Lesson · Hinglish · ⏱ 7 min padhne mein

🗂️ MLOps: experiment tracking aur model registries

Q: Naya model version production mein kharab perform kar raha hai. Sabse jaldi recover kaise hoga?

Model registry mein pichhle version pe roll back karna. Registry tested versions rakhti hai, isliye rollback ek step ka kaam hai.

🏠 Asli zindagi ki misaal

Achha cook naye biryani recipes try karte waqt har koshish likhta hai: chawal ka brand, masaalon ki quantity, pakane ka time aur taste kaisa tha. Notes ke bina woh sabse acchi wali kabhi dobara nahi bana paayega. Experiment tracking ML ke liye wahi notebook hai, aur model registry approved, label wali recipes ki shelf hai jo restaurant asal mein serve karta hai.

💡 Aasaan shabdon mein

MLOps machine learning ke liye DevOps hai: model training ko reproducible aur deployment ko reliable banaana. Experiment tracking tools (MLflow, Weights & Biases) har run ka code version, data version, parameters, metrics aur output model automatically log karte hain, taaki aap runs compare kar sako aur best wala dobara bana sako. Model registry approved model versions ko 'staging' aur 'production' jaise stages ke saath store karti hai, taaki deployment hamesha jaana-pehchaana, tested version uthaaye aur aap turant roll back kar sako. ML ke liye CI/CD model promote hone se pehle automated tests aur evals jodta hai.

🪜 Step by step

Har training run log karo: parameters, metrics, code commit aur data version (MLflow ya waisa tool).
Trained model aur uska preprocessing ek artifact ke roop mein save karo.
Acche models ko model registry mein version aur stage (staging, production) ke saath register karo.
Model tabhi promote karo jab woh fixed evaluation set pe current wale ko hara de.
Version se deploy karo, taaki rollback bas pichhle version ki taraf ishaara karna ho.

🔑 Zaroori shabd

Experiment run: Ek training koshish, uski settings aur results ke log ke saath.
Model registry: Approved models aur unke stages ka versioned store.
Reproducibility: Bilkul same model dobara bana paana.

👀 Example dekho

python

import mlflow
with mlflow.start_run(run_name="xgb-depth4"):
    mlflow.log_params({"max_depth": 4, "lr": 0.05, "data": "orders_2026_09"})
    model.fit(X_train, y_train)
    mlflow.log_metric("recall", 0.81)
    mlflow.sklearn.log_model(model, "model", registered_model_name="fraud")

Har run record hota hai, aur best wala registry mein jaata hai.

⚠️ Aam galtiyaan jinse bachna hai

Results notebooks aur spreadsheets mein rakhna. Baad mein koi reproduce ya compare nahi kar paata.
Runs ke beech evaluation data badalna, jisse scores compare nahi ho paate.

✅ Jaldi se check karo

Naya model version production mein kharab perform kar raha hai. Sabse jaldi recover kaise hoga?

Yaad se dobara train karna
Logs delete karna
Model registry mein pichhle version pe roll back karna

Jawaab dekho

C. Model registry mein pichhle version pe roll back karna: Registry tested versions rakhti hai, isliye rollback ek step ka kaam hai.

🙋 Seekhne wale yeh poochhte hain

Kya MLflow hi ek option hai?

Nahi. Weights & Biases, cloud tools (SageMaker, Vertex AI) aur doosre bhi aisa hi karte hain. Tool se zyaada aadat zaroori hai.

Data ka version kaise rakhun?

Dated snapshots, warehouse ya lakehouse mein table versions, ya DVC jaise tools use karo, aur har run ke saath version record karo.

📌 Yaad rakho

Har run log karo, har ship hua model register karo, aur hamesha roll back kar sako.

📖 Is lesson ke shabd (AI dictionary)

Poora course, aapke roadmap ke hisaab se

Progress tracking, quick checks aur step-by-step order ke saath. Yeh lesson in courses ka hissa hai:

ML Engineer · MLOps / AI Platform Engineer

Course mein seekho