The Hidden Technical Debt of ML

ML code is the tip of the iceberg

Google's classic paper Hidden Technical Debt in Machine Learning Systems makes one uncomfortable point: the model code is a tiny box in a sprawling diagram of configuration, data collection, feature extraction, serving infra, and monitoring. The debt lives in the plumbing, not the model.

Name the anti-patterns so your team can call them in review:

• Glue code — endless adapters around a vendor library.
• Pipeline jungles — scrapes and joins accreted until no one dares change them.
• Dead experimental codepaths — if EXPERIMENT_42: branches that never got cleaned up.
• Configuration debt — hundreds of flags, half undocumented.
• CACE — Changing Anything Changes Everything: a model's inputs are entangled, so tweaking one feature silently shifts the decision boundary everywhere.

Unlike normal tech debt, ML debt is invisible at the code level — it hides in data dependencies and feedback loops a linter can't see.

An ML system is an iceberg. The model is the glinting tip above the water; the nine-tenths below — data pipelines, configs, monitoring — is what sinks ships. Teams that only inspect the tip hit the rest at speed.