Technology
Python
Data pipelines, ML training and inference services — typed, tested and packaged like production software rather than notebooks in a trench coat.
The ecosystem is the argument
There is no serious alternative for data and ML work, so the question is never whether to use Python — it is whether the Python in production is held to the same standard as the rest of the estate.
We apply type hints, dependency pinning, packaging and test coverage to data code specifically because it is the code most likely to fail silently and be discovered a quarter later.
Practice
How we use it
Typed and linted
Type hints with static checking, because a data pipeline that fails loudly at the boundary beats one that produces plausible nonsense.
Pipeline engineering
Orchestrated, idempotent, restartable jobs with data-quality assertions at each stage.
ML serving
Inference services with the same latency budgets, health checks and rollback path as any other production service.
Reproducibility
Pinned environments and versioned data, so a training run from six months ago can be reproduced exactly.
Judgement
When Python is the right call
And when it is not. A technology page that only lists strengths is a brochure.
Reach for it when
- Anything touching data engineering, ML or scientific computing
- Inference services where the model ecosystem is Python-native
- Rapid prototyping ahead of a decision on production language
Look elsewhere when
- High-concurrency network services where Go or Rust fit better
- Latency floors below a few milliseconds
Sectors