Service-Level Indicators That Survive Refactors

An SLI tied to an internal function name dies the day someone refactors. Durable SLIs are tied to user-facing behaviour, not code shape.

Why SLIs rot

An SLI defined as ‘processOrder() latency at p99’ works until someone renames the function. The metric stops emitting; the SLO appears to be at 100%; nobody notices for weeks.

Refactors are a constant. SLIs that depend on code internals are coupled to that constant. Decoupling is the whole job.

Four properties of a durable SLI

• 1. Behaviour-defined, not code-defined. ‘HTTP request to /orders returning 2xx within 500ms.’ Code can be refactored arbitrarily.
• 2. Edge-measured, not internal. Measure where the user hits the system, not where the code happens to be.
• 3. Owns its semantics. The SLI definition is part of the contract; changes require explicit revision.
• 4. Tested. A change in code that breaks the SLI’s emit path fails CI.

Implementation patterns

Implement at the request boundary, ingress controller, service mesh sidecar, gateway. The metric exists at a layer that does not move with code refactors.

Where edge measurement is not possible (background workers), use a wrapper utility that the function calls; the wrapper emits the metric. Refactors that remove the wrapper are explicit.

Test suite for SLI drift

Add a test in CI that hits a known endpoint and checks that the SLI metric increments. If the metric stops emitting, CI fails before merge.

Add a synthetic check in production that exercises the SLI path on a known cadence. Recovery from any drift is automatic; the alert is loud.

Antipatterns

• SLIs defined by team that does not maintain the code. The SLI rots without an owner.
• SLI definition only in the dashboard. The dashboard is not source-of-truth; YAML in repo is.
• One SLI per microservice. User journeys span services; SLIs should too.

What to do this week

Three moves. (1) Inventory your SLIs; flag any defined by code-internal names. (2) Move one to edge measurement at the ingress. (3) Add the CI test that catches drift.