Logs are versioned event records with a data boundary

Choose a stable event class such as checkpoint.publish_failed and keep variable details in typed fields rather than generating a different event name for every path. A duration field should have one documented unit, an outcome should use a bounded vocabulary where appropriate, and identifiers should have defined scope. Preserve a human-readable body for context without forcing every query to parse its punctuation. Resource identity and instrumentation scope help distinguish which component emitted the event from the operation or tenant the event describes.

OpenTelemetry's log model separates fields including body, severity, attributes, event name, and optional trace/span context. A severity label is a classification, not an error-rate denominator or a guarantee that an event was delivered once. A retry can produce several similar records for one user operation. If deduplication is required, design an event identity and retention scope that distinguish retransmission from a genuinely repeated occurrence. Do not use a timestamp or a whole rendered message as a reliable universal uniqueness key.

Event time describes when the source says the event happened; observation time describes when a collector observed it. OpenTelemetry carries these separately because buffered or externally collected events can arrive later. Their difference can include clock offset and conversion errors as well as transport delay, so it is not automatically a precise one-way network latency. Use appropriate timing sources for local elapsed durations, retain timezone/epoch conventions for absolute timestamps, and test units explicitly. A millisecond/nanosecond mix-up can look like an extraordinary incident while only being a schema defect.

Version a schema when changing a field's meaning, type, or unit would break existing consumers. Prefer additive compatible changes when possible, document missing versus null versus empty values, and test old/new producers with queries and parsers. An event's schema version should not become an uncontrolled metrics label merely because it is useful in logs. Keep migration fixtures for historical records and handle malformed events observably rather than dropping them silently. A successful parse is weaker than correctly interpreting the intended event; consumers need both syntactic and semantic compatibility.

Define which fields may be collected, transformed, retained, and accessed. Raw authorization headers, credentials, unrestricted request bodies, or sensitive model inputs should not flow into an event merely because debugging would be convenient. Prefer explicit allowed fields and early redaction or omission at the intended boundary, then test nested attributes and exception paths. Hashing an identifier does not automatically make it anonymous or reduce its cardinality. Once a sensitive value reaches several exporters, indexes, and backups, removing only the source record does not undo that propagation.

The queue lab illustrates buffering of equal-sized log batches, not a schema validator. Reset, set both arrival λ=4 and service μ=2, predict backlog growth from four batches, then Play or Step. Reset and re-enter both λ=1 and μ=3 for a separate drainage case. Slider edits preserve elapsed time and recalculate from the initial four rather than continuing an accumulated queue at new rates. The instrument cannot parse JSON, redact data, enforce access, or guarantee delivery. It simply shows why buffered logs can be observed later than their source event.