ML Drift Monitoring Setup (Data Drift, Concept Drift)

ML Drift Monitoring Setup (Data Drift, Concept Drift)

Your ML model in production showed ROC‑AUC 0.92, but last month it dropped to 0.87. Business is complaining about reduced recommendation quality. What went wrong? Likely data drift or concept drift. We've been configuring drift monitoring for production ML systems for 5+ years — over 50 projects for fintech, e‑commerce, and ad platforms. Monitoring detects changes early and prevents model degradation before business metrics suffer.

Types of Drift

Data drift (covariate shift) — change in input feature distributions. The model sees data different from its training set. Example: seasonal shifts in buying behavior altering the distribution of "average time between purchases".

Concept drift — change in the relationship between features and the target. Example: fraud patterns evolve, and features that once reliably predicted fraud lose predictive power.

Label drift — change in the distribution of the target variable. Example: the proportion of positive examples in binary classification changes significantly.

Prediction drift — change in the distribution of model predictions. Can be monitored without labeled data.

Why Drift Monitoring Is Critical for Production ML

According to our statistics, 70% of models in production degrade within 6 months of deployment. Teams usually notice problems after 2 weeks — when business metrics have already dropped 5‑10%. Drift monitoring with properly configured alerts reduces detection time to hours. We guarantee that after implementing our solutions, you'll receive a drift notification no later than 15 minutes after the change begins.

Statistical Tests for Drift Detection

Monitoring Tools: Which to Choose?

Evidently AI — open‑source library for generating drift reports with rich visualization. Great for detailed analysis but has higher overhead. Whylogs / WhyLabs — lightweight library for logging statistical profiles in real time; minimal overhead on production inference but requires more manual dashboard setup. Arize AI, Fiddler, Arthur — commercial platforms with ready‑made dashboards and alerts, but high cost.

Tool Comparison:

How to Set Up Alerts

We integrate drift metrics into your existing monitoring stack. For example, using Prometheus and Grafana:

# Integration with Grafana Alerting
def compute_psi(expected, actual, buckets=10):
    expected_hist, _ = np.histogram(expected, bins=buckets, density=True)
    actual_hist, _ = np.histogram(actual, bins=buckets, density=True)
    # Smooth to avoid division by zero
    expected_hist = np.where(expected_hist == 0, 1e-6, expected_hist)
    actual_hist = np.where(actual_hist == 0, 1e-6, actual_hist)
    psi = np.sum((actual_hist - expected_hist) * np.log(actual_hist / expected_hist))
    return psi

# Export to Prometheus
psi_value = compute_psi(reference_feature, production_feature)
prometheus_client.Gauge('model_feature_psi', 'PSI for feature X').set(psi_value)

Alerts are configured in Grafana: PSI > 0.2 — warning, PSI > 0.25 — critical with notification in Slack/PagerDuty. We recommend multi‑threshold alerts to avoid noise.

Monitoring Without Ground Truth

A classic problem: in production, ground truth arrives with a delay or never. Without labeled data, you can monitor:

• Prediction drift — change in prediction distribution
• Feature drift — change in input feature distributions
• Confidence distribution — change in model confidence
• Business proxy metrics — e.g., CTR as proxy for recommendation quality

What's Included in Drift Monitoring Setup?

When you order this service, you get:

• Audit of your current pipeline and identification of critical points
• Selection of the best tool for your stack (Evidently AI, Whylogs, Grafana)
• Integration of drift metrics into your existing infrastructure
• Setup of alerts and dashboards in Grafana (Slack/PagerDuty)
• Runbook documentation with step‑by‑step response plan
• Team training on monitoring usage

We do the work turnkey — from analysis to deployment. Timeline: 5 to 10 business days depending on system complexity.

Drift Response Process

Upon drift detection: analyze changes in data, decide on retraining or feature engineering fixes; for concept drift, architectural changes may be needed. Monitoring without a response process is useless — it's essential to describe a runbook for each alert type in advance. We include this runbook in our deliverables.

Case Study: Quick Detection in Fintech

On a fraud detection project for a fintech company, we set up Whylogs + Grafana monitoring. Within the first month, an alert triggered on a feature with PSI of 0.3 — the model's predictions were starting to shift. Our team analyzed the change within two hours, identified that a new payment method was altering transaction patterns, and retrained the model. The client avoided a 10% increase in false positives. Previously, this would have been noticed only after a week of customer complaints.

We’ll assess your project for free — contact us, and we’ll choose a solution that fits your budget and timeline.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.