Sampling deliberately creates gaps in observability’s evidentiary history.
Head sampling creates the gap before the outcome exists. Its decision relies on information available at the beginning of an execution, while later failure, latency, retries, downstream effects, and violated commitments remain unknown.
Tail sampling creates the gap after the system has captured, enriched, serialized, transported, buffered, assembled, and evaluated the trace. A retention policy selects the histories worth preserving and erases the remainder.
Each discarded trace marks an execution absent from the durable trace-level record.
The Problem of Synthetic Silence
Sampling makes absence ambiguous.
A missing trace may indicate that no execution occurred, instrumentation failed, context propagation broke, transport lost the record, or a sampling policy deleted it. The surviving trace set cannot explain missing executions.
The architecture manufactures silence.
An evidence-based approach draws meaning from presence, contradiction, expectation, and silence. Silence becomes evidence when an actor was expected to speak, the deadline passed, and the reporting path stayed credible.
Sampling injects synthetic silence into the same channel.
Intentional deletion becomes indistinguishable from operational inactivity. The absence of data loses its evidentiary significance, resulting in disrupted continuity across executions.
The Cost of Enrichment
A trace begins as execution and ends as inventory.
Each span accumulates attributes, resource metadata, baggage, events, exception details, service identities, deployment information, correlation identifiers, and domain-specific context. Every field increases descriptive resolution, payload size, cardinality, processing cost, and variation.
Enrichment produces denser records.
Interpretation produces state.
The tracing pipeline concentrates on detailed historical representations for storage, indexing, correlation, reconstruction, and later analysis. Record richness and retention cost rise together. Volume eventually crosses the economic boundary, and sampling restores affordability by puncturing the history.
Those gaps carry epistemic consequences.
Evidence At Subject Scale
Statistical sampling estimates population characteristics, including throughput, latency shapes, and failure rates.
Subject-level evidence supports a different class of judgment:
- whether this payment remains authorized
- whether this dependency continues to honor its commitment,
- whether this deployment is degrading,
- or whether an expected witness has fallen silent.
A population estimate might report that 0.7 percent of requests failed.
A situated account must establish what happened to this subject, which commitments were affected, whose testimony supports that conclusion, what remains uncertain, and how the situation changed.
Evidentiary continuity requires each relevant observation to update the system before its raw form expires.
Semantic Compression
Repeated successful executions can sustain a sign of continuity. Retries can produce a sign of instability. Diverging latency can weaken confidence in a commitment. A failed operation can change the status of a subject. Contradictory outcomes can escalate a situation.
Thousands of events can be absorbed into a few meaningful state transitions: event → sign → status → situation
- Event: A recorded occurrence.
- Sign: An interpretation of what that occurrence indicates.
- Status: The current condition attributed to a subject.
- Situation: A reconciled operational account of what matters now.
Each stage reduces volume by increasing interpretation. Raw records can expire after their evidentiary effect enters the system’s durable state. The representation disappears while the knowledge survives.
Semantic compression preserves knowledge while reducing representation.
The originating component is the first authoritative location for transformation. It knows the operation being attempted, the local commitment involved, the outcome it observed, and the deviation from its expectations. A later reconciler can combine that testimony with evidence from other components to establish status and situation.
Sampling leaves excluded observations unreconciled at the level of subjects, commitments, and situations. Their representations disappear together with any effect they might have had on status, confidence, or situation.
The resulting gap becomes permanent.
The Receipt
Sampling exposes two structural failings within a distributed tracing system:
- Record production exceeds the economics of retention.
- Responsibility for semantic reconciliation remains unassigned.
The pipeline enriches records, transports them, selects a minority, and erases the remainder.
A sign-based observability pipeline communicates at the rate of semantic change.
- Stable conditions persist as state.
- Significant changes propagate because their meaning had changed.
- Communication volume follows the rate of change rather than the rate of execution.
The tracing pipeline captured the event without assigning learning responsibility; sampling confirms no meaning.