Sampling without losing edge cases
Uniform 1% sampling is easy to explain and often wrong for libraries. The failures that matter — weird locale paths, uncommon runtimes, first calls after a major bump — are exactly the ones random sampling erases.
Tiered keep rules
Keep 100% of error-class events (with aggressive PII scrubbing). Sample successful happy-path events more aggressively. Always keep the first productive use per install ID for a bounded window. This pattern shows up repeatedly in our Observatory Lab audits.
Privacy as a design constraint
Edge-case capture tempts teams to log arguments and file paths. Don’t. Hash where you must correlate; drop where you cannot justify. Korean and EU cohorts in our programs share a checklist we publish in course materials — consent, retention, and cross-border notes belong next to the schema, not in a forgotten wiki.
Prove the sampler
Once a quarter, compare a full-capture canary environment against production samples. If rare classes vanish, adjust tiers before you trust the adoption scorecard again.