Stripe's engineering team describes how they built a Selective Test Execution (STE) system for their 50-million-line Ruby monorepo with ~1.2 million test units. Rather than static analysis (which fails with Ruby's dynamic nature), STE uses a C++ shared library loaded via LD_PRELOAD to intercept open syscalls and record which files each test actually accesses at runtime. This file-level dependency graph is stored as roaring bitmaps in a compact selection index. When a new build starts, changed files are detected via hashdeep (including generated artifacts), mapped to bitmap indexes, and unioned to find impacted tests. The result: only ~5% of tests run on average (median 0.5%), using less than 10% of the compute of a full run. The system includes guardrails for edge cases like directory-globbing tests, previously failing tests, and linters. Build baseline metadata is stored in MongoDB with Monotonic Revision IDs for fast, reproducible selection.
