Metrics disappear after a pipeline change

Understand the likely causes

• Possible causes include no source observations, changed resource or metric names, temporality or unit migration, an inactive pipeline component, filter removal, export failure, delayed buffering, or a query that still selects the old identity. These possibilities need different repairs and cannot be distinguished by panel color alone.

Freeze the contract and change boundary

Record the last known good and first known bad intervals, source and collector revisions, metric contract, backend, dashboard query, and environment scope. Keep credentials out of diagnostic notes and configuration excerpts. Compare old and new identity and units before assuming that data stopped. Use approved read-only evidence in a real environment; this editorial guide runs only a synthetic stage classifier. Do not restart or reconfigure a live collector simply to see whether the panel returns.

Locate the first missing stage

Trace the expected signal through source emission, receiver acceptance, processing, export, and query selection. The offline fixture distinguishes an export gap from a query-selection gap using known synthetic booleans. It cannot inspect an SDK or collector. In a real pipeline, align stage counters and freshness with queue changes, retries, and resets; counts need not balance instantaneously. An accepted transport batch is not proof of backend persistence, and a successful export is not proof that the current query selects the resulting identity.

def stage(emitted, accepted, exported, selected):
    if not emitted:
        return "source-gap"
    if not accepted:
        return "transport-or-receiver-gap"
    if not exported:
        return "processing-or-export-gap"
    if not selected:
        return "identity-or-query-selection-gap"
    return "inspect-rendering-and-freshness"
print(stage(True, True, False, False))
print(stage(True, True, True, False))

processing-or-export-gap
identity-or-query-selection-gap

Check identity, temporality, and transformations

Compare resource mapping, normalized names, attribute filters, histogram representation, and schema revisions with the consumer query. A removed attribute can cause a selector to match nothing, while a renamed service can split history. If the underlying quantity changed, do not patch the panel by selecting both streams and adding them blindly. Review whether old and new populations overlap. Check documented component availability and active pipeline wiring for the exact collector distribution rather than copying a configuration example from another release.

Repair the narrow failed contract

Choose the correction supported by evidence: restore required emission, repair authorized pipeline wiring, revise a documented transform, resolve export access through the owner, or migrate the query identity. Preserve a reviewed rollback path for any real configuration change and avoid broad simultaneous tuning. Keep missing and stale states visible during recovery. If evidence remains incomplete, report unknown coverage rather than converting absence to zero. A lower error chart after data disappeared is not an improvement in service behavior.

Keep it from recurring

• Version telemetry contracts and consumer queries together, retain migration fixtures, and monitor pipeline acceptance, rejection, export, queue pressure, and freshness. Review changes to filtering and resource mapping before rollout. Keep a visible coverage state so future loss cannot masquerade as healthy zero activity.