I posed three separate questions about my favourite MLOps tools for a) data and pipeline versioning, b) experiment tracking, and c) model deployment and monitoring, supplementing the post with relevant hashtags.
I shared a short advisory note on 2022‑04‑11 15:01 UTC (attributed to Valliappa Lakshmanan) listing five situations where a feature store can be omitted.
- Known by the client.
- In a data warehouse.
- Not time dependent.
- Needed only for batch serving.
- Computationally inexpensive.
I decided to adopt data versioning alongside code versioning to make experiments reproducible and to quickly rewind data after dead‑ends.
- DVC
By versioning data I can trace model lineage, reset to a clean data snapshot when an experiment hits a dead‑end, and keep the ML workflow reproducible. This aligns with my broader push for MLOps rigor.
- Pachyderm
- DeltaLake
- Git LFS