MLOps Plan — OfertaFlash SaaS

Estado actual vs objetivos

p95 Latencia actual

400ms

⚠ Objetivo: < 80ms

Error rate estimado

~0.3%

⚡ Objetivo: < 0.05%

Coste LLM mensual

~580€

💸 Objetivo: < 200€/mes

Drift detection

—

✗ Sin monitorización

Pipeline de despliegue propuesto

Export
ONNX

✓ Listo

Docker
FastAPI

→ Sprint 1

MLflow
Registry

→ Sprint 1

Staging
K8s

Sprint 2

Canary
5%

Sprint 2

Monitor
Drift

Sprint 3

RAG
Chatbot

Sprint 4

LLM
Cost Ctrl

Sprint 4

🐳

Container — FastAPI + XGBoost

        Dockerfile
FROM python:3.11-slim

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY model/xgb_reco_v3.onnx /app/model/
COPY src/ /app/src/

ENV MODEL_PATH=/app/model/xgb_reco_v3.onnx
ENV REDIS_URL=redis://redis:6379

HEALTHCHECK --interval=30s --timeout=3s \
  CMD curl -f http://localhost:8080/health || exit 1

EXPOSE 8080
CMD ["uvicorn", "src.server:app", \
     "--host", "0.0.0.0", "--port", "8080", \
     "--workers", "4"]

🗄️

Feature Store — Feast + PostgreSQL

        Python
from feast import Entity, FeatureView

user = Entity("user_id", value_type=INT64)

reco_features = FeatureView(
    name="ofertaflash_user_signals",
    entities=["user_id"],
    ttl=timedelta(days=1),
    features=[
        Feature("clicks_30d",    INT64),
        Feature("cart_abandon_rate",FLOAT),
        Feature("avg_price_viewed", FLOAT),
        Feature("pref_category",   STRING),
    ],
    online=True,  # Redis-backed online store
    source=pg_source
)

# Materialize: cada hora via Airflow DAG
store.materialize_incremental(
    end_date=datetime.now()
)

📉

Monitorización de Drift — KS Test

        Python
from scipy.stats import ks_2samp
import mlflow

def monitor_input_drift(ref_window, live_window):
    features = ["clicks_30d", "avg_price_viewed",
                "cart_abandon_rate"]
    alerts = []
    for feat in features:
        stat, p = ks_2samp(ref_window[feat], live_window[feat])
        psi = compute_psi(ref_window[feat], live_window[feat])
        mlflow.log_metric(f"psi_{feat}", psi)
        if psi > 0.2:    # Critical threshold
            alerts.append({"feature": feat, "psi": psi,
                           "severity": "CRITICAL"})
        elif psi > 0.1: # Warning threshold
            alerts.append({"feature": feat, "psi": psi,
                           "severity": "WARNING"})
    return alerts

Feature	PSI Simulado	Estado	Acción
clicks_30d	0.04	OK	—
avg_price_viewed	0.12	WARNING	Evaluar retrain
cart_abandon_rate	0.23	CRITICAL	Retrain inmediato
pref_category	0.07	OK	—

🔍

RAG Chatbot — Soporte de Producto

        Python
from qdrant_client import QdrantClient
from anthropic import Anthropic

client = QdrantClient(host="qdrant", port=6333)
llm = Anthropic()

def answer_support_query(query: str, user_id: str):
    # 1. Embed query → retrieve top-5 docs
    hits = client.search(
        collection_name="ofertaflash_docs",
        query_vector=embed(query),
        limit=5,
        score_threshold=0.75
    )
    context = "\n\n".join([h.payload["text"] for h in hits])
    # 2. Call LLM with context (Haiku: 5x cheaper)
    resp = llm.messages.create(
        model="claude-haiku-4-5",
        max_tokens=512,
        messages=[{"role":"user",
                   "content": PROMPT.format(ctx=context, q=query)}]
    )
    track_cost(resp.usage)  # ← cost control
    return resp.content[0].text

Qdrant (self-hosted) text-embedding-3-small Chunk: 512 tok / 50 overlap Haiku 4.5 (coste reducido 5x)

🚨

Umbrales de Alerta

Métrica	Warning	Critical
p95 latencia	> 80ms	> 150ms
Error rate	> 0.05%	> 0.5%
PSI drift	> 0.10	> 0.20
Accuracy drop	> 2%	> 5%
Coste LLM/día	> 6€	> 10€
Cache hit rate	< 40%	< 20%

🔄

Triggers de Reentrenamiento

Trigger	Detección	Acción
Semanal	Cron Lunes 02:00	Retrain completo
PSI > 0.2	Monitor 6h	Retrain urgente
Accuracy < 78%	Evaluación diaria	Retrain + A/B
+50K registros	Conteo incremental	Fine-tune
Black Friday	Calendario manual	Pre-train especial

💰

Control de Costes LLM

Proveedor/Modelo	Input	Output
Claude 3 Opus	$15/1M	$75/1M
→ Actual (sin ctrl)	~580€/mes
Claude Haiku 4.5	$0.25/1M	$1.25/1M
→ Propuesto + caché	~95€/mes

Ahorro estimado: 485€/mes

Migrar a Haiku + prompt caching + response cache Redis 24h

📅

Roadmap del Proyecto (8 semanas)

Semana 1-2 · Sprint 1

Docker + FastAPI + MLflow

Containerizar modelo XGBoost, exponer endpoint /predict, configurar MLflow tracking + model registry. CI/CD GitHub Actions → ECR.

Semana 3-4 · Sprint 2

K8s Staging + Canary Deploy

EKS cluster (3 nodos), manifiestos HPA (2-8 réplicas), canary 5% con NGINX Ingress. Target: p95 < 80ms validado en staging.

Semana 5-6 · Sprint 3

Drift Monitoring + Feature Store

Feast online store (Redis), dashboard Grafana con PSI por feature, alertas PagerDuty, pipeline reentrenamiento automático con Airflow.

Semana 7-8 · Sprint 4

RAG Chatbot + Cost Control

Qdrant self-hosted, ingestión docs soporte (1.200 páginas), migración a Claude Haiku, response caching Redis, budget tracker Slack alert.

☸️

Kubernetes HPA Manifest

        YAML · k8s
apiVersion: apps/v1
kind: Deployment
metadata:
  name: ofertaflash-reco
  labels: {app: reco, version: v3}
spec:
  replicas: 3
  selector:
    matchLabels: {app: reco}
  template:
    spec:
      containers:
      - name: reco
        image: 123456.dkr.ecr.eu-west-1.amazonaws.com/reco:v3.2.1
        resources:
          requests: {cpu: "500m", memory: "1Gi"}
          limits:   {cpu: "2",    memory: "3Gi"}
        readinessProbe:
          httpGet: {path: /health, port: 8080}
---
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
spec:
  minReplicas: 2
  maxReplicas: 8   # Pico viernes: 8K req/h
  metrics:
  - type: Resource
    resource:
      name: cpu
      target: {averageUtilization: 60}

Proyección de rendimiento post-migración

62ms

p95 latencia

0.03%

error rate

95€

coste LLM/mes

🗃️

Comparativa Vector Databases

Database	Hosting	Latencia	Coste	Recomendación
Qdrant	Self-hosted	Muy baja	$0/mes	✓ Elegido
Pinecone	Managed	Baja	~$70/mes	Backup
Weaviate	Ambos	Baja	Variable	Alternativa
pgvector	Self-hosted	Media	$0/mes	Si <100K docs
Chroma	Self-hosted	Baja	$0/mes	Solo dev

🛠️

Stack Técnico Final Propuesto

Capa	Herramienta	Justificación
Serving	FastAPI + Uvicorn	REST, 4 workers
Orquestación	EKS (K8s)	HPA automático
Experiments	MLflow	Open source, S3 backend
Features	Feast + Redis	Online <10ms
Monitoring	Grafana + PagerDuty	Alertas en <5min
Vector DB	Qdrant	Self-hosted, $0
LLM	Claude Haiku 4.5	5x más barato
Cache	Redis 24h TTL	Hit rate objetivo 55%

📊

Resumen ROI del Proyecto

-84%

Reducción latencia p95
(400ms → 62ms)

-84%

Reducción coste LLM
(580€ → 95€/mes)

< 24h

Alert drift antes de
degradación visible

Auto

Reentrenamiento
sin intervención manual

8 sem

Tiempo a producción
completo