Skip to main content

Frequently Asked Questions

Common questions about getting started, data, training, deployment, monitoring, and collaboration.

Getting Started​

Q: Do I need to learn a new framework to use Michelangelo?

A: No. If you're using the UI, it's entirely point-and-click. If you're coding, Michelangelo uses familiar tools:

  • Python for model code (PyTorch, TensorFlow, scikit-learn, XGBoost all work)
  • Ray for distributed computing
  • Standard data formats (Parquet, CSV, JSON)
  • Decorators (@uniflow.task, @uniflow.workflow) to integrate your existing code

Q: Can I use my existing Python ML code?

A: Yes! Wrap your training functions with @uniflow.task() decorator and you're ready to go. Example:

from michelangelo.uniflow.plugins.ray import RayTask

@uniflow.task(config=RayTask(head_cpu=2, head_memory="4Gi"))
def train_model(data_path: str):
    # Your existing training code here
    model = train_my_model(data_path)
    return model

Q: How do I migrate from my current ML stack?

A: Start small:

  1. Pick one model to migrate (not your most critical one)
  2. Use Michelangelo's data prep → training → deployment workflow
  3. Compare results with your existing pipeline
  4. Gradually migrate more models as you gain confidence

Data & Features​

Q: Where does my training data come from?

A: Multiple sources:

  • Upload CSV/Parquet files directly to the plugged-in storage
  • Connect to data warehouses (Snowflake, BigQuery, Redshift)
  • Use Spark/Ray for large-scale data processing
  • Reference existing datasets in Michelangelo's data catalog

Q: Can I use feature stores with Michelangelo?

A: Yes, Michelangelo integrates with feature stores or you can manage features within the platform using the data prep pipelines and inference.

Q: What data formats are supported?

A: Parquet (recommended). CSV, JSON, Avro. For custom formats, use Uniflow tasks to handle data loading.

Training & Deployment​

Q: What compute resources are available?

A: Michelangelo provides:

  • CPU-only instances for lightweight models
  • Single-GPU instances (V100, A100) for deep learning
  • Multi-GPU clusters for distributed training
  • Ray clusters for data-parallel and model-parallel training

Q: How long does it take to deploy a model?

A: Deployment time varies:

  • Online inference: ~5-10 minutes (container build + rollout)
  • Batch predictions: Immediate (scheduled jobs)
  • Testing in sandbox: <2 minutes

Q: Can I do A/B testing?

A: Yes. Deploy multiple model versions to the same endpoint with traffic splitting. Monitor metrics per variant and gradually shift traffic to the winner.

Q: What happens if my model training fails?

A: Uniflow automatically:

  • Retries transient failures (network issues, spot instance preemption)
  • Preserves logs and intermediate outputs for debugging
  • Sends notifications (email, Slack) on terminal state — see Pipeline Notifications

Monitoring & Operations​

Q: How do I monitor model performance in production?

A: Michelangelo provides:

  • Model Excellence Scores tracking accuracy, latency, throughput
  • Data drift detection comparing training vs. production distributions
  • Custom metrics you define and track
  • Alerts when metrics degrade beyond thresholds

Q: Can I roll back a model deployment?

A: Every deployment is versioned (current_revision and candidate_revision). For automatic rollback on a failed rollout, set with_rollback_trigger on a BlastUpdate deployment strategy — the controller will revert to the last healthy revision if health gates fail. To roll back manually today, update the Deployment resource to point at a prior revision; there isn't a dedicated rollback CLI command yet.

Q: How do I debug predictions?

A: Multiple approaches:

  • Request tracing: See exact features used for a specific prediction
  • Batch debugging: Run model on test inputs via UI
  • Local testing: Pull deployed model and run locally with same inputs

Cost & Scaling​

Q: How does Michelangelo handle scaling?

A: Automatically:

  • Online inference autoscales based on request volume
  • Batch jobs use spot instances for cost savings
  • Ray clusters elastically scale workers based on workload

Q: What if my dataset doesn't fit in memory?

A: Use Michelangelo's Ray integration for out-of-core processing. Data is streamed from storage (S3, HDFS) and processed in chunks.

Collaboration & Governance​

Q: How do multiple team members collaborate?

A: Michelangelo provides:

  • Shared projects with role-based access control (see Project Management)
  • Model lineage tracking from training data through deployment
  • Versioning for models, pipelines, and pipeline runs
  • Notifications to keep teammates informed of pipeline events (see Notifications)

For operator-side RBAC and identity-provider setup, see the Authentication guide.

Q: Is my data secure?

A: Yes. Michelangelo enforces:

  • Role-based access control (RBAC)
  • Encryption at rest and in transit
  • Audit logs for all operations
  • Audit logging and controls to support SOC 2, GDPR, and HIPAA compliance

See the Compliance Guide for configuration steps specific to each framework.

Q: Can I use Michelangelo for regulated industries (healthcare, finance)?

A: Yes, with proper configuration. Michelangelo supports:

  • Data residency requirements (region-specific storage)
  • Audit trails for model decisions
  • Explainability tools for model interpretability

What's next?​