Spaces:
Sleeping
Sleeping
File size: 24,206 Bytes
df31aa1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 | """
A/B Testing Framework for Cognexa ML Service
Provides a complete A/B testing system for:
- Feature experiments (UI variants, algorithm variants)
- Model comparison (comparing prediction model versions)
- Notification strategies (channel, timing, content)
- Recommendation variants (algorithm A vs B)
Implements:
- User bucketing (consistent hash-based assignment)
- Statistical significance testing
- Early stopping (O'Brien-Fleming bounds)
- Multiple comparison correction (Bonferroni)
- Experiment lifecycle management
"""
from __future__ import annotations
import hashlib
import json
import logging
import uuid
from dataclasses import dataclass, asdict, field
from datetime import datetime, timedelta
from enum import Enum
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
from scipy import stats as scipy_stats
from statistical_analysis import compare_groups, _power_analyzer
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Enums
# ---------------------------------------------------------------------------
class ExperimentStatus(str, Enum):
DRAFT = "draft"
RUNNING = "running"
PAUSED = "paused"
COMPLETED = "completed"
STOPPED_EARLY = "stopped_early"
class ExperimentType(str, Enum):
FEATURE_FLAG = "feature_flag" # on/off split
MODEL_COMPARISON = "model_comparison"
NOTIFICATION = "notification"
UI_VARIANT = "ui_variant"
RECOMMENDATION = "recommendation"
class AllocationStrategy(str, Enum):
HASH = "hash" # deterministic, consistent
RANDOM = "random"
# ---------------------------------------------------------------------------
# Data Structures
# ---------------------------------------------------------------------------
@dataclass
class Variant:
"""A single variant (arm) in an experiment."""
variant_id: str
name: str # e.g. "control", "treatment_a"
description: str
allocation_percent: float # 0-100
config: Dict[str, Any] = field(default_factory=dict)
is_control: bool = False
@dataclass
class Experiment:
"""A/B experiment definition."""
experiment_id: str
name: str
description: str
experiment_type: str # ExperimentType
status: str # ExperimentStatus
variants: List[Variant]
primary_metric: str # e.g. "task_completion_rate"
secondary_metrics: List[str]
allocation_strategy: str # AllocationStrategy
traffic_percent: float # 0-100 fraction of users to include
min_sample_size: int # required per variant before analysis
alpha: float # significance level
target_power: float
created_at: str
started_at: Optional[str]
ended_at: Optional[str]
owner: str
tags: List[str] = field(default_factory=list)
metadata: Dict[str, Any] = field(default_factory=dict)
@dataclass
class ExperimentAssignment:
"""Records which variant a user is assigned to."""
assignment_id: str
experiment_id: str
user_id: str
variant_id: str
variant_name: str
assigned_at: str
is_exposed: bool = True # first exposure recorded
@dataclass
class MetricObservation:
"""A single metric measurement from an experiment participant."""
observation_id: str
experiment_id: str
user_id: str
variant_id: str
metric_name: str
value: float
recorded_at: str
@dataclass
class VariantResult:
"""Statistical result for a single variant."""
variant_id: str
variant_name: str
n_observations: int
mean: float
std: float
median: float
is_control: bool
@dataclass
class ExperimentResult:
"""Full statistical analysis result for an experiment."""
experiment_id: str
experiment_name: str
primary_metric: str
status: str
control_result: VariantResult
treatment_results: List[VariantResult]
# Significance
p_value: float
is_significant: bool
alpha: float
# Effect size
effect_size: float
relative_uplift: float # % improvement over control
# Power
current_power: float
recommended_sample_size: int
is_adequately_powered: bool
# Decision
winner: Optional[str] # variant_name of winner, None if no winner yet
recommendation: str
analyzed_at: str
@dataclass
class EarlyStoppingDecision:
"""Result of early stopping evaluation."""
should_stop: bool
reason: str
current_p_value: float
obrien_fleming_threshold: float
current_n: int
planned_n: int
interim_fraction: float
# ---------------------------------------------------------------------------
# User Bucketing
# ---------------------------------------------------------------------------
class UserBucketizer:
"""Consistent hash-based user assignment to experiment variants."""
def assign_variant(
self,
user_id: str,
experiment_id: str,
variants: List[Variant],
traffic_percent: float = 100.0,
strategy: str = AllocationStrategy.HASH,
) -> Optional[Variant]:
"""
Assign a user to a variant. Returns None if user not in experiment traffic.
"""
if strategy == AllocationStrategy.RANDOM:
bucket = np.random.uniform(0, 100)
else:
# Deterministic: hash(user_id + experiment_id)
key = f"{user_id}:{experiment_id}"
h = int(hashlib.md5(key.encode()).hexdigest(), 16)
bucket = (h % 10000) / 100.0 # 0.00 - 99.99
# Check traffic inclusion
if bucket >= traffic_percent:
return None
# Assign to variant based on allocation percentages
cumulative = 0.0
# Normalize allocation within traffic slice
total_alloc = sum(v.allocation_percent for v in variants)
for variant in variants:
cumulative += (variant.allocation_percent / total_alloc) * traffic_percent
if bucket < cumulative:
return variant
return variants[-1] # fallback
# ---------------------------------------------------------------------------
# Experiment Storage
# ---------------------------------------------------------------------------
class ExperimentStore:
"""Filesystem-backed store for experiments and observations."""
def __init__(self, data_dir: str = "data/ab_testing"):
self.data_dir = Path(data_dir)
self.data_dir.mkdir(parents=True, exist_ok=True)
self.experiments_file = self.data_dir / "experiments.json"
self.assignments_file = self.data_dir / "assignments.json"
self.observations_file = self.data_dir / "observations.json"
self._experiments: Dict[str, Experiment] = self._load_experiments()
self._assignments: List[ExperimentAssignment] = self._load_assignments()
self._observations: List[MetricObservation] = self._load_observations()
# -- Experiments ----------------------------------------------------------
def _load_experiments(self) -> Dict[str, Experiment]:
if not self.experiments_file.exists():
return {}
try:
with open(self.experiments_file) as f:
data = json.load(f)
return {
eid: Experiment(
**{k: v for k, v in exp.items() if k != "variants"},
variants=[Variant(**v) for v in exp.get("variants", [])],
)
for eid, exp in data.items()
}
except Exception as e:
logger.warning("Could not load experiments: %s", e)
return {}
def save_experiment(self, experiment: Experiment):
self._experiments[experiment.experiment_id] = experiment
with open(self.experiments_file, "w") as f:
json.dump(
{eid: asdict(exp) for eid, exp in self._experiments.items()},
f, indent=2
)
def get_experiment(self, experiment_id: str) -> Optional[Experiment]:
return self._experiments.get(experiment_id)
def list_experiments(self, status: Optional[str] = None) -> List[Experiment]:
exps = list(self._experiments.values())
if status:
exps = [e for e in exps if e.status == status]
return sorted(exps, key=lambda e: e.created_at, reverse=True)
# -- Assignments ----------------------------------------------------------
def _load_assignments(self) -> List[ExperimentAssignment]:
if not self.assignments_file.exists():
return []
try:
with open(self.assignments_file) as f:
return [ExperimentAssignment(**a) for a in json.load(f)]
except Exception as e:
logger.warning("Could not load assignments: %s", e)
return []
def save_assignment(self, assignment: ExperimentAssignment):
self._assignments.append(assignment)
if len(self._assignments) > 100000:
self._assignments = self._assignments[-100000:]
with open(self.assignments_file, "w") as f:
json.dump([asdict(a) for a in self._assignments], f)
def get_user_assignment(
self, user_id: str, experiment_id: str
) -> Optional[ExperimentAssignment]:
for a in reversed(self._assignments):
if a.user_id == user_id and a.experiment_id == experiment_id:
return a
return None
# -- Observations ----------------------------------------------------------
def _load_observations(self) -> List[MetricObservation]:
if not self.observations_file.exists():
return []
try:
with open(self.observations_file) as f:
return [MetricObservation(**o) for o in json.load(f)]
except Exception as e:
logger.warning("Could not load observations: %s", e)
return []
def save_observation(self, obs: MetricObservation):
self._observations.append(obs)
if len(self._observations) > 500000:
self._observations = self._observations[-500000:]
with open(self.observations_file, "w") as f:
json.dump([asdict(o) for o in self._observations], f)
def get_observations(
self,
experiment_id: str,
metric_name: str,
) -> Dict[str, List[float]]:
"""Return dict: variant_id -> list of metric values."""
result: Dict[str, List[float]] = {}
for obs in self._observations:
if obs.experiment_id == experiment_id and obs.metric_name == metric_name:
result.setdefault(obs.variant_id, []).append(obs.value)
return result
# ---------------------------------------------------------------------------
# Statistical Analyzer
# ---------------------------------------------------------------------------
class ExperimentAnalyzer:
"""Analyzes experiment results and computes statistical significance."""
def analyze(
self,
experiment: Experiment,
observations: Dict[str, Dict[str, List[float]]], # variant_id -> metric -> values
) -> ExperimentResult:
primary = experiment.primary_metric
obs_by_variant = observations.get(primary, {})
control_variant = next(
(v for v in experiment.variants if v.is_control), experiment.variants[0]
)
control_values = obs_by_variant.get(control_variant.variant_id, [])
treatment_variants = [v for v in experiment.variants if not v.is_control]
def _summarize(v: Variant) -> VariantResult:
vals = obs_by_variant.get(v.variant_id, [])
return VariantResult(
variant_id=v.variant_id,
variant_name=v.name,
n_observations=len(vals),
mean=round(float(np.mean(vals)), 4) if vals else 0.0,
std=round(float(np.std(vals, ddof=1)), 4) if len(vals) > 1 else 0.0,
median=round(float(np.median(vals)), 4) if vals else 0.0,
is_control=v.is_control,
)
ctrl_result = _summarize(control_variant)
treat_results = [_summarize(v) for v in treatment_variants]
# Primary comparison: control vs best treatment
best_treat = max(treat_results, key=lambda r: r.mean) if treat_results else None
best_values = obs_by_variant.get(best_treat.variant_id, []) if best_treat else []
p_value = 1.0
effect_size = 0.0
if len(control_values) >= 2 and len(best_values) >= 2:
test_result = compare_groups(control_values, best_values, test="auto", alpha=experiment.alpha)
p_value = test_result["p_value"]
effect_size = test_result["effect_size"]
is_significant = p_value < experiment.alpha
relative_uplift = (
(best_treat.mean - ctrl_result.mean) / ctrl_result.mean * 100
if ctrl_result.mean != 0 and best_treat else 0.0
)
# Power
total_n = sum(len(v) for v in obs_by_variant.values())
per_variant_n = max(1, total_n // max(1, len(experiment.variants)))
power_result = _power_analyzer.compute_sample_size(
effect_size=max(0.01, abs(effect_size)),
alpha=experiment.alpha,
power=experiment.target_power,
)
is_powered = per_variant_n >= power_result.required_sample_size
# Winner
winner = None
if is_significant and is_powered and best_treat and relative_uplift > 0:
winner = best_treat.variant_name
elif is_significant and is_powered and best_treat and relative_uplift < 0:
winner = control_variant.name
recommendation = self._recommendation(
is_significant, is_powered, winner, per_variant_n,
power_result.required_sample_size, relative_uplift
)
return ExperimentResult(
experiment_id=experiment.experiment_id,
experiment_name=experiment.name,
primary_metric=primary,
status=experiment.status,
control_result=ctrl_result,
treatment_results=treat_results,
p_value=round(p_value, 6),
is_significant=is_significant,
alpha=experiment.alpha,
effect_size=round(effect_size, 4),
relative_uplift=round(relative_uplift, 2),
current_power=round(power_result.current_power, 4),
recommended_sample_size=power_result.required_sample_size,
is_adequately_powered=is_powered,
winner=winner,
recommendation=recommendation,
analyzed_at=datetime.utcnow().isoformat(),
)
def check_early_stopping(
self,
current_p: float,
current_n: int,
planned_n: int,
alpha: float = 0.05,
) -> EarlyStoppingDecision:
"""O'Brien-Fleming alpha spending for early stopping."""
if planned_n <= 0 or current_n <= 0:
return EarlyStoppingDecision(False, "Insufficient data", 1.0, alpha, 0, planned_n, 0.0)
fraction = min(1.0, current_n / planned_n)
# O'Brien-Fleming boundary: alpha_spent = alpha * (2 - 2*Phi(z_alpha / sqrt(fraction)))
# Simplified: threshold scales with 1/sqrt(fraction)
z_alpha = scipy_stats.norm.ppf(1 - alpha / 2)
if fraction < 0.01:
threshold = 1e-6 # essentially never stop very early
else:
obf_z = z_alpha / np.sqrt(fraction)
threshold = 2 * scipy_stats.norm.sf(obf_z) # two-tailed p-value threshold
should_stop = current_p < threshold
reason = (
f"Early stopping: p={current_p:.4f} < OBF threshold={threshold:.4f}"
if should_stop
else f"Continue: p={current_p:.4f} >= OBF threshold={threshold:.4f} at {fraction:.0%} interim"
)
return EarlyStoppingDecision(
should_stop=should_stop,
reason=reason,
current_p_value=round(current_p, 6),
obrien_fleming_threshold=round(threshold, 6),
current_n=current_n,
planned_n=planned_n,
interim_fraction=round(fraction, 4),
)
def _recommendation(
self,
significant: bool,
powered: bool,
winner: Optional[str],
current_n: int,
required_n: int,
uplift: float,
) -> str:
if not powered:
return (
f"Continue collecting data. Need {required_n} per variant, "
f"currently at {current_n}."
)
if not significant:
return "No significant difference. Consider continuing or stopping (null hypothesis)."
if winner:
return (
f"Ship variant '{winner}' - statistically significant "
f"({'↑' if uplift > 0 else '↓'}{abs(uplift):.1f}% vs control)."
)
return "Re-examine experiment design."
# ---------------------------------------------------------------------------
# A/B Testing Manager (main API)
# ---------------------------------------------------------------------------
class ABTestingManager:
"""High-level manager for A/B experiments."""
def __init__(self):
self.store = ExperimentStore()
self.bucketizer = UserBucketizer()
self.analyzer = ExperimentAnalyzer()
def create_experiment(
self,
name: str,
description: str,
primary_metric: str,
variants: List[Dict[str, Any]],
experiment_type: str = ExperimentType.FEATURE_FLAG,
secondary_metrics: Optional[List[str]] = None,
traffic_percent: float = 100.0,
min_sample_size: int = 100,
alpha: float = 0.05,
target_power: float = 0.80,
owner: str = "system",
tags: Optional[List[str]] = None,
) -> Experiment:
experiment_id = str(uuid.uuid4())
parsed_variants = [
Variant(
variant_id=str(uuid.uuid4()),
name=v["name"],
description=v.get("description", ""),
allocation_percent=v.get("allocation_percent", 100.0 / len(variants)),
config=v.get("config", {}),
is_control=v.get("is_control", False),
)
for v in variants
]
# If no control marked, mark first as control
if not any(v.is_control for v in parsed_variants):
parsed_variants[0].is_control = True
exp = Experiment(
experiment_id=experiment_id,
name=name,
description=description,
experiment_type=experiment_type,
status=ExperimentStatus.DRAFT,
variants=parsed_variants,
primary_metric=primary_metric,
secondary_metrics=secondary_metrics or [],
allocation_strategy=AllocationStrategy.HASH,
traffic_percent=traffic_percent,
min_sample_size=min_sample_size,
alpha=alpha,
target_power=target_power,
created_at=datetime.utcnow().isoformat(),
started_at=None,
ended_at=None,
owner=owner,
tags=tags or [],
)
self.store.save_experiment(exp)
logger.info("Created experiment %s (%s)", name, experiment_id)
return exp
def start_experiment(self, experiment_id: str) -> bool:
exp = self.store.get_experiment(experiment_id)
if not exp:
return False
exp.status = ExperimentStatus.RUNNING
exp.started_at = datetime.utcnow().isoformat()
self.store.save_experiment(exp)
return True
def stop_experiment(self, experiment_id: str, early: bool = False) -> bool:
exp = self.store.get_experiment(experiment_id)
if not exp:
return False
exp.status = ExperimentStatus.STOPPED_EARLY if early else ExperimentStatus.COMPLETED
exp.ended_at = datetime.utcnow().isoformat()
self.store.save_experiment(exp)
return True
def get_variant_for_user(
self, user_id: str, experiment_id: str
) -> Optional[Dict[str, Any]]:
"""
Assign (or retrieve cached assignment) for a user in an experiment.
Returns dict with variant info or None if not in experiment.
"""
exp = self.store.get_experiment(experiment_id)
if not exp or exp.status != ExperimentStatus.RUNNING:
return None
# Check existing assignment
existing = self.store.get_user_assignment(user_id, experiment_id)
if existing:
variant = next((v for v in exp.variants if v.variant_id == existing.variant_id), None)
if variant:
return {"variant_id": variant.variant_id, "variant_name": variant.name,
"config": variant.config, "is_control": variant.is_control}
# New assignment
variant = self.bucketizer.assign_variant(
user_id, experiment_id, exp.variants,
exp.traffic_percent, exp.allocation_strategy
)
if variant is None:
return None
assignment = ExperimentAssignment(
assignment_id=str(uuid.uuid4()),
experiment_id=experiment_id,
user_id=user_id,
variant_id=variant.variant_id,
variant_name=variant.name,
assigned_at=datetime.utcnow().isoformat(),
)
self.store.save_assignment(assignment)
return {"variant_id": variant.variant_id, "variant_name": variant.name,
"config": variant.config, "is_control": variant.is_control}
def record_metric(
self,
experiment_id: str,
user_id: str,
metric_name: str,
value: float,
) -> bool:
"""Record a metric observation for a user in an experiment."""
assignment = self.store.get_user_assignment(user_id, experiment_id)
if not assignment:
return False
obs = MetricObservation(
observation_id=str(uuid.uuid4()),
experiment_id=experiment_id,
user_id=user_id,
variant_id=assignment.variant_id,
metric_name=metric_name,
value=value,
recorded_at=datetime.utcnow().isoformat(),
)
self.store.save_observation(obs)
return True
def analyze_experiment(self, experiment_id: str) -> Optional[Dict[str, Any]]:
"""Run statistical analysis on an experiment."""
exp = self.store.get_experiment(experiment_id)
if not exp:
return None
all_metrics = {exp.primary_metric} | set(exp.secondary_metrics)
observations: Dict[str, Dict[str, List[float]]] = {}
for metric in all_metrics:
observations[metric] = self.store.get_observations(experiment_id, metric)
result = self.analyzer.analyze(exp, observations)
# Check early stopping
total_n = sum(
len(v) for v in observations.get(exp.primary_metric, {}).values()
)
n_per_variant = max(1, total_n // max(1, len(exp.variants)))
stopping = self.analyzer.check_early_stopping(
result.p_value, n_per_variant,
exp.min_sample_size, exp.alpha
)
return {
**asdict(result),
"early_stopping": asdict(stopping),
}
def list_experiments(self, status: Optional[str] = None) -> List[Dict[str, Any]]:
exps = self.store.list_experiments(status)
return [asdict(e) for e in exps]
# ---------------------------------------------------------------------------
# Singleton
# ---------------------------------------------------------------------------
_manager_instance: Optional[ABTestingManager] = None
def get_ab_manager() -> ABTestingManager:
global _manager_instance
if _manager_instance is None:
_manager_instance = ABTestingManager()
return _manager_instance
|