VictoriaMetrics Data Model

VictoriaMetrics uses the Prometheus data model: a metric name plus a set of label key-value pairs identifies a series, and each series holds timestamped float values.

http_requests_total{job="api", instance="10.0.1.5:8080", method="GET", status="200"}

Ingestion

It accepts several protocols, which makes it a drop-in replacement in most stacks:

• Prometheus remote write — the usual path.
• Prometheus scraping, performed by VictoriaMetrics itself (vmagent or -promscrape.config).
• InfluxDB line protocol, Graphite, OpenTSDB, CSV and JSON import.

# prometheus.yml
remote_write:
  - url: http://victoriametrics:8428/api/v1/write
    queue_config:
      max_samples_per_send: 10000
      capacity: 20000

vmagent sits in front for scraping, relabeling, buffering during backend outages and fan-out to several destinations. Its on-disk buffer is what prevents data loss during a backend restart.

MetricsQL

MetricsQL is a superset of PromQL: existing queries work, with additional functions and some behavioural differences that reduce common PromQL surprises.

# Rate over a window, aggregated by status.
sum(rate(http_requests_total{job="api"}[5m])) by (status)
 
# 99th percentile latency from a histogram.
histogram_quantile(0.99, sum(rate(http_request_duration_seconds_bucket[5m])) by (le))
 
# MetricsQL additions.
rollup_rate(http_requests_total[5m])

Stream aggregation

Aggregating at ingestion is the primary defence against cardinality growth. VictoriaMetrics can do it in vmagent or at the storage layer, before the raw series are stored:

# stream aggregation config
- match: 'http_requests_total'
  interval: 1m
  outputs: [total]
  without: [instance]        # drop per-instance detail, keep the service-level series

This turns thousands of per-instance series into a handful of per-service ones, permanently.

Relabeling

metric_relabel_configs:
  # Drop a high-cardinality label entirely.
  - regex: 'request_id|trace_id'
    action: labeldrop
  # Drop metrics you never query.
  - source_labels: [__name__]
    regex: 'go_gc_.*'
    action: drop

Storage characteristics

VictoriaMetrics stores data in per-month partitions with columnar compression, and merges parts in the background — an LSM-like design. Two practical consequences:

• Retention is a partition drop, so it is cheap.
• Free disk space is needed for merges, so plan capacity above the steady-state size.

Out-of-order and duplicate samples are handled: it accepts them and deduplicates on read according to the configured deduplication interval, which is what makes highly-available Prometheus pairs straightforward to consolidate.