Gradual Soundness: Lessons from Static Python

Context: Gradually-typed languages allow typed and untyped code to interoperate, but typically come with significant drawbacks. In some languages, the types are unreliable; in others, communication across type boundaries can be extremely expensive; still others only limited forms of interoperability. The research community is actively seeking a sound, fast, expressive approach gradual typing. Inquiry: This paper describes Static Python, language developed by engineers at Instagram that has proven itself reasonably production. Python's essentially programmer-tunable combination concrete transient approaches from literature. Concrete provide full soundness low performance overhead, impose nonlocal constraints. Transient sound shallow sense easier use; they help bridge gap between code. Approach: We evaluate its current state develop model captures essence types. draw upon personal communication, bug reports, Python regression test suite this model. Knowledge: Our main finding arises mix an effective way lower maintenance cost approach. also find method-based JIT technology eliminate costs On more technical level, two contributions: evaluation Python. process formalization found several errors implementation, including fatal errors. Grounding: implemented PLT Redex tested using property-based tests 265 suite. includes small core convey ideas soundness. claims based on production experience web server. Migrations server have caused 3.7\% increase requests handled per second maximum CPU load. Importance: first whose piece-meal application realistic codebase consistently improved performance. Other designers may wish replicate approach, especially those who currently maintain unsound path