Downsample metrics before writing

Collect at high resolution, keep the raw data briefly, and store compact statistical summaries long-term. This configuration collects CPU and memory metrics every 10 seconds, writes the raw metrics to one database, and writes 5-minute min, max, and mean summaries to another.

Configuration

[agent]
  interval = "10s"

[[inputs.cpu]]
  totalcpu = true

[[inputs.mem]]

[[aggregators.basicstats]]
  ## Summarize each 5-minute window.
  period = "5m"
  stats = ["min", "max", "mean"]

  ## Keep the raw metrics flowing to the outputs.
  drop_original = false

  ## Rename the aggregates so outputs can route on the name.
  name_suffix = "_5m"

[[outputs.influxdb_v3]]
  ## Raw metrics: everything except the renamed aggregates.
  urls = ["http://localhost:8181"]
  token = "
AUTH_TOKEN
"
  database = "
RAW_DATABASE_NAME
"
  namedrop = ["*_5m"]

[[outputs.influxdb_v3]]
  ## Aggregates only.
  urls = ["http://localhost:8181"]
  token = "
AUTH_TOKEN
"
  database = "
SUMMARY_DATABASE_NAME
"
  namepass = ["*_5m"]

Replace the following:

• AUTH_TOKEN: your InfluxDB authorization token
• RAW_DATABASE_NAME: the database for full-resolution metrics, typically with a short retention period
• SUMMARY_DATABASE_NAME: the database for long-term summaries

How it works

• aggregators.basicstats collects matching metrics into 5-minute windows and emits <field>_min, <field>_max, and <field>_mean fields at the end of each period.
• drop_original = false (the default) lets the raw metrics continue to the outputs alongside the aggregates. The result is two streams in one pipeline: raw cpu and mem metrics every 10 seconds, plus cpu_5m and mem_5m aggregates every 5 minutes.
• name_suffix = "_5m" renames the aggregate metrics, which is what makes them routable: the raw output excludes them with namedrop = ["*_5m"], and the summary output selects only them with namepass = ["*_5m"].
• Each output batches and delivers independently, so a slow or unavailable destination doesn’t block the other.

To keep only the summaries and discard raw data entirely, set drop_original = true and remove the raw output. See Why processors run twice for how aggregates flow through the pipeline.

Example output

Raw metrics, every 10 seconds:

cpu,cpu=cpu-total,host=host1 usage_idle=92.4,usage_user=4.2 1709572230000000000
mem,host=host1 used_percent=64.2 1709572230000000000

Aggregates, every 5 minutes:

cpu_5m,cpu=cpu-total,host=host1 usage_idle_min=88.1,usage_idle_max=94.6,usage_idle_mean=91.7,usage_user_min=3.1,usage_user_max=6.8,usage_user_mean=4.4 1709572500000000000
mem_5m,host=host1 used_percent_min=63.8,used_percent_max=65.1,used_percent_mean=64.3 1709572500000000000

Extend this example

• Swap or add aggregators for different summaries: quantiles, histograms, or final values. See Common aggregator examples.
• Scope the aggregator to specific measurements with namepass on the aggregator itself.
• For more routing patterns, see Route metrics to different outputs.