Notes · 8 February 2026
SLA language ops can defend
Vague uptime promises collapse under the first late extract. Write levels your instrumentation can actually prove.
Product managers ask for ‘reliable overnight data.’ Ops hears that as a wish. An SLA that can be defended names the dataset, the clock, the acceptable lag, and the signal that proves the promise held.
Instrumentation passes fail when the written SLA talks about platform availability while consumers care about row freshness. Align the words with the probes. If you cannot measure it by 09:15, do not promise it by 09:00.
Hong Kong teams we work with often settle on a small set of data-product SLAs rather than a single estate-wide percentage. Fewer promises, each with an owner and an alert, beat a polished number nobody can explain in an incident review.