Telegraf is a server-based agent that collects and sends metrics and events from databases, systems, and IoT sensors. Telegraf is an open source, plugin-driven tool for the collection and output of data.

To view metrics gathered by Telegraf and stored in a hypertable in a Timescale Cloud service.

Before you import your data:

To create a Telegraf configuration that exports data to a hypertable in your service:

  1. Setup your service connection string

    This variable hold the connection information for the target Timescale Cloud service.

    In Terminal on the source machine, set the following:

    export TARGET=postgres://tsdbadmin:<PASSWORD>@<HOST>:<PORT>/tsdb?sslmode=require

    You find the connection information for your Timescale Cloud service in the configuration file you downloaded when you created the service.

  2. Generate a Telegraf configuration file

    In Terminal, run the following:

    telegraf --input-filter=cpu --output-filter=postgresql config > telegraf.conf

    telegraf.conf configures a CPU input plugin that samples various metrics about CPU usage, and the PostgreSQL output plugin. telegraf.conf also includes all available input, output, processor, and aggregator plugins. These are commented out by default.

  3. Test the configuration

    telegraf --config telegraf.conf --test

    You see an output similar to the following:

    2022-11-28T12:53:44Z I! Starting Telegraf 1.24.3
    2022-11-28T12:53:44Z I! Available plugins: 208 inputs, 9 aggregators, 26 processors, 20 parsers, 57 outputs
    2022-11-28T12:53:44Z I! Loaded inputs: cpu
    2022-11-28T12:53:44Z I! Loaded aggregators:
    2022-11-28T12:53:44Z I! Loaded processors:
    2022-11-28T12:53:44Z W! Outputs are not used in testing mode!
    2022-11-28T12:53:44Z I! Tags enabled: host=localhost
    > cpu,cpu=cpu0,host=localhost usage_guest=0,usage_guest_nice=0,usage_idle=90.00000000087311,usage_iowait=0,usage_irq=0,usage_nice=0,usage_softirq=0,usage_steal=0,usage_system=6.000000000040018,usage_user=3.999999999996362 1669640025000000000
    > cpu,cpu=cpu1,host=localhost usage_guest=0,usage_guest_nice=0,usage_idle=92.15686274495818,usage_iowait=0,usage_irq=0,usage_nice=0,usage_softirq=0,usage_steal=0,usage_system=5.882352941192206,usage_user=1.9607843136712912 1669640025000000000
    > cpu,cpu=cpu2,host=localhost usage_guest=0,usage_guest_nice=0,usage_idle=91.99999999982538,usage_iowait=0,usage_irq=0,usage_nice=0,usage_softirq=0,usage_steal=0,usage_system=3.999999999996362,usage_user=3.999999999996362 1669640025000000000
  4. Configure the PostgreSQL output plugin

    1. In telegraf.conf, in the [[outputs.postgresql]] section, set connection to the value of $TARGET.

      connection = "<VALUE OF $TARGET>"
    2. Use hypertables when Telegraf creates a new table:

      In the section that begins with the comment ## Templated statements to execute when creating a new table, add the following template:

      ## Templated statements to execute when creating a new table.
      # create_templates = [
      # '''CREATE TABLE {{ .table }} ({{ .columns }})''',
      # ]
      # table_template=`CREATE TABLE IF NOT EXISTS {TABLE}({COLUMNS}); SELECT create_hypertable({TABLELITERAL},by_range('time', INTERVAL '1 week'),if_not_exists := true);`

      The by_range dimension builder was added to TimescaleDB 2.13.

This section shows you how to generate system metrics using Telegraf, then connect to your service and query the metrics hypertable.

  1. Collect system metrics using Telegraf

    Run the following command for a 30 seconds:

    telegraf --config telegraf.conf

    Telegraf uses loaded inputs cpu and outputs postgresql along with global tags, the intervals when the agent collects data from the inputs, and flushes to the outputs.

  2. View the metrics

    1. Connect to your Timescale Cloud service:

      psql $TARGET
    2. View the metrics collected in the cpu table in tsdb:

      SELECT*FROM cpu;

      You see something like:

      time | cpu | host | usage_guest | usage_guest_nice | usage_idle | usage_iowait | usage_irq | usage_nice | usage_softirq | usage_steal | usage_system | usage_user
      ---------------------+-----------+----------------------------------+-------------+------------------+-------------------+--------------+-----------+------------+---------------+-------------+---------------------+---------------------
      2022-12-05 12:25:20 | cpu0 | hostname | 0 | 0 | 83.08605341237833 | 0 | 0 | 0 | 0 | 0 | 6.824925815961274 | 10.089020771444481
      2022-12-05 12:25:20 | cpu1 | hostname | 0 | 0 | 84.27299703278959 | 0 | 0 | 0 | 0 | 0 | 5.934718100814769 | 9.792284866395647
      2022-12-05 12:25:20 | cpu2 | hostname | 0 | 0 | 87.53709198848934 | 0 | 0 | 0 | 0 | 0 | 4.747774480755411 | 7.715133531241037
      2022-12-05 12:25:20 | cpu3 | hostname| 0 | 0 | 86.68639053296472 | 0 | 0 | 0 | 0 | 0 | 4.43786982253345 | 8.875739645039992
      2022-12-05 12:25:20 | cpu4 | hostname | 0 | 0 | 96.15384615371369 | 0 | 0 | 0 | 0 | 0 | 1.1834319526667423 | 2.6627218934917614

      To view the average usage per CPU core, use SELECT cpu, avg(usage_user) FROM cpu GROUP BY cpu;.

For more information about the options that you can configure in Telegraf, see the PostgreQL output plugin.

Keywords

Found an issue on this page?Report an issue or Edit this page in GitHub.