thoughtpolice/remote-tsdb-clickhouse

remote-tsdb-clickhouse stores timeseries data in ClickHouse.

Implements both Prometheus remote writer (to store metrics) and Prometheus remote reader (use metrics from ClickHouse directly in Prometheus)

Install

go install github.com/jamessanford/remote-tsdb-clickhouse@latest

Create destination table

Use clickhouse client to create this table:

CREATE TABLE metrics.samples
(
    `updated_at` DateTime CODEC(DoubleDelta, LZ4),
    `metric_name` LowCardinality(String),
    `labels` Array(LowCardinality(String)),
    `value` Float64 CODEC(Gorilla, LZ4),
    INDEX labelset (labels, metric_name) TYPE set(0) GRANULARITY 8192
)
ENGINE = MergeTree
ORDER BY (metric_name, labels, updated_at)
SETTINGS index_granularity = 8192

This works well with over 30 billion metrics, even when searching by label, although cardinality of my dataset is low at 16032 unique metrics+labels. Including label values, it takes approximately 1 byte per value for my dataset (1 gigabyte per billion metrics)

The labelset index granularity is set to 8192 (8192*8192 rows) on purpose for queries like has(labels, 'job=omada') while still providing performance with many rows that match.

Configure Prometheus remote writer

In your prometheus.yaml:

remote_write:
 - url: "http://localhost:9131/write"

Configure Prometheus remote reader

In your prometheus.yaml:

remote_read:
 - url: "http://localhost:9131/read"

Query data with Prometheus

The above configuration will use remote-tsdb-clickhouse to backfill data not present in Prometheus.

If you'd like to query remote-tsdb-clickhouse immediately, consider this configuration:

remote_read:
 - url: "http://localhost:9131/read"
   read_recent: true
   name: clickhouse
   required_matchers:
     remote: clickhouse

Then issue queries with the added {remote="clickhouse"} label.

remote-tsdb-clickhouse will remove the {remote="clickhouse"} label from incoming requests by default, see --help.

Query ClickHouse directly with Grafana

Use the ClickHouse Data Plugin for Grafana pointed at ClickHouse (eg localhost:8123)

Queries to get you started:

$perSecondColumns(arrayConcat([metric_name], labels), value)
FROM metrics.samples
WHERE
    metric_name='go_memstats_alloc_bytes_total'
    AND has(labels, 'job=omada')

$perSecondColumns(arrayConcat([metric_name], arrayFilter(x -> x LIKE 'name=%', labels)), value * 8)
FROM metrics.samples
WHERE
    metric_name='omada_station_transmit_bytes_total'

SELECT
    $timeSeries as t,
    metric_name,
    labels,
    max(value)
FROM $table
WHERE
    metric_name='go_goroutines'
    AND has(labels, 'job=omada')
    AND $timeFilter
GROUP BY
    metric_name,
    labels,
    t
ORDER BY t

SELECT
    t,
    if(runningDifference(max_0) < 0, nan, runningDifference(max_0) / runningDifference(t / 1000)) AS max_0_Rate
FROM
(
    SELECT
        $timeSeries AS t,
        max(value) as max_0
    FROM $table
    WHERE metric_name='go_memstats_alloc_bytes_total'
    AND has(labels, 'job=omada')
    AND $timeFilter
    GROUP BY t
    ORDER BY t
)

Importing existing data

You may export TSDB data from Prometheus and reinsert it into ClickHouse.

Use a modified promtool command to dump one day at a time.

Note that promtool tsdb writes to your TSDB directory, so run it against a read-only snapshot.

promtool tsdb dump \
  --min-time=$(date -u -d '2021-12-16' +%s)001 \
  --max-time=$(date -u -d '2021-12-17' +%s)000 \
  /zfs/tsdbsnap1/jsanford/prom2/bin/data \
|  clickhouse client \
  --query 'INSERT INTO metrics.samples FORMAT TabSeparated'

You may significantly speed up the bulk import by running many in parallel.

Importing one day a time makes it easy to delete and reimport data, eg

ALTER TABLE metrics.samples DELETE WHERE updated_at > 1656806400 AND updated_at <= 1656892800

Let ClickHouse settle for 30 minutes or so after bulk importing data before determining what CPU usage will look like long term.

thoughtpolice / remote-tsdb-clickhouse