mlops-pipeline-design

Software CI assumes a git sha plus a lockfile determines the build.

SKILL.md

GitHub

MLOps Pipeline Design (Verbose)

Core Patterns

Reproducibility Requires Data Versioning, Not Just Code Versioning

Software CI assumes a git sha plus a lockfile determines the build. ML breaks that: the same commit trained on Tuesday's table and Thursday's table produces two different models, both labelled with the same sha.

A reproducible run pins code, data, and configuration under one identifier.

@dataclass(frozen=True)
class RunFingerprint:
    git_sha: str            # code, including the feature library
    dataset_uri: str        # s3://lake/churn/snapshot_date=2026-06-01/
    dataset_sha256: str     # hash of the sorted file manifest (path, size, etag)
    params_hash: str        # hash of the resolved config, defaults included
    env_hash: str           # hash of requirements.lock / conda env

    def run_key(self) -> str:
        return sha256("|".join(astuple(self)).encode()).hexdigest()[:16]

Hash a manifest (sorted paths plus etags) rather than streaming terabytes: it is cheap, stable, and detects files added, removed, or rewritten under you.

The most effective structural change is making stage one materialize an immutable snapshot:

CREATE TABLE churn_snapshot_2026_06_01 AS
SELECT * FROM events WHERE event_ts < TIMESTAMP '2026-06-01 00:00:00';

Mutable sources make the "reproducible" pipeline a fiction. Re-running last quarter's job against SELECT * FROM events trains on data that did not exist then — and if those rows post-date the labels, you have manufactured leakage that only shows up in production.

Stage Decomposition and Caching

ingest → validate → build_features → train → evaluate → register(staging)

Each stage declares its inputs and outputs, so the orchestrator (Kubeflow, Airflow, Metaflow, Dagster) can skip any stage whose inputs are unchanged.

Stage Cache key Typically re-runs when
ingest snapshot date New snapshot
validate dataset hash New data
build_features dataset hash + feature-lib sha Data or transform change
train features hash + params hash Hyperparameter sweep
evaluate model hash + holdout hash New model

Correct decomposition turns a hyperparameter sweep from hours of repeated ingestion into minutes of training. One script doing everything makes each experiment pay full price, which is why teams stop experimenting.

Fix the evaluation holdout by rule, not by random split per run:

# ✅ Deterministic, stable across runs and comparable between models
holdout = df[hash_bucket(df.entity_id, 100) < 10]           # fixed 10%
train_df = df[hash_bucket(df.entity_id, 100) >= 10]

# ❌ A fresh random split per run makes two runs' AUCs incomparable
train_df, holdout = train_test_split(df, test_size=0.1)     # no seed, no rule

For time-series or any temporally ordered target, split by time, never at random: a random split lets the model see the future for the same entity.

Data Validation as a Blocking Gate

def validate(df: pd.DataFrame) -> None:
    failures = []
    if len(df) < MIN_ROWS:
        failures.append(f"row count {len(df)} < {MIN_ROWS} — partial upstream load")
    if (unknown := (~df["plan"].isin(KNOWN_PLANS)).mean()) > 0.01:
        failures.append(f"unseen plan categories at {unknown:.3f}")
    if abs(df["label"].mean() - BASE_RATE) > 0.05:
        failures.append(f"label base rate {df['label'].mean():.3f} vs {BASE_RATE:.3f}")
    if failures:
        raise DataValidationError("; ".join(failures))

The rule is that validation fails the run. Emitting a warning into a dashboard nobody reads is how a model gets trained on a half-loaded partition and shipped.

The subtle checks are the valuable ones: a shifted label base rate usually means the label-generation join changed; a spike in unseen categories means an upstream enum grew. Both train cleanly and degrade in production.

Registry and Promotion

The training pipeline registers; it does not promote. Separating them is what keeps a bad Tuesday-night run from reaching users.

result = evaluate(model, holdout)
champion = registry.get_model("churn", alias="production")
champion_metrics = evaluate(champion, holdout)      # SAME holdout, both models

if result.auc < champion_metrics.auc - REGRESSION_TOLERANCE:
    raise PipelineFailed(f"AUC regression: {result.auc:.4f} vs {champion_metrics.auc:.4f}")

for segment, m in result.by_segment.items():        # aggregate wins can hide harm
    if m.auc < champion_metrics.by_segment[segment].auc - SEGMENT_TOLERANCE:
        raise PipelineFailed(f"segment regression in {segment}")

registry.register("churn", model, stage="staging", fingerprint=fp.run_key())

Per-segment comparison is the check most pipelines lack. A model can gain 0.004 AUC overall while losing badly on a small but important cohort.

Feature Definitions vs Materialization

One definition, two materializations, is the structural answer to training/serving skew (covered operationally in the ML deployment skill).

The offline path must be point-in-time correct: joining a label at time t to feature values as of t, not as of now. Naively joining a current feature table to historical labels is the single most common cause of an offline AUC that does not survive contact with production.

# ✅ as-of join: each label row gets the feature value that existed at its timestamp
training_df = store.get_historical_features(
    entity_df=labels[["user_id", "label_timestamp", "label"]],
    features=["orders_30d", "account_age_days"],
)

Orchestration and Triggers

Trigger Fires on Good for Watch out for
Schedule Cron Stable, slow-moving domains Burns compute when nothing changed
Data arrival New partition landed Batch pipelines Needs a reliable completion signal
Drift alarm Monitoring signal Fast-moving distributions Retraining on drifted data can bake in the drift
Code merge Feature-library change Any pipeline Requires a fast, trustworthy eval gate
Manual Human decision High-stakes models Becomes the bottleneck

Retraining on a drift alarm deserves care: if the drift is a genuine distribution change, retraining is right; if it is an upstream data bug, retraining hard-codes the bug into the model. Gate automatic retraining on validation passing first.

Common Anti-Patterns

❌ Versioning the code but not the data. ✅ Record the dataset URI and manifest hash in every run; snapshot mutable sources as stage one.

❌ Reading SELECT * FROM events at training time. ✅ Read an immutable, date-partitioned snapshot so old runs stay reproducible.

❌ A single monolithic train.py that ingests, transforms, trains, and uploads. ✅ Discrete stages with explicit inputs, outputs, and content-hash caching.

❌ Random train/test split regenerated per run. ✅ Deterministic hash-bucket split, fixed across runs — and split by time whenever the target is temporally ordered, since a random split leaks the future.

❌ Data quality checks that log a warning. ✅ Checks that raise and fail the run before compute is spent.

❌ The training job promoting its own output to production. ✅ Register to staging; promotion is a separate, gated decision.

❌ Absolute metric thresholds as the quality gate. ✅ Compare against the live champion on the identical holdout, overall and per segment.

❌ Feature transforms defined inside the training script. ✅ A shared feature library or feature store with point-in-time-correct historical retrieval.

❌ Hyperparameters edited in the source and lost after the run. ✅ Resolved config logged as run parameters, defaults included.

MLOps Pipeline Checklist

  • Every run records git sha, dataset URI, dataset hash, params, and env hash
  • Mutable sources snapshotted immutably as the first stage
  • Pipeline decomposed into independently re-runnable, cacheable stages
  • Data validation raises and blocks the run
  • Row-count, null-rate, category-cardinality, and label-base-rate checks in place
  • Holdout split deterministic and stable across runs; time-based where the target is temporally ordered
  • Historical feature retrieval is point-in-time correct
  • Feature definitions shared between training and serving
  • Evaluation compares against the live champion on the same holdout
  • Per-segment regression check, not just an aggregate metric
  • Training registers to staging; promotion is a separate decision
  • Retraining triggers defined beyond a bare cron schedule
  • Automatic retraining gated on data validation passing
  • Any past model rebuildable from its recorded fingerprint alone