MLOps game mode (board)

[pshenok]

Add an MLOps-themed game mode that teaches machine learning infrastructure concepts using the same tower defense mechanics.

Concept:
Instead of web traffic, players handle ML workloads - training jobs, inference requests, data pipelines. Build and scale ML infrastructure to survive increasing demand.

Potential services:

Service	Function	Equivalent
Feature Store	Stores/serves features	S3-like
Training Cluster	Processes training jobs	EC2/GPU
Model Registry	Stores trained models	Database
Inference Endpoint	Serves predictions	API compute
Data Pipeline	ETL/preprocessing	WAF-like filter
GPU Pool	Expensive compute	Premium EC2
Model Cache	Reduces inference latency	CDN

Traffic types:

• 🟢 Inference requests (need endpoint + model)
• 🟠 Training jobs (need GPU + data + registry)
• 🟣 Bad data / adversarial inputs (need validation)

MLOps-specific mechanics:

• Model versioning (deploy v1 vs v2)
• A/B testing traffic splits
• GPU costs (expensive but fast)
• Cold start for models (loading time)
• Batch vs real-time inference
• Training queue management
• Data drift detection

Educational value:
Teaches concepts like:

• Why feature stores exist
• Model serving at scale
• Training vs inference trade-offs
• Cost management for GPU workloads
• ML pipeline orchestration

Implementation approach:

• Could be separate game mode selectable from main menu
• Or entirely separate "board" with different service set
• Reuse core mechanics (traffic, queues, connections)
• New visuals/colors for ML theme

Open questions:

• Separate mode or integrated with cloud services?
• How to visualize model training (takes longer than request processing)?
• Should models "degrade" over time (concept drift)?

Priority: Low (fun expansion after core game is solid)

Discussion:
Would love input from ML engineers - what concepts would be most valuable to teach? What's confusing for newcomers to MLOps?