WEB / WEB-21.03

Logs, metrics, and traces

Observe business outcomes, signal production, sampling, and observer visibility separately.

CORE · RED · BLUE · ARTIFACT · PRACTICE

Read every scenario without motion

Signal production

Public operation
The fixed public-note input is known; instrumentation, local exporter, and observer know no completion record yet.
Public business service
Public input known
Local instrumentation
Signals not recorded
Local export adapter
Local sink empty
Observer projection
Observer has no record
Business result
The own public-read adapter produces status200 and Tea title; telemetry delivery is a separate event from business completion.
Public business service
Tea result completed
Local instrumentation
Signals not recorded
Local export adapter
Local sink empty
Observer projection
Observer has no record
Metric increment
The selected operation/status counter increments from local input; a metric aggregate does not replace an individual log record.
Public business service
Tea result completed
Local instrumentation
Counter incremented
Local export adapter
Local sink empty
Observer projection
Observer has no record
Context relation
Supplied logical spans share traceAlias and an exact parentAlias relationship; they are not a measured network trace.
Public business service
Tea result completed
Local instrumentation
Root / child recorded
Local export adapter
Local sink empty
Observer projection
Observer has no record
Redact / project
Known fake authorization/session aliases are excluded by the output allowlist; the source dictionary is unchanged.
Public business service
Tea result completed
Local instrumentation
Allowlisted log ready
Local export adapter
Local sink empty
Observer projection
Observer has no record
Local stream write
The own structured record is written to a StringIO stream; no real collector or remote telemetry packet is used.
Public business service
Tea result completed
Local instrumentation
Allowlisted log ready
Local export adapter
Own stream written
Observer projection
Observer has no record
Independent observe
An independent JSON parser reads event/operation/correlation fields; seeded aliases are absent from output.
Public business service
Tea result completed
Local instrumentation
Allowlisted log ready
Local export adapter
Own stream written
Observer projection
Parsed public record

Unsampled trace

Sampling decision
The selected branch supplies sampled=false; this decision does not say that the business operation never occurs.
Public business service
Public input known
Local instrumentation
sampled=false supplied
Local export adapter
Local sink unchanged
Observer projection
No request trace
Read completes
The public-read adapter produces normal status200; trace selection does not alter this finite result.
Public business service
Tea result completed
Local instrumentation
sampled=false supplied
Local export adapter
Local sink unchanged
Observer projection
No request trace
Counter still counts
The own request counter includes the sampled=false operation; metric counting and trace inclusion are separate decisions.
Public business service
Tea result completed
Local instrumentation
Metric +1 recorded
Local export adapter
Local sink unchanged
Observer projection
No request trace
Trace omitted
Declared sampling leaves retained trace0; an omitted trace is not a synthetic failure event.
Public business service
Tea result completed
Local instrumentation
Trace deliberately absent
Local export adapter
Local sink unchanged
Observer projection
No request trace
Knowledge boundary
The observer does not equate no trace with no event; the verified sampling reason remains beside business completion.
Public business service
Tea result completed
Local instrumentation
Trace deliberately absent
Local export adapter
Local sink unchanged
Observer projection
No trace; event occurred

Exporter failure

Read input known
The export-failure branch starts with an own public read and declared best-effort diagnostics policy; it is not a mandatory audit.
Public business service
Public input known
Local instrumentation
Record not produced
Local export adapter
Failing sink supplied
Observer projection
Observer empty
Business completes
The public-note read produces a successful result; logging-sink failure is neither a business mutation nor a permission verdict.
Public business service
Business success200
Local instrumentation
Record not produced
Local export adapter
Failing sink supplied
Observer projection
Observer empty
Record projection
Only public allowlisted fields enter the local record; raw seeded aliases are not included in sensitive output.
Public business service
Business success200
Local instrumentation
Public record ready
Local export adapter
Failing sink supplied
Observer projection
Observer empty
Exporter failure
The own failing sink raises OSError; an actual local exception is observed without claiming a remote exporter outage.
Public business service
Business success200
Local instrumentation
Public record ready
Local export adapter
Own OSError observed
Observer projection
Observer empty
Visibility missing
The observer retains exported record0 and visibilityConfirmed=false; business success is not turned into a health guarantee.
Public business service
Business success200
Local instrumentation
Public record ready
Local export adapter
Own OSError observed
Observer projection
export0 / visibility gap
Scope / policy
The explicit public-read policy preserves the result without a general guarantee about compliance audit completeness.
Public business service
Success / diagnostic gap
Local instrumentation
Public record ready
Local export adapter
Own OSError observed
Observer projection
export0 / visibility gap

Missing signal

No observation
The missing case supplies no evidence about an application event, exporter result, or recent counter sample.
Public business service
Operation unobserved
Local instrumentation
No observation supplied
Local export adapter
Sink state unknown
Observer projection
Observer has no evidence
Window undefined
An empty measurement window leaves the error fraction undefined; absence of observation is not written as rate0.
Public business service
Operation unobserved
Local instrumentation
No observation supplied
Local export adapter
Sink state unknown
Observer projection
Denominator0 / undefined
Health unknown
Health unknown proves neither that the application is down nor healthy; it states only an evidence boundary.
Public business service
Operation unobserved
Local instrumentation
No observation supplied
Local export adapter
Sink state unknown
Observer projection
Health unknown
No event inference
The observer asserts neither no event nor no failure; a missing record cannot confer authorization or business success.
Public business service
Operation unobserved
Local instrumentation
No observation supplied
Local export adapter
Sink state unknown
Observer projection
No absence proof

ACTORS · STATE · DATA FLOW

Observability is a question beyond instrumentation counts

Observability uses external signals to understand internal system behavior and why a selected outcome occurred. Many log lines or an attractive dashboard are not enough. Choose the question first: did a public note read finish, which dependency waited, where did an error occur, and did the observation pipeline receive data? Bind instrumentation, signal semantics, and correlation to that question. Telemetry outcomes do not become authoritative business state.

A log records an event in structured or unstructured form with time, event type, and bounded context. Metrics aggregate selected measurements; request counts, error counts, and histograms are different kinds. A cumulative counter can increase while process resets or exporter temporality affect interpretation. A gauge describes a current observation and a histogram a distribution. Raw URLs or user identifiers as labels can create cardinality cost. This lesson uses finite operation/status dimensions without vendor tuning numbers.

A trace connects operations through spans and parent/context relationships. A public-read root can have a local-note-read child; shared context and the correct parent convey causality. A correlation identifier is not a credential or permission grant. trace-demo and span-root are inert teaching aliases, not a claimed trace wire format. The logical10–14 root duration4 and11–13 child duration2 are authored clock values, not measured latency or captured production traces.

An event need not appear in every signal

Instrumentation, sampling, local buffers, exporters, and collector/backend stages differ. A request can occur without a retained trace under selected sampling. Separate metric instrumentation can still count it. Head sampling decides early and does not guarantee selection of a later error; tail sampling can inspect more completed spans but has its own buffering and availability limits. The fixture uses a supplied sampled boolean, not an OpenTelemetry sampler or collector.

The finite window contains three local public calls: two successes and one unavailable result. Request total is3 and error total1, while only two selected requests retain trace records. Missing one trace does not make the metric total wrong. Error fraction1/3 uses that selected denominator; no global rate follows without window, reset, and label scope. An empty window with denominator0 has undefined rate, which differs from zero measured errors.

Producing a record, redacting it, submitting it to an exporter, and storing it for observation are separate facts. A normal own stream exports structured public data for independent JSON parsing. In the failure case a selected local sink raises OSError: business success remains, but observed exports are0 and visibility is unknown. This public-read policy is not a general compliance/audit permission; required audit events need their own retention and durability contracts.

Read every scenario without motion

Trace absence

Unsampled successful call
The unsampled call has status 200, businessCompleted true and metricIncrement 1, but sampled false and retainedTraces 0. Trace absence does not mean the request did not occur.
Diagnostic ledger
Business complete / metric+1 / trace0
Observation policy
Not yet evaluated
Observer outcome
Not yet evaluated
Window denominator
The 200/503/200 window has requestTotal 3 and errorTotal 1. Using retainedRequestTraces 2 as denominator changes scope; sampling does not remove calls from metrics.
Diagnostic ledger
Business complete / metric+1 / trace0
Observation policy
Error fraction1/3 / retained traces2
Observer outcome
Not yet evaluated
Empty input is not a rate
Missing denominator 0 has windowRate null. Zero observations do not yield an error fraction of 0 or health without signal-coverage evidence.
Diagnostic ledger
Business complete / metric+1 / trace0
Observation policy
Error fraction1/3 / retained traces2
Observer outcome
No denominator / rate null / health unknown
Read span relationships
span-child refers to span-root in trace-demo, with logical interval 11–13 inside root 10–14. Authored duration is not measured milliseconds or latency.
Diagnostic ledger
Logical parent relation / no wall-clock latency
Observation policy
Error fraction1/3 / retained traces2
Observer outcome
No denominator / rate null / health unknown
Bound causal claims
Equal operation names or nearby timestamps do not establish parent context. Trace alias and exact parent relation retain scope without diagnosing a production bottleneck from logical duration.
Diagnostic ledger
Logical parent relation / no wall-clock latency
Observation policy
Correlation relation ≠ full causal diagnosis
Observer outcome
No denominator / rate null / health unknown
Name absence correctly
Verified sampling explains known trace absence; missing observation lacks that explanation. No signal is not collapsed into no event or successful health.
Diagnostic ledger
Logical parent relation / no wall-clock latency
Observation policy
Correlation relation ≠ full causal diagnosis
Observer outcome
Sampling absence ≠ unexplained visibility gap

Export and privacy

Public log projection
Fake input includes demo-token-alias and demo-session-alias. The allowlist exports only event, operation and correlationAlias without mutating the source object.
Diagnostic ledger
Three exported fields / fake aliases excluded
Observation policy
Not yet evaluated
Observer outcome
Not yet evaluated
Narrow redaction scope
Known-field projection is not a secret detector for every nested format. Trace attributes and metric dimensions do not automatically pass through the same filter.
Diagnostic ledger
Three exported fields / fake aliases excluded
Observation policy
Known-field projection / other signals separate
Observer outcome
Not yet evaluated
Export-sink failure
exportFailure retains businessCompleted true and status 200 with exportOk false and exportedRecords 0. Successful public reading does not set visibilityConfirmed true.
Diagnostic ledger
Three exported fields / fake aliases excluded
Observation policy
Known-field projection / other signals separate
Observer outcome
Business200 / export false / visibility false
Policy qualification
This public diagnostic uses explicit best-effort policy. Mandatory security audit requires separate treatment of export failure; the example is not a universal fail-open rule.
Diagnostic ledger
Best-effort public diagnostics only
Observation policy
Known-field projection / other signals separate
Observer outcome
Business200 / export false / visibility false
Do not justify secret collection
Closing a correlation gap does not require Authorization, sessions or private bodies. Fake seed values are not credentials; minimal public comparison fields suffice.
Diagnostic ledger
Best-effort public diagnostics only
Observation policy
Minimum fields / no raw credential collection
Observer outcome
Business200 / export false / visibility false
Report two outcomes separately
The public result completed while diagnostic export failed. The record proves neither server authorization nor backend delivery; local sink failure narrows observer coverage.
Diagnostic ledger
Best-effort public diagnostics only
Observation policy
Minimum fields / no raw credential collection
Observer outcome
Business result known / diagnostic gap retained

ASSUMPTION · BOUNDARY · CONTROL

Reading no signal as no event or healthy

An absent trace does not prove that an operation never occurred. Sampling false, missing context, export failure, or retention windows can explain the gap. The condition examined is an observation gap, not an attack workflow. In the unsampled branch the business response completes and metrics increment while trace records remain0. The control separates outcomes from recording decisions without fabricating another success or failure event.

A metric0 does not prove no errors unless instrumentation and observation-window coverage are known. Producing error fraction0 from empty input hides the denominator. Health assessment combines expected signal production, recent observations, and user-flow results. The missing case retains unknown observer state; a dashboard or completed static scene confers neither real application health nor caller authority.

More collected data is not always better

Copying bodies, Authorization, session identifiers, or private attributes wholesale into logs can expose secrets and personal data. The fixture seeds only fake demo-token-alias and demo-session-alias values. Its output allowlist retains event, operation, and correlationAlias without mutating the source dictionary. This is not discovery of real secrets or a detector for every nested log format; independent reading checks only that known fixture fields do not reach the sink.

Span attributes and metric dimensions also collect data. A log filter does not automatically apply to other signals. High-cardinality raw identifiers affect memory, cost, and investigation scope. Operational telemetry and authoritative security audits can have different completeness and retention guarantees. Separating export failure from business results must not dismiss mandatory audit gaps; this example selects only a public read and an explicit best-effort diagnostic policy.

Read every scenario without motion

Signal controls

Three-call inventory
Adapter statuses are 200/503/200 with 3 requests, 2 successes and 1 error. Counter scope is this finite window, not a global service rate or SLA.
Public telemetry records
Window3 / success2 / error1
Signal comparison
Not yet evaluated
Coverage outcome
Not yet evaluated
Sampling comparison
Sampling true/true/false retains two traces while the third call increments metrics. Outcome and retention remain separate; trace count is not the request denominator.
Public telemetry records
Window3 / success2 / error1
Signal comparison
Calls3 / retained traces2
Coverage outcome
Not yet evaluated
Structured-field control
The exported record contains only three allowed fields. Fake authorization/session aliases are absent and sourceMutated false; no broader anonymization success is added.
Public telemetry records
Window3 / success2 / error1
Signal comparison
Calls3 / retained traces2
Coverage outcome
Allowed fields only / source unchanged
Span-context control
Root has null parent, child refers to span-root, and both use trace-demo. End-minus-start yields 4/2 logical units, kept separate from real timing claims.
Public telemetry records
Exact parent relation / logical durations4,2
Signal comparison
Calls3 / retained traces2
Coverage outcome
Allowed fields only / source unchanged
Three visibility states
Unsampled has a verified recording decision, exportFailure an observed sink OSError, and missing no observation. The same zero count carries different knowledge.
Public telemetry records
Exact parent relation / logical durations4,2
Signal comparison
Sampling / observed export error / unknown
Coverage outcome
Allowed fields only / source unchanged
Observer-health boundary
Finite log/count comparisons remain within local-adapter scope. They do not claim collector, remote exporter, SDK or production-log execution; visibility and business outcome remain separate.
Public telemetry records
Exact parent relation / logical durations4,2
Signal comparison
Sampling / observed export error / unknown
Coverage outcome
Local signal evidence / remote coverage unmeasured

Observation gap

Empty observer view
Missing has observationPresent false and health unknown. No observer record establishes neither source-event absence nor service health.
Public telemetry records
No observation / health unknown
Signal comparison
Not yet evaluated
Coverage outcome
Not yet evaluated
Missing coverage conditions
Sampling, export success and retained window are unknown. Copying normal export or unsampled explanations would invent a verified cause for this missing state.
Public telemetry records
No observation / health unknown
Signal comparison
Sampling/export/window missing
Coverage outcome
Not yet evaluated
Do not compute a rate
Denominator 0 and windowRate null are retained. Reporting zero errors adds an instrumentation-completeness assumption; businessEventAbsenceProved remains false.
Public telemetry records
No observation / health unknown
Signal comparison
Sampling/export/window missing
Coverage outcome
No fraction / event absence unproved
Required observation binding
Signal purpose, active instrumentation, sampling, export outcome and observation window are needed. A public correlation alias binds to the same action without secrets or private bodies.
Public telemetry records
Needed: instrumentation/sampling/export/window
Signal comparison
Sampling/export/window missing
Coverage outcome
No fraction / event absence unproved
Bound causal conclusions
Nearby clocks or equal titles do not prove event completion. Business results need their receiver and telemetry receipt its sink; neither fills the other gap.
Public telemetry records
Needed: instrumentation/sampling/export/window
Signal comparison
Business and telemetry recipients distinct
Coverage outcome
No fraction / event absence unproved
Retain unknown honestly
This review records a visibility gap without inventing events or health. Rates and outcomes can be reevaluated when new evidence identifies window and recipient boundaries.
Public telemetry records
Needed: instrumentation/sampling/export/window
Signal comparison
Business and telemetry recipients distinct
Coverage outcome
Visibility unknown / collect scoped signal evidence

IMPLEMENTATION · SECURITY CONTROL · VERIFICATION

Define what each signal counts and carries

Define log purpose, safe output fields, and correlation aliases rather than raw secrets or entire upstream objects. Record counter-reset scope and operation/status dimensions. Histogram or duration interpretation needs units, a window, and a source clock; authored logical ticks are not milliseconds or a benchmark. Show shared traceAlias and an exact parentAlias relationship for root/child spans. Equal names alone do not establish parent context.

The native check derives cumulative totals from actual public-call adapter decisions and compares them against separately parsed result lists. Sampled=false requests remain in metrics. An own StringIO sink produces structured logs, and an independent JSON parser checks selected fields. Fake seeds are absent from output while the source object remains equal. This is Python/local-writer evidence, not an SDK exporter/backend or a production-log capture.

Show visibility gaps beside business outcomes

Use separate branches for export errors, unsampled traces, and missing observations. A local successful public read can survive exporter failure while the signal store remains incomplete; record a diagnostic gap rather than health success. The unsampled branch has a declared explanation for trace0, whereas a missing observation may lack that verified explanation. Best-effort policy here concerns operational public reads; security-sensitive audit retention and access policy need separate design.

The selector detects the inconsistent combination exportOk=false and visibilityConfirmed=true. The degraded-visibility row yields count=0 and the inconsistent positive row count=1; missing rows remain insufficient evidence. A completed business response establishes neither instrumentation recording nor remote telemetry receipt. Observe request outcome, recording and export receipt separately for the same work.

NOW APPLY IT TO THE CASE

If the trace is absent, was the request absent too?

Three public calls have statuses 200, 503 and 200, while only the first two traces are retained. A public read can complete with 200 despite an export-sink error. Counters, sampled traces and exported records cover different scopes; correlation aliases are not wire trace IDs.

Evaluate the case data journey together with environment, dependency and telemetry context. Does an operational-success label establish every application outcome?

Open the related case file ↗

READ THE EXAMPLE TOGETHER

Public telemetry / visibility ledger

Select a field and examine what it supports and where the evidence ends.

Read the raw artifact and field explanations

{
  "schema": {
    "fields": [
      "event",
      "operation",
      "correlationAlias"
    ],
    "classification": "public synthetic diagnostic only",
    "recordIsVendorFormat": false,
    "traceAliasesAreWireIds": false
  },
  "normal": {
    "rawFakeInput": {
      "event": "public-read-completed",
      "operation": "read-public-note",
      "correlationAlias": "request-demo",
      "authorizationAlias": "demo-token-alias",
      "sessionAlias": "demo-session-alias"
    },
    "exported": {
      "event": "public-read-completed",
      "operation": "read-public-note",
      "correlationAlias": "request-demo"
    },
    "businessResult": {
      "status": 200,
      "title": "Tea note"
    },
    "exportedRecords": 1,
    "sourceMutated": false
  },
  "window": {
    "adapterStatuses": [
      200,
      503,
      200
    ],
    "sampled": [
      true,
      true,
      false
    ],
    "requestTotal": 3,
    "successTotal": 2,
    "errorTotal": 1,
    "retainedRequestTraces": 2,
    "errorFraction": {
      "numerator": 1,
      "denominator": 3
    },
    "scope": "finite own adapter window; not global service rate"
  },
  "spans": [
    {
      "traceAlias": "trace-demo",
      "spanAlias": "span-root",
      "parentAlias": null,
      "operation": "read-public-note",
      "logicalStart": 10,
      "logicalEnd": 14,
      "logicalDuration": 4
    },
    {
      "traceAlias": "trace-demo",
      "spanAlias": "span-child",
      "parentAlias": "span-root",
      "operation": "local-note-read",
      "logicalStart": 11,
      "logicalEnd": 13,
      "logicalDuration": 2
    }
  ],
  "unsampled": {
    "status": 200,
    "metricIncrement": 1,
    "sampled": false,
    "retainedTraces": 0,
    "businessCompleted": true
  },
  "exportFailure": {
    "status": 200,
    "businessCompleted": true,
    "exportOk": false,
    "exportedRecords": 0,
    "visibilityConfirmed": false,
    "observedFailure": "own sink OSError",
    "policy": "explicit best-effort public diagnostics; not mandatory audit"
  },
  "missing": {
    "observationPresent": false,
    "health": "unknown",
    "businessEventAbsenceProved": false,
    "windowRate": null,
    "denominator": 0
  },
  "limits": {
    "actualProductionLogsRead": false,
    "otelSdkExecuted": false,
    "collectorExecuted": false,
    "remoteExporterExecuted": false,
    "realSecretsRead": false,
    "realLatencyMeasured": false,
    "authorizationVerified": false
  },
  "selector": {
    "language": "XPath",
    "xml": "<audit><telemetry exportOk=\"false\" visibilityConfirmed=\"false\"/></audit>",
    "xpath": "count(/audit/telemetry[@exportOk='false' and @visibilityConfirmed='true'])",
    "expected": 0,
    "positiveFixtureXml": "<audit><telemetry exportOk=\"false\" visibilityConfirmed=\"true\"/></audit>",
    "positiveExpected": 1,
    "scope": "Own inert decision observations, not vendor log fields.",
    "missingFieldBehavior": "Missing observation is insufficient evidence; zero matches alone proves no healthy runtime/control."
  }
}

normal/schema: Own public output schema only; seeded fake aliases redacted.
window/spans: Finite counters and authored logical parent intervals have different scopes.
unsampled/exportFailure/missing: Recording decisions, sink errors, and no evidence carry different knowledge.
selector/limits: No production capture, SDK/backend, real secret, or latency claim.

Actual own local adapters, structured stream, and counts; spans are logical authored context fixtures. No real logs/secrets, OpenTelemetry SDK/collector/backend, remote exporter, or latency capture.

Two reality checks

No trace or error log means no request or error occurred.

Sampling, export, retention, or observation gaps can explain absence; no signal is not proof of no event.

A correlation ID and log success prove authority or business commit.

Correlation links context; permissions and authoritative business results are verified by their own owners.

YOUR TURN

Work through this story.

Predict the outcome, find the evidence boundary, and explain it in your own words.

A request completed with status 200, but sampled=false and retainedTraces=0. Does that prove it was omitted from metrics too?

Yes. Without a retained trace, the request never occurred.
No. metricIncrement=1 and businessCompleted=true.

Read the explained answer

No. metricIncrement=1 and businessCompleted=true. — No trace is retained for this request, but the metric increments. Trace count is not total request count.

The public-diagnostic export failed. Are returning the response and making the log visible the same success?

Selected lesson artifact fields; teaching example

exportFailure.status: 200
exportFailure.businessCompleted: true
exportFailure.exportOk: false
exportFailure.visibilityConfirmed: false
exportFailure.policy: explicit best-effort public diagnostics; not mandatory audit

No. The business operation completes; export visibility is not confirmed.
A 200 response proves every log sink received the record.

Read the explained answer

No. The business operation completes; export visibility is not confirmed. — Even with status=200 and businessCompleted=true, exportOk=false and visibilityConfirmed=false. This policy does not stop the response, but provides no log-delivery evidence.

The window has three requests, one error and two retained traces. Why is missing observation with denominator=0 not an error rate of zero?

Read the explained answer

For the supplied window, errorFraction is 1/3. Sampling leaves two retained traces while the metric denominator is three. The missing branch has observationPresent=false and windowRate=null; denominator=0 is not evidence of a zero error rate. Trace aliases provide logical correlation, not user identity or wire-ID verification. Fake session/authorization aliases outside the public allowlist are not exported, but this is not a full-anonymization guarantee.

□ I did not substitute trace count for the metric denominator.

□ I did not turn missing observation into a zero error rate or healthy-service verdict.

Open the application console · request / response / browser

Also open request, URL, cookie and DOM examples

Sources and scope

OpenTelemetry — Observability primerWhat is Observability? · Reliability and metrics · Understanding distributed tracing
OpenTelemetry — LogsLogs: timestamped record of an event · Structured logs · Log Record Exporter
OpenTelemetry — MetricsMetrics: runtime measurement · Metric Instruments · Aggregation · Cardinality limits
OpenTelemetry — TracesContext Propagation · Spans · Span Context
OpenTelemetry — SamplingHead Sampling · Tail Sampling
OWASP — Logging Cheat SheetEvent attributes · Data to exclude · Verification · Protection

YOUR NOTES / THIS BROWSER

Logs, metrics, and traces

1. Core · Mechanism

Signal production

Unsampled trace

Exporter failure

Missing signal

Observability is a question beyond instrumentation counts

An event need not appear in every signal

2. Red · Trust & Impact

Trace absence

Export and privacy

Reading no signal as no event or healthy

More collected data is not always better

3. Blue · Secure development & defense

Signal controls

Observation gap

Define what each signal counts and carries

Show visibility gaps beside business outcomes

If the trace is absent, was the request absent too?

Two reality checks

A request completed with status 200, but sampled=false and retainedTraces=0. Does that prove it was omitted from metrics too?

The public-diagnostic export failed. Are returning the response and making the log visible the same success?

The window has three requests, one error and two retained traces. Why is missing observation with denominator=0 not an error rate of zero?

Sources and scope