---

name: run-ab-test-models locale: caveman source_locale: en source_commit: 82c77053 fence_basis_commit: 82c77053 translator: "Julius Brussee homage — caveman" translation_date: "2026-05-03" description: > Design and execute A/B tests for ML models in production using traffic splitting, statistical significance testing, and canary/shadow deployment strategies. Measure performance differences and make data-driven decisions about model rollout. Use when validating a new model version before full rollout, comparing candidate models trained with different algorithms, measuring business metric impact of model changes, or when regulatory requirements mandate gradual rollout. license: MIT allowed-tools: Read Write Edit Bash Grep Glob metadata: author: Philipp Thoss version: "1.0" domain: mlops complexity: intermediate language: multi tags: ab-testing, canary, shadow-deployment, traffic-splitting, statistical-significance, experimentation

Run A/B Test for Models

See Extended Examples for complete configuration files and templates.

Run controlled experiments comparing model versions with traffic split + statistical analysis.

When Use

• Deploy new model version, want validate before full rollout
• Compare multiple candidate models (different algorithms, features)
• Test impact of hyperparameter changes on business metrics
• Measure model performance in prod without risk full traffic
• Regulatory needs gradual rollout (medical ML)
• Judge cost-performance tradeoffs between model sizes

Inputs

• Required: Champion model (current prod)
• Required: Challenger model(s) (new version to test)
• Required: Traffic allocation % (e.g., 5% to challenger)
• Required: Success metrics (business + ML)
• Required: Min sample size or test duration
• Optional: Guardrail metrics (latency, error rate thresholds)
• Optional: User segments for stratified testing

Steps

Step 1: Design Experiment

Define test parameters, success criteria, statistical needs.

# ab_test/experiment_config.py
from dataclasses import dataclass
from typing import List, Dict
import numpy as np
from scipy.stats import norm


@dataclass
# ... (see EXAMPLES.md for complete implementation)

Got: Experiment config with stat-sound sample size calc, typical 5-10k samples per variant for 5-10% MDE.

If fail: Sample too large? Up traffic allocation, extend duration, or accept larger MDE; verify baseline metric estimate; consider sequential testing.

Step 2: Implement Traffic Splitting

Set up routing — randomly assign requests to models.

# ab_test/traffic_router.py
import hashlib
import random
from typing import Dict, Optional
from dataclasses import dataclass
import logging

logger = logging.getLogger(__name__)
# ... (see EXAMPLES.md for complete implementation)

Got: Consistent user-to-variant assignment, accurate traffic split matches configured %, all assignments logged.

If fail: Verify hash uniform (test 10k user IDs), check user_id stable across requests (not session_id), logs capture all predictions, validate split in first 1000 requests.

Step 3: Implement Shadow Deployment (Optional)

Run challenger in parallel without affecting users (shadow mode).

# ab_test/shadow_deployment.py
import asyncio
from typing import Dict, Any
import logging
from concurrent.futures import ThreadPoolExecutor
import time

logger = logging.getLogger(__name__)
# ... (see EXAMPLES.md for complete implementation)

Got: Champion served at normal latency, challenger logged async without blocking, prediction diffs captured.

If fail: Set challenger timeout < champion SLA, handle challenger errors gracefully, monitor memory (two models loaded), consider sampling (log 10% of shadow predictions).

Step 4: Collect and Analyze Metrics

Gather data, run statistical tests.

# ab_test/analysis.py
import pandas as pd
import numpy as np
from scipy import stats
from typing import Dict, Tuple
import logging

logger = logging.getLogger(__name__)
# ... (see EXAMPLES.md for complete implementation)

Got: Stat test results with p-values, CIs, clear decision (rollout/keep/inconclusive), typical after 7-14 days or sample size.

If fail: Verify ground truth labels available (delayed analysis maybe), check sample ratio mismatch (SRM = assignment bugs), enough sample size, look for novelty/primacy effects in early data, consider sequential testing if fixed-horizon too slow.

Step 5: Monitor Guardrail Metrics

Continuous check challenger does not violate safety thresholds.

# ab_test/guardrails.py
import pandas as pd
import logging
from typing import Dict, List

logger = logging.getLogger(__name__)


# ... (see EXAMPLES.md for complete implementation)

Got: Guardrail violations detected within 5-15 min, auto stop if critical thresholds breached (latency, errors), alerts to team.

If fail: Verify thresholds realistic (not too tight), monitoring loop runs continuous, check stop_experiment() updates routing, test alert delivery.

Step 6: Make Rollout Decision

From results, decide rollout challenger.

# ab_test/rollout_decision.py
import logging
from typing import Dict
from dataclasses import dataclass

logger = logging.getLogger(__name__)


# ... (see EXAMPLES.md for complete implementation)

Got: Clear decision (full/gradual rollout, keep champion, extend test) with justification + action items.

If fail: Decision unclear? Subgroup analysis (segment, time, device), check interaction effects, review business context (2% lift worth eng cost?), consult stakeholders.

Checks

• Traffic split matches configured % (within 1%)
• Same user always to same variant
• Sample size calc reasonable (5-50k per variant)
• Stat tests produce p-values consistent with manual calc
• Guardrail violations trigger alerts within 5 min
• Shadow deployment shows <5% prediction divergence
• Reports include CIs
• Rollout decision documented

Pitfalls

• Sample ratio mismatch (SRM): Observed split differs from configured (95/5 becomes 92/8) = assignment bug; check hash uniformity
• Peeking: Check results before sample size inflates Type I error; use sequential testing or wait for pre-set end date
• Novelty effect: Users respond different to new model at first; run 2+ weeks for steady state
• Carryover effects: Prev variant exposure affects current; use new users or washout
• Multiple testing: Many metrics = false positive risk; correct with Bonferroni or single primary metric
• Insufficient power: Small allocation = months to detect; balance power with risk
• Ignore segments: Aggregate lift hides negative on important segments; subgroup analysis
• Attribution errors: Outcome metrics attributed to predictions (not other system changes)

See Also

• deploy-ml-model-serving - Model deployment infra, versioning
• monitor-model-drift - Post-rollout monitoring