AP-008 — Real time where it matters

Real-time capability is a genuine differentiator. It is not a universal requirement.

Every decision to build real-time infrastructure must be justified by a specific use case and evaluated against its cost and complexity. Real-time pipelines are significantly more complex than batch. That complexity must earn its place.

Required — no debate: - Payment authorisation — balance and fraud score must be current at the moment of authorisation - Sanctions screening — must reflect the current list, not yesterday's - KYC gate on account activation — must reflect current verification status - Capital and liquidity breach alerts — must fire before the breach, not after the report

Valuable but not required: - Transaction categorisation — seconds, not milliseconds, is fine - Customer insight cards — 60-second refresh is sufficient - Credit pre-approval — nightly batch is correct

Not appropriate: - Analytical reporting — Snowflake batch is the right tool - Model retraining — weekly cadence is sufficient - Audit log analysis — batch is correct

The staleness test: What is the cost of this data being 60 seconds stale? 5 minutes? If the honest answer is "it does not matter much," batch is the right choice.

KISS check: This principle has the tightest relationship with KISS of all nine. Real-time is where well-intentioned engineers add the most unnecessary complexity. Default to batch. Evolve to real-time when the use case demands it.

See also: Data architecture for the write-back pattern and latency targets.

Relationship to other principles

Principle	Relationship
AP-001 KISS	Real-time is the most common source of avoidable complexity — KISS is the strongest check on this principle.
AP-005 Customer driven	Payment authorisation latency is a direct customer experience metric — real-time earns its place here.
AP-006 Cost effective	Real-time infrastructure carries a significant cost premium that must be justified by a specific use case.
AP-009 Robust and serviceable	Real-time pipelines have more failure modes than batch — operational simplicity favours batch where latency permits.

See the full architectural principles index.