbose logging. Multiple invocations increase
                              the verbosity.
  -h, --help                Show this help message and exit.
  -V, --version             Print version information and exit.

Options may also be provided via an @-file:

Note that --jmx-service-url, --jmx-user, --jmx-password, --cql-address, --cql-user and --cql-password are only applicable to the standalone version -- the agent does not use JMX or CQL connections.

To protect the JMX password and prevent it from showing up in ps, top and other utilities, use an @-file that contains --jmx-password=PASSWORD.

When run as an agent, command line options must be provided as part of the -javaagent flag, with an equals sign (=) separating the JAR path and the agent options. Multiple options, or option arguments can be separated by commas (,) or spaces. Commas are preferred as the whitespace quoting rules of are quite complex. Options with values containing whitespace must be quoted appropriately. Alternatively use an @-file (see above).

For example, to change the agent listening port to 1234 and exclude some metrics:

JVM_OPTS="$JVM_OPTS -javaagent:$CASSANDRA_HOME/lib/cassandra-exporter-agent-<version>.jar=--listen=:1234,[email protected]$CASSANDRA_CONF/prometheus-exclusions"




JMX is slow, really slow. JMX adds significant overhead to every method invocation on exported MBean methods, even when those methods are called from within the same JVM. On a 300-ish table Cassandra node, trying to collect all exposed metrics via JVM resulted in a collection time that was upwards of 2-3 seconds. For exporters that run as a separate process there is additional overhead of inter-process communications and that time can reach the 10's of seconds.

cassandra-exporter on the same node collects all metrics in 10-20 milliseconds.

Best practices

The exporter attempts to follow Prometheus' best practices for metric names, labels and data types.

Cassandra has keyspace- and node-level metrics that are aggregates of the per-table metrics. By default, only a subset of these aggregate metrics, specifically histograms, are exposed by cassandra-exporter. All other keyspace- and node-level metrics are skipped in favour of only exporting the per-table metrics. The rationale behind this is that apart from the histograms, the aggregate metrics are essentially duplicates. If they are needed they may be computed on-the-fly via PromQL or once, at scrape time, using Prometheus recording rules.

Unlike the metrics exported via JMX, where each table metric has a unique name, Cassandras metrics are coalesced when appropriate so they share the same exported metric family name, opting for labels to differentiate individual time series. For example, each table level metric has a constant name and at minimum a table & keyspace label, which allows for complex PromQL queries.

For example the cassandra_table_operation_latency_seconds[_count|_sum] summary metric combines read, write, range read, CAS prepare, CAS propose and CAS commit latency metrics together into a single metric family. A summary exposes percentiles (via the quantile label), a total count of recorded samples (via the _count metric), and (if available, NaN otherwise) an accumulated sum of all samples (via the _sum metric).

Individual time-series are separated by different labels. In this example, the operation type is exported as the operation label. The source keyspace, table, table_type (table, view or index), table_id (CF UUID), and numerous other metadata labels are available.



These metrics can then be queried:

sum(cassandra_table_operation_latency_seconds_count) by (keyspace, operation) # total operations by keyspace & type
Element Value
{keyspace="system",operation="write"} 13989
{keyspace="system",operation="cas_commit"} 0
{keyspace="system",operation="cas_prepare"} 0
{keyspace="system",operation="cas_propose"} 0
{keyspace="system",operation="range_read"} 10894
{keyspace="system",operation="read"} 74
{keyspace="system_schema",operation="write"} 78
{keyspace="system_schema",operation="cas_commit"} 0
{keyspace="system_schema",operation="cas_prepare"} 0
{keyspace="system_schema",operation="cas_propose"} 0
{keyspace="system_schema",operation="range_read"} 75
{keyspace="system_schema",operation="read"} 618

Global Labels

The exporter does attach global labels to the exported metrics. These may be configured with the --global-labels (or disabled via --no-global-labels) CLI option.

These labels are:

These labels allow aggregation of metrics at the cluster, data center and rack levels.

While these labels could be defined in the Prometheus scrape config, we feel that having these labels be automatically applied simplifies things, especially when Prometheus is monitoring multiple clusters across numerous DCs and racks.

Exported Metrics

See the Exported Metrics wiki page for a list.

We suggest viewing the metrics endpoint (e.g., http://localhost:9500/metrics) in a browser to get an understanding of what metrics are exported by your Cassandra node.

Unstable, Missing & Future Features

See the project issue tracker for a complete list.

Please see for Instaclustr support status of this project.