An SLI starts with eligible events and a measurement boundary

State the event unit: a user operation, an HTTP attempt, a completed task, or a requested byte range. Then define eligibility, including routes, tenants, environments, and any deliberate exclusions. Define good as a measurable predicate, such as a successful final result within a deadline. A status-only success indicator and a success-plus-latency indicator answer different questions. If expected invalid input is excluded, document the classification and verify it cannot hide genuine service failures. Eligibility changes should be reviewed as measurement-contract changes rather than quiet dashboard fixes.

The numerator must be a subset of the same denominator under the same window and source boundary. An internal server can miss failures before traffic reaches it, while a client can observe network or gateway behavior that the server never sees. A completion-only measurement can also miss work that never completes unless timeouts and abandoned operations are resolved according to policy. Choose where starts, terminal outcomes, rejections, and deadlines are recorded, then reconcile in-progress work. The measurement should reflect the promised experience rather than only the requests convenient for one process to count.

One user operation can generate several attempts, so an attempt-failure counter is not automatically a user-failure counter. Both are useful, but label the unit. Similarly, latency only for successful operations can look excellent while failed requests wait for long timeouts. Decide whether the indicator counts all eligible outcomes and which latency definition reflects the promise. Record duration using a suitable local timer at the chosen boundary, with consistent units and exactly one appropriate terminal classification per event. A retry log emitted several times must not multiply the denominator.

Align aggregation across instances and time. Pool good and eligible counts, not unweighted success percentages. Use compatible reset-aware rate or increase calculations for counter windows and retain the limits of sparse scrapes and lost accounting. Events that start before a window and finish after it require a defined assignment rule. Measurement coverage is separate from the indicator value: missing traffic, missing telemetry, and genuinely zero eligible traffic are different states. A default replacement of every missing series with zero can make an outage appear as an undefined or falsely healthy ratio.

Use known fixtures for a direct success, a retry that eventually succeeds, a final failure, an upstream rejection, an exceeded deadline, cancellation, and a request whose telemetry export fails. Predict each fixture's eligible and good contributions before checking the dashboard. Compare client and server measures to explain boundary differences rather than forcing them to agree. Include a no-traffic interval and a collection outage as distinct tests. Keep SLI metric collection separate from outcome-enriched trace selection unless a documented valid estimator and denominator supply the promised population measurement.

The sampling lab illustrates a hypothetical good-event fraction with strong independence assumptions. Reset, set both p=0.97 and n=100, predict mean 0.97 and variance 0.000291, then Play or Step. The exact binomial distribution is not an observed production window and playback does not draw random events. Its normal overlay is approximate, and its Berry–Esseen statement is about standardized CDF error rather than each bar. Real traffic can be correlated and systematically unmeasured, so the instrument cannot certify SLI coverage or turn a chosen p into evidence of actual reliability.