wide-events

traditional := many(log-lines) → grep(services) → hope wide := one(event) → query(structured) → answer

optimize(querying) ∧ ¬optimize(writing)

implementation := OTel spans + annotations wide-events := mental-model for annotation strategy

annotate(span, context) where context = {
business ∪ user ∪ technical ∪ outcome
}

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

scattered-logs := console.log("step1") >> console.log("step2") >> ... low-dimensionality := span.set("success", true) ∧ |fields| < 5 technical-only := {http.status, db.queries} ∧ ¬{user, business} 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)

retain(100%) := errors ∨ slow(>p99) ∨ vip retain(1-5%) := success ∧ fast 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

cardinality := |unique values| (userId=high, httpMethod=low) dimensionality := |fields per event| (more → better) wide-event := canonical-log-line := one comprehensive record deciding(span-annotations) reviewing(instrumentation-coverage) debugging(incidents) → "what context was missing?" planning(new-service-observability) choosing(fields-to-index)

Reference: See Article.md for full article by Boris Tane (loggingsucks.com)