Assessing a T-fleet takes **two complementary measures** — together they decide targeted-vs-fleet-wide. **Measure 1 — repeat offenders (who to act on).** Per-instance `burn_ratio > 1` sustained, confirmed by % of 5-min periods over baseline (persistence separates chronic from one-off). Output is a discrete list. **Exclude these from Measure 2**, so the aggregate reads the *remaining* fleet cleanly. **Measure 2 — rest-of-fleet envelope (is targeted enough?).** "Within envelope" = the remaining fleet is net credit-accruing (`burn_ratio < 1` with margin). Aggregate across mixed sizes two ways: **mean of per-instance `(avg_CPU − baseline)`** (broad drift, equal weight) and **pooled `Σspent ÷ Σearned`** (total credit economy, size-weighted). Report both — they diverge when a few large instances dominate. **Trap — the mean hides a second tier.** After removing offenders, the rest can be bimodal: most low, a sub-tier creeping toward `burn_ratio = 1`. The mean looks fine while that sub-tier becomes your next offenders. Also report a **high percentile (p90)** of the rest; mean healthy *and* p90 < 1 = genuinely within envelope. **Read both measures at projected load, not just current** (apply the growth factor): | Rest-of-fleet | Decision | |---|---| | Within envelope now **and** under forecast | Targeted suffices — fix offenders only | | Within now, breaches under forecast | Fleet-wide needed | | Breaching now | Fleet-wide, urgent | The offender/rest cut is a **threshold choice, not a fact** — instances at `burn_ratio ≈ 1.0–1.1` are ambiguous; document where you drew the line. Builds on [[Credit Burn Ratio for Burstable Fleets]] and [[Sizing Capacity for Forecasted Load]]; feeds [[EC2 Capacity Planning for Burstable Fleets]]. --- *Source: synthesized from [[Credit Burn Ratio for Burstable Fleets]], [[Sizing Capacity for Forecasted Load]], and [[EC2 Capacity Planning for Burstable Fleets]].*