You Probably Don’t Need Real-Time Analytics

"We need it real-time" appears in almost every analytics requirements meeting, near the top, unexamined. It is worth examining, because the delta between a real-time pipeline and an hourly batch is roughly 5–10x in build cost, a permanent tax in operational fragility — and, for most commerce decisions, zero difference in outcome.

Match latency to the decision, not the dashboard

The test is one question: what would you do differently in the next hour if you knew this number now? Budget reallocations are weekly decisions. SKU pricing reviews are daily. Cohort analysis is monthly. The executive glancing at revenue at 15:00 acts identically whether the number is live or 40 minutes stale — declare the freshness and nobody suffers. Building streaming infrastructure so a dashboard tile updates between meetings is engineering as interior design.

The hidden costs compound: streaming pipelines fail in exciting ways (ordering, late events, backfills that batch handles trivially), on-call load rises, and every schema change becomes a production migration. Hourly dbt runs fail boringly and rerun by button. Boring is a feature you pay to keep.

Where real-time genuinely earns it

Honest exceptions, all sharing one trait — a machine acts on the data within seconds:

• Inventory sync across channels. The oversell window is measured in seconds; this is the canonical commerce streaming case — though note it is an operational pipeline, not analytics.
• Fraud and abuse checks in checkout flows — the decision happens mid-transaction or not at all.
• Stock-aware advertising: pausing campaigns the moment a hero SKU hits threshold is free money and needs minutes, not hours.
• Peak-day war rooms — Black Friday genuinely benefits from live order/error/conversion telemetry, because interventions (kill a broken discount, reroute fulfilment) happen within the hour. Build the war-room view as an event mode, not as the everyday architecture.
• Operational alerting on money-losing anomalies (site errors, payment failure spikes, feed disapprovals) — alerts want low latency even when dashboards do not.

Note what the list excludes: every revenue dashboard, every marketing report, every cohort view, essentially all "analytics" as normally meant.

The architecture that keeps everyone happy

Split the two concerns cleanly. Operational events (orders, stock, errors) flow through the event bus that your integration layer already needs — machines subscribe where seconds matter. Analytics runs batch on the right-sized stack: hourly for trading-critical models, daily for the rest. The dashboards read the warehouse; the robots read the bus; each side is simple because it is not pretending to be the other.

When someone says "real-time", hand them the one-question test. Nine times out of ten the honest answer is "hourly is fine" — and the tenth time, now you know exactly what to build and why it is worth its cost.