lance-format/lance

Lance is a mature, infrastructure-grade project with strong adoption signals (6.2k stars, 621 forks, active velocity of 0.33/hr suggests ongoing development). It solves a concrete pain point—100x faster random access vs Parquet for AI workloads—with a specialization that transcends simple algorithmic contribution. The defensibility stems from: (1) network effects through deep integrations with ecosystem leaders (DuckDB, Polars, PyArrow, PyTorch), creating data gravity and switching costs; (2) the format itself is a commodity once adopted—users build on Lance, not around it; (3) the 1370-day history indicates battle-tested stability and community trust. Frontier risk is medium because: OpenAI/Anthropic/Google have less incentive to fork or reimplement a storage format (it's infrastructure commodity), but they could integrate Lance support or build competing vector-first formats if their internal needs diverge. However, Lance's polish, ecosystem alignment, and open governance make it more likely they'd adopt than compete. The novelty is novel_combination—vector indexing + columnar storage + versioning + AI-specific optimizations are known ideas, but Lance's specific design (Arrow-compatible, distributed, format-versioned) for this exact use case is non-obvious. Production implementation depth confirmed by 621 forks and real-world usage.