Tally

Appearance

Goals and design principles ​

Tally does reporting, metering and rating for cloud platforms. It collects events, metrics and inventory data from a cloud, records the usage it finds in neutral units, rates that usage with a versioned pricing model and exports the result as statements. A billing system, a cost dashboard, a capacity plan or a chargeback workflow reads those statements. Every platform reaches the system through a provider pattern: the platform contributes a thin integration layer, and the core behind it stays shared and platform-independent. OpenStack is the one provider that exists, and the pattern is built for more (the providers that do not exist yet).

Goals ​

  • Collect runtime data from cloud resources centrally, whatever platform the resources live on.
  • Record the lifecycle events of cloud resources without gaps.
  • Export inventory data from platform-specific sources, a database or an API.
  • Meter usage in neutral, platform-independent units: minutes, counts and sizes.
  • Rate metered usage by applying a configurable pricing model.
  • Stay extensible for further cloud platforms and services.

Design principles ​

  • Platform-agnostic data model: one schema for metrics and one for events, the same across every platform and service (architecture and the provider pattern).
  • Provider pattern: each platform registers with its resource types, its exporters and its event collectors (architecture and the provider pattern).
  • VictoriaMetrics as metrics store: every metric goes to VictoriaMetrics, which answers PromQL and MetricsQL and takes remote write (architecture and the provider pattern).
  • Project as a first-class entity: a project is registered with the platform it belongs to, and a cross-platform dependency is a directed relation carrying metadata (project registry, relations and exclusive attribution).
  • Dual ingestion with reconciliation: events carry the real-time picture, and a periodic sync against the platform API catches what the events missed (dual ingestion and reconciliation).
  • Events as the single source of truth: the append-only event history is authoritative and current_resources is a derived projection a replay rebuilds at any time (events as the source of truth).
  • Billing-grade ingestion: delivery is at-least-once, ingestion is idempotent and deduplicates on event_id, and a late or out-of-order event is folded in by projection replay (events as the source of truth).
  • Cloud instance dimension: platform names the platform type and cloud names the concrete installation, so two OpenStack clouds are two first-class clouds (architecture and the provider pattern).
  • Metering separated from rating: usage is recorded in neutral units first, and pricing is applied as a separate step (metering separated from rating).
  • Reproducible billing: pricing models are versioned, project relations are temporally valid, a finalized billing period is immutable and a correction is a delta against it, so re-processing a past period always yields the same result (billing period lifecycle and corrections).
  • Decimal money arithmetic: monetary values are computed in decimal types under defined rounding rules, never in binary floats (money and rounding).