18 Jun 2026 · Methods
Sampling is not suppression
When traffic spikes, analytics teams reach for whichever lever is closest. Mixing measurement choices with governance acts is how silent bias enters the warehouse.
Sampling means you intentionally keep a fraction of events to control cost or noise, and you still know how to reweight. Suppression means you drop a class of events or properties because they should not exist downstream — PII, prohibited categories, a consent flag that is off. One is a statistician’s dial. The other is a steward’s decision.
We see them collapsed into a single SDK toggle labelled “reduce volume.” That toggle then gets copied into the warehouse as if it were a random sample. Funnels look smoother. Rare but sensitive events disappear. Nobody wrote down which rule fired.
Write the rule in the contract
On the Analytics Data Contract we insist that sampling and suppression occupy different lines. Sampling: rate, seed or hash method, whether identity is preserved, how dashboards reweight. Suppression: the condition (consent off, property matches a pattern, event family retired), the owner who may change the condition, and whether dropped rows are deleted or held in a restricted store.
If you cannot explain the difference to a new analyst on Monday, you do not have a method. You have a cost-saving accident.
Bias you will not see in Looker
Sampling that keys on user id can quietly undercount power users. Suppression that keys on a language or region field can quietly remove a PDPA-relevant population from QA dashboards while leaving them in a marketing extract. The desk review for this is boring on purpose: we read the condition aloud and ask who is missing.
The week-six lecture in the flagship path is shorter than students want. That is a known thin spot — several reviews mention it. The remedy is a marked example on your own stream, not more slides. Bring it to the Desk if the cohort calendar is full.