How do I plan an exit from an observability platform?

Inventory the objects people depend on

Raw logs, metrics, traces, and profiles are only part of an observability product. Teams depend on saved queries, alert definitions, dashboard variables, service catalogs, access controls, annotations, and links in incident records. Record owners and retention obligations for each. A migration that moves bytes while losing their interpretation can make prior incidents effectively unreadable.

Separate ingest portability from result portability

An OpenTelemetry-compatible export path may reduce producer lock-in, but platform-specific query languages, derived views, sampling rules, and storage layouts remain. Test a sample of raw and transformed data through the proposed path. Check units, temporality, resource identity, trace relationships, and timestamps after import. A successful HTTP transfer does not show that the next investigator will reach the same answer.

The old incident link becomes a dead end

In a fictional migration, a team switches exporters and immediately closes the old account. An incident report from last month links to a dashboard with a custom error-rate calculation; the link now fails and the new platform’s calculation excludes a tenant. The exit plan retains a governed historical view, records the formula, and tests an equivalent query before old access is retired.

Use an exit readiness register

Track each dependency through a measurable disposition.

Object	Portability test	Disposition
Source telemetry	Export and import sample	Migrate or retain
Derived metric	Compare formula and population	Rebuild and verify
Active alert	Shadow and cutover rehearsal	Transfer owner
Historical link	Open an old incident path	Preserve or document gap
Access rule	Test least-privilege users	Recreate policy

Dual-run with a clear interpretation rule

Send a bounded signal set to both systems while the old alert path remains authoritative. Compare counts and incident queries after accounting for sampling and ingestion delay. Record which platform is the source of truth for pages, corrections, and historical reports during each stage. Without that rule, differing values can trigger confusion or duplicate incidents instead of revealing a migration defect.

Retire only after the obligation is satisfied

Keep rollback target, contract dates, export rights, and a named owner for unmigrated history. Test a provider outage and a return to the old path before closing it. Some proprietary views may not be transferable; document the limitation and retain a lawful read-only archive if needed. Retiring an account is a separate decision from moving new telemetry.

Evidence boundary for platform exit

• OpenTelemetry vendor guidance: OpenTelemetry defines vendor support and implementation qualifications. Protocol support does not make saved queries or history portable.
• OpenTelemetry Collector resiliency: OpenTelemetry documents sending queues and failure recovery constraints. Those mechanisms do not guarantee complete dual-run equivalence.

The incident-link failure is invented. Export rights, retention, and historical access need contract and technical verification.