Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
118 lines (95 loc) · 3.17 KB

ARCHITECTURE.md

File metadata and controls

118 lines (95 loc) · 3.17 KB

Architecture

Overview

ModelServe is a fraud detection inference API. Client sends credit card number → FastAPI fetches features from Redis (Feast) → runs model from MLflow → returns prediction.

Components

Component Technology Purpose
API FastAPI Serves predictions at /predict
Model Registry MLflow Stores trained models, tracks experiments
Feature Store Feast + Redis Fast feature lookups at inference time
Monitoring Prometheus + Grafana Metrics, dashboards, alerts
Database PostgreSQL MLflow metadata storage
CI/CD GitHub Actions Test → Build → Deploy to AWS

Architecture Diagrams

Local Development (Docker Compose)

flowchart TB
  subgraph dc["Docker Compose network: modelserve"]
    FEAST[("Feast repo\nfeast_repo")]
    PG[(PostgreSQL :5432)]
    RD[(Redis :6379)]
    MLF[MLflow :5000]
    API[FastAPI :8000]
    PROM[Prometheus :9090]
    GRAF[Grafana :3000]

    API -->|get_online_features| RD
    API -->|FeatureStore repo| FEAST
    API -->|load_model| MLF
    MLF --> PG
    MLF -->|artifacts| VOL[mlflow_artifacts]
    PROM -->|scrape /metrics| API
    GRAF -->|query| PROM
  end

  DEV[Developer] --> API
  DEV --> GRAF
  DEV --> MLF
Loading

Prediction Request Flow

sequenceDiagram
  participant C as Client
  participant A as FastAPI
  participant F as Feast/Redis
  participant M as MLflow

  C->>A: POST /predict {entity_id}
  A->>F: get_online_features(cc_num)
  F-->>A: feature vector
  A->>M: load model, predict(features)
  M-->>A: prediction + probability
  A->>C: JSON response
Loading

CI/CD Pipeline

flowchart LR
  subgraph GHA[GitHub Actions]
    TEST[pytest]
    BUILD[docker build + push to ECR]
    DEPLOY[ssh EC2 + docker-compose up]
    TEST --> BUILD --> DEPLOY
  end

  subgraph AWS
    ECR[ECR repository]
    EC2[EC2 instance]
    BUILD --> ECR
    DEPLOY --> EC2
  end
Loading

Key Files

File Purpose
app/main.py FastAPI routes: /health, /predict, /metrics
app/model_loader.py Loads model from MLflow registry
app/feature_client.py Fetches features from Feast/Redis
app/metrics.py Prometheus metrics definitions
training/train.py Trains and registers model
feast_repo/feature_store.yaml Feast configuration
docker-compose.yml Local stack definition
.github/workflows/deploy.yml CI/CD pipeline
infrastructure/__main__.py Pulumi AWS infrastructure

Environment Variables

Variable Default Purpose
MLFLOW_TRACKING_URI http://localhost:5000 MLflow server
MODEL_NAME FraudDetector Model in registry
MODEL_STAGE Production Stage to load
FEAST_REPO_PATH ./feast_repo Feast project path
POSTGRES_USER mlflow Database user

Alert Rules

Alert Condition
ServiceDown up{job="fastapi"} == 0 for 1m
HighPredictionLatency p95 > 1s for 5m
HighErrorRate errors/requests > 5% for 2m

Related