You can't control a bill you can't attribute. Cost controls give every team a budget, a live meter, and a cap before they blow past it.

The most common support ticket we got last year wasn't a bug report. It was some version of "our bill doubled and nobody knows why." Usually the answer was a new service shipped with debug-level logging left on, or a label added to a metric that multiplied its cardinality by forty. The data existed to explain it. It just wasn't surfaced anywhere a human would look before the invoice arrived.

Attribution first

Cost controls starts with per-team, per-signal ingest attribution, using the same team tag most of our customers already set for ownership. No new instrumentation required for teams already tagging resources.

from ingest_usage
| where team == "payments"
| summarize bytes=sum(size_bytes) by signal, bin(1d)
| order by bytes desc

Budgets and soft caps

A team sets a monthly budget; Helix alerts at 70% and 90% of it automatically, routed through the same alert routing that pages you for an incident. Past 100%, teams can choose a soft cap (alert only) or a hard cap (start sampling down non-error traffic automatically until next month) — we default every new budget to soft, because silently dropping data is a decision a team should make on purpose, not one we make for them.

What surprised us

We expected the hard cap to be the popular option. It's used by fewer than 15% of teams who set budgets. Most teams just wanted visibility — once cost was attributed to a team and visible before the invoice, the runaway-cardinality problems mostly fixed themselves within a sprint, without anyone needing an automatic cutoff.

Rolling this out to finance, not just engineering

A quieter but important part of the launch was a monthly export that maps ingest cost back to the same team and cost-center codes finance already uses for headcount, so a platform team's observability line item stops being an unattributed lump on a cloud bill. One customer's finance team told us it was the first time they could actually explain a quarter-over-quarter spend change to their own leadership without asking engineering to reconstruct it after the fact.

  • Per-team, per-signal ingest attribution with no new instrumentation
  • Budget alerts at 70% and 90%, routed through existing alert routing
  • Optional hard caps that auto-sample down non-error traffic
  • Daily cost breakdown queryable in HelixQL like any other signal