optimize(querying) ∧ ¬optimize(writing) </philosophy>
<current-practice> implementation := OTel spans + annotations wide-events := mental-model for annotation strategyannotate(span, context) where context = { business ∪ user ∪ technical ∪ outcome } </current-practice>
<dimensionality> wide-event.fields := { identity: {traceId, spanId, service, operation} user: {userId, accountTier, accountAge, lifetimeValue} business: {featureFlags, experimentGroup, cartValue} performance: {durationMs, dbQueryCount, cacheHitRate, retryCount} outcome: {success, errorCode, httpStatus} }high-dimensionality → better-queryability high-cardinality(userId) → acceptable </dimensionality>
<anti-patterns> scattered-logs := console.log("step1") >> console.log("step2") >> ... low-dimensionality := span.set("success", true) ∧ |fields| < 5 technical-only := {http.status, db.queries} ∧ ¬{user, business} </anti-patterns> <correct-pattern> span.setAttributes({ "request.operation", "user.id", "user.tier", "cart.items", "cart.value", "feature.*", "db.query_count", "cache.hit_rate", "request.success" })∀ span → attach(identity ∪ user ∪ business ∪ performance ∪ outcome) </correct-pattern>
<tail-sampling> retain(100%) := errors ∨ slow(>p99) ∨ vip retain(1-5%) := success ∧ fast </tail-sampling> <queryability-test> before(instrument) → verify(answerable({ "failures where tier=premium ∧ feature.new_flow=true" "p99(latency) group by tier" "errors group by featureFlags" "full context for user X incident" }))¬queryable → ¬enough-context </queryability-test>
<terminology> cardinality := |unique values| (userId=high, httpMethod=low) dimensionality := |fields per event| (more → better) wide-event := canonical-log-line := one comprehensive record </terminology> <when-to-apply> deciding(span-annotations) reviewing(instrumentation-coverage) debugging(incidents) → "what context was missing?" planning(new-service-observability) choosing(fields-to-index) </when-to-apply> </wide-events>Reference: See Article.md for full article by Boris Tane (loggingsucks.com)