专业货物及培训服务器

@dataclass(frozen=True)
class FeatureDefinition:
    name: str
    entity: str
    aggregate: str
    window_days: int | None
    default: float
    version: int = 1
    owner: str = "growth-team"
    description: str = ""

    def compute(self, events: list[Event], as_of: int) -> float:
        visible = [e for e in events if e.day <= as_of]
        if self.window_days is not None:
            visible = [e for e in visible if e.day > as_of - self.window_days]
        if self.aggregate == "count":
            return float(len(visible))
        if self.aggregate == "sum":
            return round(sum(e.amount for e in visible), 2)
        if self.aggregate == "days_since_last":
            return float(as_of - max(e.day for e in visible)) if visible else self.default
        raise ValueError(f"unknown aggregate {self.aggregate!r}")

def backfill(registry: FeatureRegistry, names: list[str], events: list[Event], days: list[int]) -> list[FeatureRow]:
    rows = []
    for entity, entity_events in sorted(group_events(events).items()):
        for day in days:
            rows.append(FeatureRow(entity, day, registry.compute(names, entity_events, day)))
    return rows

def point_in_time_join(labels: list[LabelRow], rows: list[FeatureRow], ttl_days: int | None = None) -> list[JoinedRow]:
    by_entity: dict[str, list[FeatureRow]] = defaultdict(list)
    for row in rows:
        by_entity[row.entity].append(row)
    days_index = {}
    for entity, entity_rows in by_entity.items():
        entity_rows.sort(key=lambda r: r.day)
        days_index[entity] = [r.day for r in entity_rows]
    joined = []
    for label in labels:
        days = days_index.get(label.entity, [])
        pos = bisect.bisect_right(days, label.day) - 1
        match = by_entity[label.entity][pos] if pos >= 0 else None
        if match is not None and ttl_days is not None and label.day - match.day > ttl_days:
            match = None
        joined.append(JoinedRow(
            label.entity, label.day, label.label,
            match.day if match else None, match.values if match else None,
        ))
    return joined

  naive join used a snapshot from after the label day for 100% of rows
  offline accuracy, naive join: 0.948 (rule: idle > 53 days)
  offline accuracy, PIT join:   0.893 (rule: idle > 19 days)
  accuracy on day 145, trained with naive join: 0.763
  accuracy on day 145, trained with PIT join:   0.876

class OnlineStore:
    def __init__(self, ttl_days: int):
        self.ttl_days = ttl_days
        self._latest: dict[str, FeatureRow] = {}

    def write(self, row: FeatureRow) -> None:
        current = self._latest.get(row.entity)
        if current is None or row.day >= current.day:
            self._latest[row.entity] = row

    def read(self, entity: str, now: int) -> dict[str, float] | None:
        row = self._latest.get(entity)
        if row is None or row.day > now or now - row.day > self.ttl_days:
            return None
        return dict(row.values)

def detect_skew(
    training: dict[tuple[str, int], dict[str, float]],
    serving: dict[tuple[str, int], dict[str, float]],
    features: list[str],
    tolerance: float = 1e-6,
    max_mismatch_rate: float = 0.01,
) -> list[SkewReport]:
    if set(training) != set(serving):
        raise ValueError(f"{len(set(training) ^ set(serving))} keys are in only one view: log and compare the same requests")
    if not training:
        raise ValueError("no requests to compare")
    keys = sorted(training)
    reports = []
    for feature in features:
        diffs = [abs(training[k][feature] - serving[k][feature]) for k in keys]
        mismatches = sum(1 for d in diffs if d > tolerance)
        rate = mismatches / len(keys)
        mean_diff = sum(diffs) / len(keys)
        reports.append(SkewReport(feature, len(keys), rate, mean_diff, rate > max_mismatch_rate))
    return reports

python3 code/main.py
npx tsx code/main.ts

=== Skew checks on 383 logged requests, days 141 to 150 ===
  online store read  vs  training join (one definition, same snapshots)
    ok   days_since_last_purchase   mismatch   0.0%  mean |diff|    0.00
    ok   purchases_30d              mismatch   0.0%  mean |diff|    0.00
    ok   spend_30d                  mismatch   0.0%  mean |diff|    0.00
  fresh compute      vs  training join (staleness only)
    SKEW days_since_last_purchase   mismatch  83.0%  mean |diff|    2.77
    SKEW purchases_30d              mismatch  30.5%  mean |diff|    0.38
    SKEW spend_30d                  mismatch  37.1%  mean |diff|   16.77
  second code path   vs  fresh compute (code differences only)
    SKEW days_since_last_purchase   mismatch   5.5%  mean |diff|   20.07
    SKEW purchases_30d              mismatch  12.5%  mean |diff|    0.13
    SKEW spend_30d                  mismatch  12.5%  mean |diff|    5.09

from datetime import timedelta

from feast import Entity, FeatureView, Field, FileSource
from feast.types import Float32, Int64

user = Entity(name="user", join_keys=["user_id"])

user_activity_source = FileSource(
    name="user_activity_source",
    path="data/user_activity.parquet",
    timestamp_field="event_timestamp",
    created_timestamp_column="created",
)

user_activity = FeatureView(
    name="user_activity",
    entities=[user],
    ttl=timedelta(days=10),
    schema=[
        Field(name="purchases_30d", dtype=Int64),
        Field(name="spend_30d", dtype=Float32),
        Field(name="days_since_last_purchase", dtype=Int64),
    ],
    source=user_activity_source,
)

from feast import FeatureStore

store = FeatureStore(repo_path=".")

training_df = store.get_historical_features(
    entity_df=labels_df,
    features=["user_activity:purchases_30d", "user_activity:days_since_last_purchase"],
    filter_by_created_timestamp=True,
).to_df()

online = store.get_online_features(
    features=["user_activity:purchases_30d", "user_activity:days_since_last_purchase"],
    entity_rows=[{"user_id": 42}],
).to_dict()