ML Systems & MLOps

Read these first

1. 01Feedback Loops and Self-Fulfilling Predictions

   A model that acts on its own predictions can end up changing the very world it is trying to predict, sometimes proving itself right for the wrong reason, sometimes destroying the pattern it was built to exploit.

2. 02Label Latency and Delayed Feedback

   A model's prediction is instant, but knowing whether it was right often is not, and that gap between prediction and ground truth shapes almost everything about how a live ML system can be monitored and retrained.

3. 03ML Technical Debt and Hidden Coupling

   Machine learning systems accumulate a distinctive kind of technical debt: changing one feature, one label definition, or one upstream data source can silently break a model in ways ordinary software's error messages never catch.

4. 04Model Serving Architectures

   How a trained model actually gets its prediction to a trading system, batch, online, or streaming, is a design decision with real consequences for latency, cost, and how stale a prediction is allowed to be.

5. 05Training-Serving Skew

   A model can perform beautifully in backtest and then quietly underperform live, not because the model is wrong but because the data it sees in production is computed differently from the data it trained on.

Then the rest

1. 06Automated Retraining Pipelines
2. 07Batch vs Online Inference
3. 08Canary Releases for Models
4. 09Data Drift vs Concept Drift
5. 10Estimating Live Performance Without Labels
6. 11Inference Latency Budgeting
7. 12Model Quantization for Inference
8. 13Online Experimentation for Model Rollout
9. 14Pipeline Orchestration and DAGs
10. 15Point-in-Time Correctness in Feature Pipelines
11. 16Prediction Drift Monitoring
12. 17Pruning and Model Compression
13. 18Rollback and Model Kill Criteria
14. 19Streaming Feature Computation
15. 20The Population Stability Index
16. 21Training Cost and Compute Budgeting