Definition
Rate Limit Equity Index (RLEI) measures the ratio of throttling rates between the most-throttled and least-throttled consumers within the same SLA tier.
In plain language: within a tier, calculate the throttling rate for each consumer.
The Measurement Gap
A B2B platform offers paid API access at three tiers: Starter, Professional, Enterprise. Each tier has a documented rate limit. The platform added an AI-driven rate limiter six months ago that observes per-consumer call patterns and dynamically adjusts throttling within tier limits to protect overall system stability. The platform marketing team likes the change because they can describe it as "intelligent rate limiting." The platform engineering team likes it because incident rate has dropped 23%.
Then the customer success leader receives a pattern of complaints from Professional-tier customers. Several customers report that their integration is being throttled aggressively while other Professional-tier customers seem unaffected. The complaining customers have the cleanest implementations: steady call patterns, modest concurrency, polite retry logic. The customers not being throttled include ones with messier patterns: burst traffic, aggressive retries, occasional polling spikes. The AI rate limiter has learned that consumers with predictable patterns can be safely throttled (they will not break) while consumers with unpredictable patterns must be protected (they might fail catastrophically if throttled). The result is that well-behaved consumers subsidize the protection of less-well-behaved ones, within the same paid tier.
The legacy API metrics report a healthy rate limiting program.
Rate limit enforcement count: trending appropriately with traffic volume.
Throttling events per consumer (aggregate): tracked.
System stability incidents: down 23%.
Overall throughput: stable.
Per-tier rate limit compliance: 100%. No consumer exceeded their tier limit.
Customer satisfaction surveys: stable on average.
None of these report on whether throttling is being applied equitably within a tier. The aggregate throttling events per consumer would show the disparity, but only if someone looked at the distribution; the headline number was always "throttling events trending appropriately." The CSAT survey averages mask the dissatisfaction of the well-behaved consumers because the population is dominated by consumers with normal traffic who never notice.
These metrics were designed for static rate limiting where the tier's limit was applied uniformly to all consumers in the tier. With AI-driven dynamic rate limiting, the limit becomes a budget that the AI allocates differentially based on what it has learned about each consumer. The differential allocation is the optimization. Whether the differential is equitable is a question that did not arise in static rate limiting and has not been operationalized in monitoring.
There is no metric that tells you whether AI-driven rate limiting disproportionately throttles well-behaved consumers while protecting fragile ones, which means service quality inequality grows invisibly.
Formula
Rate Limit Equity Index (RLEI) measures the ratio of throttling rates between the most-throttled and least-throttled consumers within the same SLA tier. In plain language: within a tier, calculate the throttling rate for each consumer. Find the most-throttled and least-throttled consumers. RLEI is the ratio of least-throttled rate to most-throttled rate. A value near 1 means consumers in the tier are treated equally. A value far below 1 means significant disparity.
The formula:
RLEI = Throttle_rate_min_same_tier / Throttle_rate_max_same_tier
Where:
- Throttle_rate_max_same_tier = the highest throttling percentage observed within the SLA tier during the measurement period
- Throttle_rate_min_same_tier = the lowest throttling percentage observed within the SLA tier during the measurement periodUnits are a ratio from 0 to 1. A value of 1.0 means equitable throttling. Below 1.0 means inequitable. A worked example. The B2B platform runs RLEI on their Professional tier over a measurement week:Consumer throttling rates in Professional tier (percentage of requests that were throttled):| Consumer | Requests | Throttled | Throttle rate ||----------|----------|-----------|---------------|| Consumer A | 124,000 | 980 | 0.79% || Consumer B | 89,000 | 2,140 | 2.40% || Consumer C | 156,000 | 540 | 0.35% || Consumer D | 78,000 | 8,420 | 10.79% || Consumer E | 211,000 | 1,890 | 0.90% || Consumer F | 47,000 | 7,580 | 16.13% || Consumer G | 198,000 | 1,210 | 0.61% || Consumer H | 134,000 | 4,650 | 3.47% |Throttle_rate_max = 16.13% (Consumer F). Throttle_rate_min = 0.35% (Consumer C).
RLEI = 0.35 / 16.13 = 0.022
Per the catalog guidance (above 0.70 reasonable equity, below 0.40 systematic disadvantage) this RLEI of 0.022 is severely inequitable. Consumer F is being throttled 46 times more than Consumer C while paying for the same tier. The team needs to investigate the rate limiter's learning model and likely impose a floor on per-consumer allocation within a tier, accepting some reduction in system-level optimization in exchange for equity. Three edge cases need explicit treatment. First, some disparity may be justified by traffic patterns (burst-heavy consumers genuinely need different handling than steady-state consumers). Track the disparity alongside an attribution to traffic pattern; not every low RLEI requires intervention if the disparity is explainable by legitimate operational difference. Second, consumers with zero throttling produce RLEI = 0 trivially. Distinguish "no consumer was throttled" (good, exclude from calculation) from "the favored consumers had zero throttling while others were throttled" (the canonical inequity case). Third, the minimum is 5+ consumers per SLA tier; below that, the min/max ratio is dominated by single-consumer variation. Cross-references. Downstream of RLEI is
True Remediation Rate (TRR, GMI-IT-007): heavily-throttled consumers produce support tickets and complaints whose remediation often gets misattributed to consumer-side problems when the actual cause is the rate limiter. The misinterpretation to watch for: a team reports RLEI = 0.8 (above the threshold) and concludes throttling is equitable. They miss that the calculation excluded one Enterprise-tier consumer whose throttling rate was anomalously high because of a specific use case, and including them would drop RLEI to 0.3. RLEI is sensitive to outliers; report the value alongside the underlying distribution rather than only the headline ratio.
Proposed formula, pending validation pilots.