Analytics and warehouse
Connect operational work to events, business facts and metrics with explicit meaning.
Status: Design. Analytics intake, maintained metrics and disk-scale warehouse execution are planned. RURAL has no demonstrated event-rate or warehouse-capacity claim.
Bring the sky into operational work
RURAL can store website analytics, product activity and business events beside the projects and deployments that give them context. That makes useful questions possible: what changed after a release, where signups came from, or how recorded revenue changed after refunds.
A source has an explicit project binding and provenance record. A browser event, an authenticated purchase event and a provider's interval report are different kinds of evidence.
Separate events, reports and metrics
| Representation | Required meaning |
|---|---|
| Event | Stable source-scoped identity; retries should not create another fact. |
| Interval report | A replaceable observation for a named interval and definition. |
| Correction or reversal | A documented change to an earlier fact or business outcome. |
| Metric | A versioned definition with units, population and aggregation rules. |
| Result | A computed answer with source coverage, time range and freshness. |
Replacing a report of 100 with 103 must produce 103, not 203. An arriving refund is not the same operation as deleting a duplicate event. Late arrivals and provider corrections need explicit treatment.
Money carries currency and a checked numeric representation. Conversion rates require compatible populations. Unique visitors and overlapping reports cannot simply be added as ordinary counts.
Accept with a meaningful receipt
The intended intake path validates identity and quotas, commits accepted observations with their replay/progress state, and only then acknowledges the advertised durability level.
A single host can promise only a declared local failure envelope. Protection against destruction of that host requires replicated durable acceptance or another explicitly stronger mechanism. A receipt from a RAM queue, a browser's send queue or a later backup job is not equivalent.
Finite capacity requires backpressure. Producers must distinguish pending, accepted, duplicate and rejected work. A viral burst should produce visible backlog or bounded rejection rather than false success and hidden loss.
Make analysis efficient and correct
The warehouse direction uses immutable analytical segments, bounded scans and maintained projections. A projection through one position must be combined with later corrections/deletions according to entity identity and version; a naive union can double-count data.
Interactive CRUD, intake, analysis and maintenance need separate budgets so one workload cannot indefinitely starve the others. Larger-than-memory execution, compaction and retention are later acceptance gates.
The Oracle should receive the small computed answer and its evidence coordinates, rather than every event that contributed to it.