> ## Documentation Index > Fetch the complete documentation index at: https://docs.orbit.devotel.io/llms.txt > Use this file to discover all available pages before exploring further. # Analytics pipeline: from raw events to an aggregate read surface > How Orbit computes the message, voice, and cost aggregates behind the Insights pages and the Analytics API — background roll-ups over per-tenant schemas, executed against the read replica, behind one scope and one rate limit. # Analytics pipeline: from raw events to an aggregate read surface Orbit aggregates every message, call, and spend event into the datasets that back the dashboard's Insights pages and the public Analytics API. This page covers how that path is built — what the read surface exposes, why every analytics query targets the read replica, the rate limits that apply, the background producers that pre-aggregate the series, and how the dashboard consumes the same numbers. Endpoint-by-endpoint request shapes live on the [Analytics API recipes page](/channels/analytics); this page is the model behind them. ## What the analytics surface exposes Every analytics dataset lives under `/api/v1/analytics` and returns counts over a caller-chosen window: * **Delivery metrics** — sent / delivered / failed / read totals for message channels, plus a time-series of the same counters. * **Voice metrics** — answered / completed / failed call totals over the same window. * **Cost** — billed message count, spend per currency, average cost per billed message, and SMS segment counts. * **Breakdowns** — the same counters split by channel, by destination country, and by error code for failed sends. Two controls shape every aggregate query: * `group_by` buckets the time-series at `hour`, `day`, `week`, or `month` resolution. Day buckets are the default. * Filters narrow the window: `channel`, `status`, `campaign_id`, and a `start_date` / `end_date` range (or a rolling `days` lookback). Both default to the last 30 days and are capped at 365. Access is gated on one scope: API-key callers must hold `analytics:read`; dashboard sessions pass through the route's session authentication. A raw API key that holds only send scopes is rejected with 403 before any query runs. Every response uses the standard Orbit envelope: the payload under `data`, `request_id` and `timestamp` under `meta`. ## The read-replica separation invariant All analytics queries execute against the **read replica**, never the primary database. That separation is deliberate: aggregate reporting over long windows scans or joins large portions of the per-tenant message tables, and it would become contention on the same database that serves the live send path — the queue that posts outbound sends, the webhooks that mark deliveries, the dialer that owns active calls. The invariant in one sentence: **reporting load never competes with live traffic for the primary database.** It holds for the API and the dashboard alike — Insights reads are routed the same way as external API reads. One consequence for readers: the replica lags the primary by a small replication delta, so a metric you read reflects the fleet a moment ago, not the nanosecond a send finished. For analytics workloads that delta is immaterial; a minute-level aggregate sees the same answer. ## Rate limits Analytics endpoints are rate-limited at **60 requests per minute per tenant**, with the limit keyed on the organization. The cap exists because aggregate queries are heavier than CRUD reads even against the replica — they bound the burst capacity of reporting workloads without touching the per-endpoint send quotas on the live channels. A response over the cap returns `429` with a `Retry-After` header. Honor it — the limiter is per tenant, so your own dashboards and scheduled exports see the same meter in the account. ## How the roll-ups pre-aggregate the series The hour/day/week/month buckets the API returns are not always computed on-demand. A pair of background producers in the webhook worker pre-aggregate message counters into per-hour buckets so the dashboards and the most common queries read a pre-built, indexed row instead of scanning raw events at request time: * **`scheduler-stats-rollup.ts`** — aggregates the previous full hour of the tenant's message table into a keyed `(hour_utc, channel, status)` row, with the count and the spend. The upsert is idempotent (`INSERT ... ON CONFLICT DO UPDATE`), so a late delivery webhook that backfills a closed hour re-aggregates cleanly instead of duplicating. Tenants are visited in small parallel batches and the pass returns early on shutdown; a skipped tenant re-rolls the same hour on the next tick. * **`scheduler-analytics.ts`** — a broader producer that fans the same kind of per-tenant aggregation across additional analytics series: channel-engagement combined with sentiment and queue-quality roll-ups, so Insights panels that mix delivery with engagement read one aggregate instead of recomputing joins on raw rows. Both producers run per tenant on the tenant's own schema — the same isolation the [tenant isolation concept](/concepts/tenant-isolation) defines — with transient-error classification so a pgBouncer blip skips a tenant batch rather than failing the whole tick. The invariant a reader should take: **the analytics surface serves pre-aggregated per-hour rows when they exist, and falls back to a live scan when a series predates the roll-up producers** — either way, the read always goes to the replica. The producers make the dashboard fast and the API cheap to call at the hour bucket width that most reporting actually asks for. ## How the dashboard consumes these datasets The Insights pages in the dashboard read the same `/api/v1/analytics` datasets the public API serves — the charts showing outbound volume by channel, delivery-rate trend lines, per-country breakdowns, and spend per currency are rendered straight over those endpoints. The scheduled-reports surface a tenant configures in the dashboard composes the same aggregates into a recurring e-mailed export. That shared read surface is the point of the pipeline: the same pre-aggregated rows power the dashboard view a user opens, the BI pull a customer schedules, and the canned scheduled report — so a figure a stakeholder sees on a dashboard is the figure an API script can fetch, counted the same way.