Gemeinsame Kosten verteilen

Shared Cost Allocation

Not every cost arrives neatly stamped with a single tenantId. Base database clusters, reserved GPU pools, shared vector indexes, CDN base fees, and platform background services must be allocated transparently so unit economics and tenant margin calculations reflect full cost.

Workflow

1. Keep direct costs and allocated shared costs strictly separate:
   • Never overwrite or mutate a tenant's direct attributed cost with allocated overhead.
   • Store allocated overhead in distinct ledger or rollup columns (directCostMicros, allocatedSharedCostMicros, totalFullyLoadedCostMicros) linked to an explicit allocationRuleId and allocationRunId.
2. Choose an allocation driver that matches resource causality:
   • Proportional to direct usage (telemetry-driven):
     • Allocate shared database compute in proportion to each tenant's query CPU time or row operations.
     • Allocate shared worker pool base cost in proportion to each tenant's cpuCoreMs or gpuMs. -$$\text{AllocatedCost}{\text{tenant}} = \text{SharedPoolCost} \times \frac{\text{DriverMetric}{\text{tenant}}}{\sum_{k} \text{DriverMetric}_{k}}$$
   • Fixed / tier-weighted split: Allocate base platform readiness costs across active workspaces by plan tier weight or equal split when usage is zero.
   • Hybrid (committed + burst / idle capacity):
     • Attribute utilized capacity directly by usage.
     • Track unused commitment / idle capacity (idlePoolCostMicros) explicitly in a platform capacity bucket (or allocate by reserved quota share) so an empty cluster does not make a single active tenant appear to cost $5,000 for one request.
3. Guarantee conservation and zero-division safety:
   • Handle zero total driver volume gracefully (when $\sum \text{DriverMetric}_k = 0$, route 100% of the shared pool to platform:idle_capacity rather than dividing by zero).
   • Reconcile rounding remainders deterministically (e.g., largest-remainder method) so $\sum \text{AllocatedCost}_{\text{tenant}} + \text{IdlePlatformCost} \equiv \text{SharedPoolCost}$ to the exact microdollar.

Anti-Patterns to Avoid

• Distorting per-unit marginal cost with idle cluster overhead: Charging a tenant's single API call the entire cost of an otherwise idle 8-GPU node hides true marginal cost and breaks pricing analysis.
• Opaque "magic multiplier" markups: Multiplying direct costs by 1.35x without recording the underlying shared pool total and driver formula prevents finance and engineering from auditing infrastructure efficiency.