Haystack 2.0: the story of a rewrite [PyCon DE & PyData Berlin 2024]

Haystack 2.0 underwent a major rewrite to improve modularity and reduce dependencies, making installation faster and lighter

The pipeline architecture was transformed from a linear “tube” structure to a directed acyclic graph, enabling more complex workflows and decision paths

Components were made smaller and more focused, receiving only necessary inputs rather than bulky data structures, improving testability and customization

Dependencies were split between core and integrations repositories, making maintenance easier and preventing dependency conflicts

Lazy imports were implemented for optional dependencies, providing better user experience with clear error messages when additional packages are needed

The rewrite focused on reducing technical debt while maintaining backward compatibility through migration guides and documentation

Integration development became 4x easier, leading to many new integrations with services like Elasticsearch, Weaviate, and PyCon

Setting firm deadlines for rewrites is crucial to avoid perfectionism and ensure actual releases

Community feedback and use cases should drive rewrite decisions, not just technical improvements

Documentation, DevRel support, and migration tools are as important as the code changes in a successful rewrite