পাঠ ১১ · ৩৩-এর মধ্যে · মডিউল ২

Home / AI Courses / MLOps / Model registry

Model registry

Model registry — versioned lifecycle

৭ মিনিট পড়া মধ্য · Intermediate Concept + Python

এই পাঠে যা শিখবেন

Model registry-এর core concept — registered model, version, alias
Lifecycle stages ও promotion workflow
MLflow Model Registry API — register, transition, load
Governance ও audit trail patterns

১ · কেন registry

একটি team-এ ৫ data scientist, প্রত্যেকে নিজের laptop-এ model save করছেন। Production-এ deploy-এর সময় — "কোন model? কোন version? কে validated?"। Registry এই questions-এর single source of truth।

২ · Core concepts

Registry-এর তিনটি concept

Registered Model: a logical model name (e.g., "bangla-sentiment")।
Version: immutable artifact — version 1, 2, 3...।
Stage / Alias: mutable label — version 3 = Production, version 4 = Staging।

৩ · Lifecycle stages

None / Dev: just registered, no testing।
Staging: integration tested, shadow deployed।
Production: serving real traffic।
Archived: retired, kept for audit।

MLflow 2.0+ recommends aliases over stages — more flexible:

champion = current production।
challenger = candidate competing।
shadow = silent comparison।
Custom aliases for complex deployments।

৪ · MLflow Model Registry API

Python · Register & promote

import mlflow
from mlflow.tracking import MlflowClient

mlflow.set_tracking_uri("http://mlflow:5000")
client = MlflowClient()

# ১. একটি training run থেকে register
with mlflow.start_run() as run:
    mlflow.sklearn.log_model(
        model,
        artifact_path="model",
        registered_model_name="bangla-sentiment",
        signature=signature,    # input/output schema
        input_example=X_test.iloc[:5],
    )

# ২. version metadata add
client.update_model_version(
    name="bangla-sentiment",
    version=4,
    description="RF, n_est=200, AUC=0.94 on 2025-Q1 data",
)
client.set_model_version_tag("bangla-sentiment", 4, "data_version", "v3.2")
client.set_model_version_tag("bangla-sentiment", 4, "approved_by", "lead-ds")

# ৩. alias assign (modern way)
client.set_registered_model_alias(
    name="bangla-sentiment",
    alias="challenger",
    version=4,
)

# ৪. promote — challenger → champion
client.set_registered_model_alias(
    name="bangla-sentiment",
    alias="champion",
    version=4,
)

# ৫. load by alias — production code
model = mlflow.pyfunc.load_model("models:/bangla-sentiment@champion")
prediction = model.predict(input_df)

Production code-এ alias use — "champion" version automatically pulls latest blessed। Promotion = alias swap; rollback = previous version-এ alias point।

৫ · Promotion workflow

Training run → automatic register।
Smoke test pass → alias candidate।
Performance test (offline metric) → tag pass।
Shadow deployment 24 hours → metric compare।
Manual review (lead-DS, product manager)।
Canary deploy 5% traffic → 24 hours monitor।
Full rollout → alias champion।
Previous champion → alias previous (rollback ready)।

৬ · Signature & input example

Registry-তে model log করার সময় signature include — input/output schema। Production-এ schema validation, training-serving skew prevent।

Python · Signature

from mlflow.models.signature import infer_signature

# auto-infer from sample input/output
signature = infer_signature(X_train, model.predict(X_train))

mlflow.sklearn.log_model(
    model,
    artifact_path="model",
    registered_model_name="bangla-sentiment",
    signature=signature,
    input_example=X_train.iloc[:3],
)

# Production-এ load → schema enforced
model = mlflow.pyfunc.load_model("models:/bangla-sentiment@champion")
# wrong schema → ValueError
prediction = model.predict(input_df)

Schema mismatch — load time error, not silent wrong prediction। Lifesaver in CI/CD।

৭ · Registry alternatives

MLflow Model Registry (OSS): most common, self-hostable।
SageMaker Model Registry: AWS native, integrated with SageMaker training/serving।
Vertex AI Model Registry: GCP equivalent।
Azure ML Model Registry: Azure native।
Hugging Face Hub: public-facing, OSS model sharing।
Comet, Neptune, W&B: SaaS alternatives।

Registry lifecycle — version-গুলো immutable, alias-গুলো mutable। Promotion ও rollback both alias re-assignment।

Bangladesh banking-এর audit log requirement — registry built-in fit। Promotion event-গুলো registry-এ logged; regulator query-able।

ভাবনার প্রশ্ন

প্র ০১ "Promotion-এ কে control করে — DS, MLOps, business?" ভাল governance কেমন?

Promotion governance একটি collaborative process — clearly-defined roles + automated gates।

Recommended role split:

DS: training, validation metric, candidate propose।
MLOps: CI/CD gate, performance test, deployment।
Lead-DS: approve based on metrics।
Product manager: approve based on business risk।
SRE/On-call: deployment timing।

Automated gates (no human bottleneck):

Validation metric > baseline (configurable threshold)।
Schema match।
Performance regression test pass।
Fairness gate (subgroup performance parity)।
Latency < SLA।

Human gates:

Major version (e.g., new architecture) → lead approval।
Critical model (fraud, credit) → review board।
Risk classification — minor patches auto-approved, major manual।

Anti-patterns:

"Anyone can deploy" — accidents waiting।
"Only one person can deploy" — bottleneck, single point of failure।
"Human approval for every minor patch" — slow, frustration।

BD context:

Bangladesh Bank ICT guideline + SOC 2 — formal approval log।
Quarterly review board for high-risk models।

মূল উপলব্ধি: Governance = automated gates + judicious human checkpoints। Trust automation for routine; human for novel/risky। Audit log every promotion।

প্র ০২ "Rollback workflow — কীভাবে fast ও safe?"

Rollback should be the easiest operation — emergency button।

Alias-based rollback:

Champion v5 → bug found → v4-এ alias point।
Production code "champion" load → instant updated version।
Time: seconds।

Considerations:

Schema mismatch — v4 expects different feature, v5 added new। Need previous schema known।
Dependency on data version — v4 trained on v3 data; if data also changed → match।
Cache — production cache evict old prediction।

Practical rollback drill:

Quarterly — rollback simulate। Real rollback in non-prod hour।
Postmortem after each real rollback — learn।
Documentation — "rollback runbook"।

K8s integration:

Argo Rollouts — auto-rollback on metric breach।
Health check failing → previous Deployment image revert।
Combined: registry alias + K8s automation।

মূল উপলব্ধি: Rollback ease-of-use directly correlates with deployment confidence। Easy rollback → bold deployment। Painful rollback → conservative, slow iteration।

প্র ০৩ "Registry signature validation কেন important — concrete bug example।"

Signature = production safety net। Concrete example:

Bug story:

Model v5 — added new feature "user_age_group" (categorical)।
Production caller — old code, sending only old features।
Without signature: model might silently accept None, predict wrong।
With signature: load-time error — "missing column"।
CI/CD catches before production traffic।

Other catches:

Type mismatch — caller sending string, model expects int।
Range — model expects normalized [0,1], caller sending raw।
Output schema — caller expects single value, model returns multi-class।

Signature limit:

Doesn't catch semantic mismatch — same column name, different meaning।
"User_age" — int vs years vs group bucket — signature treats as int (same)।
Documentation + naming convention required।

Best practice:

Auto-infer signature from training data (mlflow infer_signature)।
Log input_example — alongside signature, "what does this look like"।
Production caller code — pydantic schema match registry signature।

মূল উপলব্ধি: Signature — fail-fast principle। Wrong input → immediate error, not subtle wrong output। Bug detection time: hours → seconds।

প্র ০৪ "Multi-region deployment — registry single source কীভাবে maintain?"

Multi-region — Bangladesh-Singapore-US — single registry replication strategy।

Pattern ১ — Single central registry:

One MLflow instance, all regions read/write।
Pros: simple, no consistency issues।
Cons: latency for distant regions, single point of failure।

Pattern ২ — Replicated registry:

Primary registry + read replicas in each region।
Promotion writes to primary; reads from local।
Pros: low read latency, better availability।
Cons: write region consistency।

Pattern ৩ — Per-region registry + sync:

Each region independent registry, automated sync।
Pros: regional independence।
Cons: drift risk, complex sync logic।

Practical Bangladesh setup:

Primary registry: Mumbai region (lowest latency from BD)।
Production model artifacts cached in each region's K8s।
"Pull at deploy time" — model fetched once, reused।

Artifact caching:

Each region S3 / GCS bucket।
Cross-region replication on artifact upload।
K8s pod-cache layer।

Consistency consideration:

"Champion alias" — globally consistent want। Eventual consistency OK seconds-delay।
For high-stakes (banking): strict consistency, slow promotion acceptable।

Bangladesh tier 1 city + tier 2:

Tier 1 (Dhaka, Chittagong): low-latency to Mumbai।
Tier 2 (Sylhet, Khulna): higher latency; local cache effective।
Service mesh with regional failover।

মূল উপলব্ধি: Multi-region — registry centralized okay, artifacts replicate। Strict global consistency rare-need; eventual consistent registry + cached artifact = practical, fast।

অনুশীলন

Register: MLflow tracking server-এ একটি model register। Alias challenger assign।

উপরের code-এ registered_model_name set, তারপর set_registered_model_alias। UI-তে "Models" tab-এ দেখুন।
Promotion: challenger → champion swap। Production code-এ "@champion" load করুন।

client.set_registered_model_alias("name", "champion", version) → load "models:/name@champion"। Always returns latest blessed।
চিন্তা: Daraz-এর recommendation model। Promotion gate-এর ৫টি criteria লিখুন।
- NDCG@10 > baseline + 1%।
- Inference latency p99 < 50ms।
- Subgroup parity (geo + demographic) within 5%।
- Shadow deployment 24h, no anomaly।
- Lead DS + product manager approval।