The Cardinality Trap: How One Label Blew Up a Metrics Bill

June 24, 2026

Placeholder content. This is not original work. This article is a stand-in used to demonstrate Markdown rendering, layout, and typography in a Zola-based site. Content is fictional and for structural testing only.

A metrics system bills you for series, not for the requests you measure. That distinction is easy to forget right up until the invoice arrives. A single innocent-looking label — added to make a dashboard “a little more useful” — can turn one time series into hundreds of thousands. This piece walks through how the multiplication works, makes it concrete with a table, and shows the blow-up at a glance.

What cardinality actually counts

A time series is uniquely identified by a metric name plus the full set of its label values. Every distinct combination is its own series, stored and indexed independently. The total is not additive — it is the product of the distinct values each label can take:

series = metrics × ∏ (distinct values per label)

So one metric with three labels of 4, 10, and 50 values is not 4 + 10 + 50 = 64 series. It is 4 × 10 × 50 = 20,000. Add a fourth label with a thousand values and you are at twenty million. The cost is geometric, and it hides behind labels that each look reasonable in isolation.

Where the blow-up comes from

The table below tracks one http_requests_total metric as labels accrete. Each row adds a label to the row above it; the rightmost column is the running product.

The first four labels are bounded and well-behaved — 7,200 series is nothing. The fifth, user_id, is unbounded: it grows with your user base and never stops. That one column is the difference between a metric and an outage of your monitoring system.

Visualizing the blow-up

Plotted on a log-ish scale, the first four rows are invisible next to the last. The bars below are scaled to each row’s series count:

http_requests_total — series by label set (log10 scale)

  method              ▏                 5
  + status            ▎                 40
  + endpoint          █▏                1,200
  + region            █▌                7,200
  + user_id           ████████▊         360,000,000
                      └────────────────────────────
                       10⁰  10²  10⁴  10⁶  10⁸

Four of the five label sets share the leftmost sliver of the chart. The unbounded label alone spans the rest. This is the signature of a cardinality problem: not a steady climb, but a flat line that suddenly goes vertical the moment an unbounded dimension enters.

How to defuse it

The fixes are unglamorous and effective:

• Never label by an unbounded identifier. User IDs, request IDs, email addresses, and full URLs belong in logs or traces, not in metric labels.
• Bucket instead of enumerate. Replace user_id with user_tier (free, pro, enterprise) — three values, not fifty thousand.
• Normalize high-variance strings. Collapse /users/8412/posts/93 to a templated /users/{id}/posts/{id} before it becomes a label.
• Reach for exemplars and traces when you genuinely need per-request detail. They attach to a low-cardinality metric without multiplying it.

Cardinality is a budget, not an afterthought. Decide what each label is worth in series before you ship it, because the metrics system will happily charge you the product whether you did the multiplication or not.