An aggregate that grew a second reason to exist

A Subscription that had been correctly scoped in 2023 and had acquired a usage counter written thousands of times a day.

the consistency question, asked again:

  plan + entitlements   must be consistent within a
                        transaction. still true.
  usage counters        written by a metering job every
                        few seconds, read monthly.
                        never consistent with anything.

the symptom that forced the question: lock waits
between a plan change and the metering job, on the
same row, at about 40 a day.

An aggregate boundary is correct at a point in time and stops being correct when the write pattern changes, which nothing prompts anybody to re-examine. The lock waits were the prompt and they had been growing for eight months — the metric that would have surfaced it earlier is lock wait time per table, which nobody graphs.