Cardinality budget

The budget is at most 2,000 series per game server, for everything that monitors it: this pack’s jobs plus the host and database exporters that run next to it. The pack’s own jobs stay far below that, and every job has a sample_limit, so a target that suddenly serves more series fails its scrape (up 0) instead of filling the database. Metrics reference lists every series.

Measured: the pack’s jobs

Measured on 2026-10-21 with scripts/smoke_exporters.py: Prometheus 3.15.0, json_exporter 0.8.0 and blackbox_exporter 0.29.0 with this repository’s configs, scraping the fake server (scrape_samples_post_metric_relabeling per target). The numbers do not depend on the server’s load or player count: the threads, the buckets and the probe phases are fixed.

Job and target	Samples per scrape	Plus Prometheus’ own series	sample_limit
fxserver_perf	51 (3 threads × (15 buckets + sum + count))	5	100
fxserver_json, module fxserver_dynamic	2	5	20
fxserver_json, module fxserver_info	2	5	20
fxserver_blackbox, module tcp_connect	8	5	60
fxserver_blackbox, module http_fxserver_info	18	5	60
fxserver_blackbox, module http_401_expected	18	5	60
Total, 6 targets	99	30

So one game server costs 129 series with the default jobs. Prometheus’ own series are up, scrape_duration_seconds, scrape_samples_scraped, scrape_samples_post_metric_relabeling and scrape_series_added, five per target. The optional fxserver_tunnel job adds one HTTPS probe (the HTTP series plus a few TLS series, not measured here, at most 60 by its limit) and five Prometheus series. The optional fxserver_versions job runs once per Prometheus, not per server: 1 sample per scrape (measured in the same run) and five Prometheus series.

The recording rules add 4 series per thread, 12 per game server (recorded series). Alerts add two series each (ALERTS and ALERTS_FOR_STATE) only while they are pending or firing.

The whole budget per server

Part	Series	How it is kept there
This pack, default jobs	129 measured; at most 350 by the sample limits (100 + 2 × 20 + 3 × 60, plus 30 from Prometheus)	sample_limit per job
This pack, tunnel probe	at most 65	sample_limit: 60
This pack, version list	at most 15, once per Prometheus	sample_limit: 10
This pack, recording rules	12 (4 series × 3 threads)	grouped by server and thread only
Reserve: node_exporter on the game host	900	not part of this pack: set sample_limit on your job and filter collectors you do not use
Reserve: mysqld_exporter	400	the same
Reserve: tunnel connector metrics	150	the same
Reserve: textfile metrics (for example the installed build and its age)	10	the same
Total	at most 1,902

The reserves are planning values for the exporters that monitor the host and the database; this repository does not ship their configs. With the measured 129 series for the default jobs instead of their limit of 350, and without the optional jobs, 399 of the 2,000 series stay free when the reserves are used up (319 with the tunnel probe and the version list at their limits).

What is left out, and why

• One series per resource. info.json lists every started resource; a server with a few hundred resources would add as many series. The pack collects no resource series.
• One series per player. Never collected, for privacy and for cardinality: a busy server would add series with every join. The player count is one series, fxserver_clients.
• Other /perf/ metrics. The scrape job keeps only the tick-time histogram, so a server version that adds metrics to /perf/ does not grow the budget.

What changes the numbers

• Each further game server adds the same 129 series (its own targets and server label), plus 12 recorded series.
• An upgrade changes the version and build labels of fxserver_info: one new series, while the old one stops getting samples.
• Another bucket layout in a future server version changes the 51 /perf/ series; the sample_limit of 100 holds up to 31 bucket series per thread (with +Inf).

Check it on your Prometheus

# Samples per scrape for each target of the pack (compare with the table above)
scrape_samples_post_metric_relabeling{job=~"fxserver_.*"}

# Series per game server, all jobs that carry the server label
count({server="game-01"})

# New series per scrape: should be 0 in steady state; a steady stream means churn
sum by (job) (scrape_series_added{job=~"fxserver_.*"})

Prometheus’ web interface also lists the metric names with the most series under Status → TSDB status.