Usage Model

NIE-SLA ships a fixed Cloudflare free-tier usage estimation model (usage-model) that estimates daily Workers, D1, R2 and Durable Objects consumption purely from the site's own data — no Cloudflare API access at runtime.

• Model version: usage-model-v1.3.4 (code: agent/scripts/usage-model.mjs)
• Calibration: agent/scripts/usage-model-calibration.json (nie-sla-usage-calibration-v2)
• Inputs: public /api/status, optionally the admin-only /api/debug/usage-summary aggregate

Ledgers vs free limits

Ledger	Metric	Free limit (day/month)	Output field
Workers Functions	invocations	100,000 / day	workers.function_calls
Durable Objects	requests	100,000 / day	do.requests
D1	rows read / written	5,000,000 / 100,000 per day	d1.rows_read / d1.rows_written
R2	Class A / B operations	1,000,000 / 10,000,000 per month	r2.class_a / r2.class_b

D1 "query count" is not the billing metric; rows read (including index scans) and rows written are. R2 "requests distribution" is dashboard-only and never adds to Class A/B.

Usage

node scripts/usage-model.mjs --hours 24 --json
node scripts/usage-model.mjs --from 2026-09-04T00:00:00Z --to 2026-09-04T08:35:00Z --json
node scripts/usage-model.mjs --hours 24 --range-mode stress

With an admin session (NIE_SLA_ADMIN_SESSION) or a read-only token (NIE_SLA_USAGE_TOKEN, generated in Admin → Settings → Security), the model anchors route counts against the debug aggregate and raises confidence from low to medium.

Method

1. Event decomposition per request path (task polling, update checks, telemetry, probe scheduling, public traffic);
2. Route anchoring from the debug aggregate when available (12% uncertainty);
3. Calibration multipliers fitted against a same-topology observed Cloudflare window (Workers calls ×0.73, D1 rows read ×1.28 / written ×0.45, DO requests ×2.05, refit per release);
4. Point estimate + operational envelope (optionally a stress ceiling) — never a billing guarantee.

Calibration status and capacity

• 2026-09-04 first reconciliation (00:00–08:35 UTC): Workers +10%, R2 B -11% inside the envelope; D1 writes underestimated by 36%, reads overestimated by 78% — superseded by the v1.3.4 refit below.
• 2026-09-17 v1.3.4 refit (offline-paired-fit, operational envelope): Workers calls ×0.73, R2 A ×1.43 / B ×2.06, D1 rows read ×1.28 / written ×0.45, DO requests ×2.05, including the D1 index write amplification (d1_index_write_multiplier=3); use --range-mode stress for ceiling planning.
• Plan A (live): probe cadence is carried by the R2 status state; the D1 targets schedule mirror is now a coarse per-target flush every 30 minutes (TARGET_SCHEDULE_FLUSH_SEC).
• 80-node write-side work (v1.1.62 / v1.1.64, live): probe buckets fall back to D1 only when the buffer is unavailable; the agent-state fallback writes only while buffered state is disabled; both per-minute schedulers share one target scan; every agent's live telemetry buffer now lives inside the shared WSS Durable Object, removing one extra DO call per agent report.
• Capacity: at 80 VPS + 5 TCP pings + 1 Cloudflare + 2 latency nodes the optimized path keeps D1 writes, D1 reads and DO requests well inside the free tier; exact ratios still require a full-UTC-day reconciliation — the model never replaces the billing statement.

Change discipline

Any change to report intervals, WSS/DO paths, D1 SQL profiles or public caching requires a model version bump and recalibration; pure UI/copy changes do not. The model is a reconciliation and capacity tool, not a Cloudflare billing replacement.

v1.3.3 debug-log scope: debug_logs keeps only login/account-security events and failed agent tasks; polling and admin operations are no longer logged, and every route's debugLog row writes drop to zero.

v1.3.4 output refit: Workers / DO / R2 / D1 output multipliers were refitted against observed same-topology windows, now including D1 index write amplification.