Usage Metering & Billing Pipeline  ·  View 18 of 22  ·  Operations

Observability

Five signal types across five pipeline stages, and the row that makes this a billing platform.

Editable source SVG draw.io All views
Ingest Meter Rate Invoice Evidence Correctness Accept vs ack gap Dedup hit rate Re-rate difference Total vs sum of lines Replay determinism Latency p99 accept, 50 ms Freshness p95, 5 min Cap decision p99, 10 ms Run complete, 6 h Evidence query p95 Volume Events/s per producer Periods closed Lines priced Invoices issued Exports served Money Accepted, not yet billed Provisional vs closed Δ Revenue per meter Credit memo value Disputed value Cost to meter Cost per M events Cost per meter Rating compute Run cost Cold retrieval Observability — Signal by Pipeline Stage The Money row is what makes this a billing platform rather than a pipeline: accepted-but-unbilled and provisional-versus-closed are the two numbers that catch lost revenue before a customer does. v 1.0 · owner Architecture Practice · date 2026-10

Decisions

  • The Money row is what distinguishes this from an event pipeline. Accepted-but-unbilled and provisional-versus-closed delta are the two numbers that catch lost revenue before a customer does, and neither is a conventional service metric.
  • Correctness is a first-class signal row: dedup hit rate, re-rate difference, invoice total versus the sum of its lines, and replay determinism. Each one tests an assumption the architecture rests on.
  • Cost is measured per meter, so a high-cardinality dimension that costs more to meter than it earns can be found and removed rather than argued about.

Targets

  • Accuracy: at most 1 in 10^6 accepted events lost or double-counted after reconciliation (stated assumption).
  • Pipeline cost below 0.5% of metered revenue, reported monthly (stated assumption).
  • Freshness p95 5 minutes; cap decision p99 10 ms; run complete in 6 hours (stated assumptions).

Risks

  • Replay determinism cannot be sampled cheaply; a scheduled replay of a known account is the only honest test, and it costs compute every day it runs.
  • Four of these twenty-five signals require a comparison against an independent source. Without the Phase 3 reconciliation against the serving path's own counters, the platform is partly checking its own homework.