TRAFFIC SPLITTING AND ROUTING
Production A/B testing for AI models requires traffic splitting at the inference gateway. Envoy and Kong Gateway support weighted routing with 1 percent granularity. A typical deployment routes 95 percent traffic to the current model (control) and 5 percent to the candidate model (treatment). Traffic splitting must maintain user session consistency using consistent hashing on request ID or user token.
Shadow traffic patterns deploy the candidate model receiving 100 percent mirrored traffic without affecting production responses. This enables latency and resource measurement without user impact but cannot measure quality metrics requiring actual responses. Shadow deployments limit 80th percentile traffic to 200 QPS for cost control. Active A/B testing requires minimum 500 requests per variant for statistical validity at 80 percent power.
| Traffic Pattern | Production Impact | Statistical Power | Risk | Cost |
|---|---|---|---|---|
| Shadow (100% mirror) | None | None (no response) | Minimal | 2x inference cost |
| Canary (5% traffic) | Minimal if poor | 80% at 5K reqs | Limited blast radius | 1.05x inference cost |
| A/B (50/50 split) | 50% users affected | 95% at 2K reqs | Equal blast radius | 2x inference cost |
| Multi-armed bandit | Adaptive | Faster convergence | Optimized exploration | 1.1-1.5x inference cost |
| Interleaved | Minimal (relative) | 98% at 1K reqs | Complex isolation | 2x inference cost |
QUALITY AND PERFORMANCE METRICS
A/B testing for LLM inference measures both quality and performance dimensions. Quality metrics: task-specific accuracy via ground truth labels, BLEU/ROUGE scores for generation tasks, embedding cosine similarity between model outputs, and human feedback ratings sampled at 1-5 percent of requests. Performance metrics: P50/P95/P99 latency, throughput, TTFT, and GPU memory utilization.
Statistical significance testing uses a two-tailed t-test at p < 0.05 with minimum detectable effect of 2 percent for quality metrics and 5 percent for performance metrics. Sequential testing with alpha spending function enables continuous monitoring without inflating false positive rate. Bayesian A/B testing with Beta prior provides 30-50 percent faster convergence for low-traffic model variants.
AUTOMATED ROLLBACK AND GUARDRAILS
Automated rollback triggers protect production quality. Typical guardrails: P99 latency increase exceeding 20 percent over 5-minute window triggers immediate rollback. Quality metric decline exceeding 3 percent over 15-minute window triggers investigation with rollback at 5 percent decline. Error rate exceeding 1 percent triggers immediate rollback. Rollback completes within 60 seconds via inference gateway configuration update.
Progressive rollout with automated gates stages deployment: 1 percent traffic for 30 minutes, expand to 5 percent for 2 hours, expand to 25 percent for 4 hours, expand to 100 percent. Automated gating at each stage requires all metrics within threshold for the duration. Progressive rollout reduces production incidents by 70 percent compared to direct 100 percent rollout.
INFRASTRUCTURE REQUIREMENTS
A/B testing infrastructure adds 15-25 percent to inference deployment complexity. Requirements include: dedicated GPU instances for the candidate model during testing, inference gateway with traffic routing support, metric aggregation pipeline with 60-second latency, and dashboard for real-time metric comparison. At 1,000 requests per second, the metric pipeline processes 86.4 million events daily.
Cost of A/B testing infrastructure: additional GPU capacity for candidate model (50-100 percent of production capacity during test), observability pipeline ($500-$2,000/month), and engineering time for test design and analysis (40-80 hours per test). Despite these costs, teams running systematic A/B tests achieve 2-3x faster model improvement velocity.