#!/usr/bin/env python3
"""Lint a feature-store design spec (JSON, or YAML when PyYAML is installed) before generating code.
The spec describes entities, sources, feature views, features and feature services (see
references/spec-format.md; print an example with gen_feast_repo.py --template). Rules:
FS001 error entity without join_keys / feature view referencing an unknown entity or source
FS002 error source without timestamp_field (point-in-time joins impossible)
FS003 error duplicate feature name inside a view, or duplicate view / entity / source names
FS004 error feature without dtype, or a dtype outside the supported set
FS005 error feature service references an unknown view or feature
FS006 warning ttl missing, or shorter than the source's update cadence (online lookups return nothing)
FS007 warning online view without freshness_sla, or freshness_sla looser than ttl
FS008 warning names not snake_case, or windowed aggregates without a window suffix (_7d, _24h)
FS009 warning feature or view without description / owner
FS010 warning likely PII (email, phone, name, address, ssn, dob, ip) without pii: true
FS011 info view with more than 50 features (split by source cadence or entity)
FS012 warning label-like feature name (churned, is_fraud, target, label, outcome) inside a feature view
Usage:
python lint_feature_spec.py features.json
python lint_feature_spec.py features.yaml --json
Exit codes: 0 no errors, 1 errors, 2 read error.
"""
import argparse
import json
import re
import sys
DTYPES = {"int32", "int64", "float32", "float64", "string", "bool", "bytes", "unix_timestamp",
"array<int64>", "array<float32>", "array<float64>", "array<string>"}
UNITS = {"s": 1, "m": 60, "h": 3600, "d": 86400, "w": 604800}
CADENCE = {"realtime": 1, "streaming": 1, "minutely": 60, "hourly": 3600, "daily": 86400, "weekly": 604800,
"monthly": 2_678_400}
SNAKE = re.compile(r"^[a-z][a-z0-9_]*$")
AGG = re.compile(r"(count|sum|mean|avg|max|min|std|n_|nuniq|distinct|total)", re.I)
WINDOW = re.compile(r"_(\d+)(m|h|d|w|min|hr|days?|weeks?)$|_(lifetime|all_time|prior)$", re.I)
PII = re.compile(r"(email|phone|first_name|last_name|full_name|address|ssn|social_security|dob|birth|passport|"
r"ip_address|^ip$|_ip$|card_number|iban)", re.I)
LABELY = re.compile(r"^(churned|is_fraud|fraud|target|label|outcome|y|defaulted|converted)$", re.I)
def seconds(v):
if v is None:
return None
if isinstance(v, (int, float)):
return float(v)
m = re.fullmatch(r"\s*(\d+(?:\.\d+)?)\s*([smhdw])\s*", str(v))
return float(m.group(1)) * UNITS[m.group(2)] if m else None
def load(path):
text = sys.stdin.read() if path == "-" else open(path, encoding="utf-8").read()
if path.endswith((".yaml", ".yml")):
try:
import yaml
except ImportError:
raise ValueError("YAML spec needs PyYAML (pip install pyyaml), or convert it to JSON")
return yaml.safe_load(text)
return json.loads(text)
def lint(spec):
out = []
def add(code, sev, where, msg):
out.append({"code": code, "severity": sev, "where": where, "message": msg})
ents = {}
for e in spec.get("entities", []):
n = e.get("name", "?")
if n in ents:
add("FS003", "error", f"entity {n}", "duplicate entity name")
ents[n] = e
if not e.get("join_keys"):
add("FS001", "error", f"entity {n}", "no join_keys")
srcs = {}
for s in spec.get("sources", []):
n = s.get("name", "?")
if n in srcs:
add("FS003", "error", f"source {n}", "duplicate source name")
srcs[n] = s
if not s.get("timestamp_field"):
add("FS002", "error", f"source {n}", "no timestamp_field: point-in-time joins and TTL cannot work")
views = {}
all_feats = {}
for v in spec.get("feature_views", []):
n = v.get("name", "?")
where = f"view {n}"
if n in views:
add("FS003", "error", where, "duplicate feature view name")
views[n] = v
for en in v.get("entities", []):
if en not in ents:
add("FS001", "error", where, f"unknown entity {en!r}")
src = srcs.get(v.get("source"))
if not src:
add("FS001", "error", where, f"unknown source {v.get('source')!r}")
if not SNAKE.match(n):
add("FS008", "warning", where, "view name is not snake_case")
if not v.get("owner"):
add("FS009", "warning", where, "no owner")
if not v.get("description"):
add("FS009", "warning", where, "no description")
ttl = seconds(v.get("ttl"))
if ttl is None:
add("FS006", "warning", where, "no ttl: stale rows are served forever (or nothing, depending on the store)")
elif src and CADENCE.get(str(src.get("cadence", "")).lower()) and ttl < CADENCE[src["cadence"].lower()]:
add("FS006", "warning", where, f"ttl {v['ttl']} is shorter than the source cadence {src['cadence']!r}: "
"online reads go empty between refreshes")
if v.get("online"):
sla = seconds(v.get("freshness_sla"))
if sla is None:
add("FS007", "warning", where, "online view without freshness_sla")
elif ttl is not None and sla > ttl:
add("FS007", "warning", where, f"freshness_sla {v['freshness_sla']} is looser than ttl {v['ttl']}")
feats = v.get("features", [])
if len(feats) > 50:
add("FS011", "info", where, f"{len(feats)} features: consider splitting by source cadence or entity")
names = set()
for f in feats:
fn = f.get("name", "?")
fw = f"{n}:{fn}"
if fn in names:
add("FS003", "error", fw, "duplicate feature name in view")
names.add(fn)
all_feats[fw] = f
dt = str(f.get("dtype", "")).lower()
if not dt:
add("FS004", "error", fw, "no dtype")
elif dt not in DTYPES:
add("FS004", "error", fw, f"dtype {dt!r} not in {sorted(DTYPES)}")
if not SNAKE.match(fn):
add("FS008", "warning", fw, "feature name is not snake_case")
elif AGG.search(fn) and not WINDOW.search(fn) and not f.get("window"):
add("FS008", "warning", fw, "aggregate without a window suffix (_7d, _24h, _lifetime) or window field")
if not f.get("description"):
add("FS009", "warning", fw, "no description")
if PII.search(fn) and not f.get("pii"):
add("FS010", "warning", fw, "looks like PII: set pii: true and restrict access, or drop it")
if LABELY.match(fn):
add("FS012", "warning", fw, "label-like name inside a feature view: labels belong in the training "
"spine, not in served features")
for svc in spec.get("feature_services", []):
sn = svc.get("name", "?")
for ref in svc.get("features", []):
view, _, feat = ref.partition(":")
if view not in views:
add("FS005", "error", f"service {sn}", f"unknown view {view!r}")
elif feat and feat != "*" and f"{view}:{feat}" not in all_feats:
add("FS005", "error", f"service {sn}", f"unknown feature {ref!r}")
return out
def main():
ap = argparse.ArgumentParser(description="Lint a feature store spec.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("spec", help="JSON/YAML spec or - for stdin (JSON)")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
spec = load(a.spec)
if not isinstance(spec, dict):
raise ValueError("spec must be a mapping")
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
findings = lint(spec)
errors = sum(f["severity"] == "error" for f in findings)
if a.json:
print(json.dumps({"errors": errors, "findings": findings}, indent=2))
else:
for f in findings:
print(f"{f['code']} [{f['severity']}] {f['where']}: {f['message']}")
nv = len(spec.get("feature_views", []))
nf = sum(len(v.get("features", [])) for v in spec.get("feature_views", []))
print(f"\n{nv} view(s), {nf} feature(s): {len(findings)} finding(s), {errors} error(s)")
return 1 if errors else 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.