Skip to content

Backend Overview

Tracelet supports multiple experiment tracking backends, allowing you to choose the platform that best fits your needs.

Supported Backends

Backend Type Hosting Best For
MLflow Open Source Self/Cloud Traditional ML workflows
ClearML Enterprise/SaaS SaaS/Self Enterprise MLOps
Weights & Biases SaaS/Open Source SaaS/Self Deep learning research
AIM Open Source Self Lightweight tracking

Choosing a Backend

MLflow

  • ✅ Best for: Traditional ML, production deployments
  • ✅ Strengths: Model registry, serving, mature ecosystem
  • ❌ Limitations: Basic visualization, manual setup

ClearML

  • ✅ Best for: Enterprise teams, automated pipelines
  • ✅ Strengths: Rich UI, automatic logging, pipeline orchestration
  • ❌ Limitations: Complex setup, resource intensive

Weights & Biases

  • ✅ Best for: Deep learning research, collaboration
  • ✅ Strengths: Best-in-class visualization, sharing, reports
  • ❌ Limitations: SaaS dependency, pricing for teams

AIM

  • ✅ Best for: Simple tracking, local development
  • ✅ Strengths: Lightweight, fast queries, local-first
  • ❌ Limitations: Fewer features, smaller ecosystem

Backend Comparison

Feature Matrix

Feature MLflow ClearML W&B AIM
Metrics Logging ✅ ✅ ✅ ✅
Hyperparameters ✅ ✅ ✅ ✅
Artifacts ✅ ✅ ✅ ⚠️
Model Registry ✅ ✅ ✅ ❌
Visualizations ⚠️ ✅ ✅ ✅
Collaboration ⚠️ ✅ ✅ ⚠️
Auto-logging ⚠️ ✅ ✅ ❌
Pipeline Orchestration ❌ ✅ ⚠️ ❌
Self-hosting ✅ ✅ ✅ ✅
Free Tier ✅ ✅ ✅ ✅

Legend: ✅ Full support, ⚠️ Limited support, ❌ Not supported

Multi-Backend Support

Use multiple backends simultaneously:

import tracelet

# Log to both MLflow and W&B
tracelet.start_logging(
    backend=["mlflow", "wandb"],
    exp_name="multi_backend_experiment",
    project="comparison_study"
)

# All metrics go to both platforms
writer = SummaryWriter()
writer.add_scalar("loss", 0.5, 1)  # → MLflow + W&B

Benefits:

  • Backup: Redundant logging prevents data loss
  • Comparison: Evaluate different platform features
  • Migration: Gradual transition between platforms
  • Team preferences: Support different tool preferences

Performance Comparison

Typical overhead per logged metric:

Backend Latency Memory Notes
MLflow ~5ms Low Local file-based
ClearML ~15ms Medium Rich automatic logging
W&B ~20ms Medium Network-dependent
AIM ~2ms Low Optimized for speed

Getting Started

  1. Choose your backend based on your needs
  2. Install dependencies for your chosen backend
  3. Configure authentication if using hosted services
  4. Start logging with a simple example

Quick setup links:

Migration Between Backends

Switching backends is easy with Tracelet:

# Change from MLflow to W&B
# tracelet.start_logging(backend="mlflow")  # Old
tracelet.start_logging(backend="wandb")     # New

# Your existing TensorBoard code remains unchanged!
writer.add_scalar("loss", loss_value, step)

See our Migration Guide for detailed instructions.

Next Steps