How fresh does our data actually need to be?
Work backwards from the decision. If a human acts on the number once a day, an hourly sync is already faster than the decision cycle. Real-time earns its complexity only when the action is time-bound — a missed call that needs a callback in minutes, a job that needs dispatching now. Everything else usually wants reliable and reconciled over fast, because near-real-time pipelines are considerably harder to make correct.
The question that settles it
Ask what you would do differently if you knew this ten minutes sooner. If the honest answer is nothing, the data does not need to be real-time and you have just saved yourself a substantial amount of engineering and ongoing fragility.
The answer is rarely uniform across a business. A missed-call alert is worth minutes because the customer is calling the next company right now. A weekly revenue-by-campaign report is not improved at all by being computed continuously. Both requirements can be true in the same company, and they should be served by different pipelines.
Three tiers most businesses actually need
- Event-driven, minutes. Missed calls, new leads, emergency jobs, appointment cancellations. Anything where a person must respond while the customer is still deciding.
- Hourly. Dispatch load, board health, pacing against the day's targets. Fast enough to change today, cheap enough to be reliable.
- Nightly and reconciled. Revenue, attribution, marketing performance, anything compared across periods. This tier should be correct rather than current, because it is the tier people argue about.
What real-time actually costs you
The cost is not compute. It is correctness. A batch pipeline has a window in which to reconcile, deduplicate, wait for late-arriving records and validate before publishing. A streaming pipeline publishes on partial information by definition, so it must handle corrections after the fact, and correction logic is where the bugs live.
Real-time alerting also fires on incomplete data. An alert that a location's revenue collapsed, sent at 10am because invoices post in the afternoon, teaches people to ignore alerts. That is a worse outcome than no alert.
The hybrid, and the rule that keeps it honest
Most operators land in the same place: event-driven for actions, batch for reporting. That works, with one requirement — the two must not contradict each other. If the alert says eleven missed calls and the morning report says eight, both lose credibility regardless of which is right.
The fix is that both read the same definitions and the same reconciled store, with the fast path clearly labeled as provisional. We build intelligence briefs on the reconciled tier and real-time notifications on the event tier, and we make the relationship between them explicit rather than leaving people to discover the discrepancy themselves. Choosing the right tier per use case is one of the most consequential decisions in an integration architecture.
Topics: latency · real-time · sync design · architecture
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.