What we Eval in the Shadows
summary
Study of 240,327 R scripts (15,401 packages) records 49,296,059 eval calls and exercises ~18k eval call-sites, showing eval is pervasive and varied. Only 6% of calls have side-effects originating from 6.3% of sites, producing 7.9M writes (66% updates, 31% new definitions, 3% removals), and 43 functions cause 90% of side-effects. Most expressions are function calls (40%) or variable lookups (27%), reflection is the common source, and 82% use caller’s environment, making many effects local but some non-modular and brittle.
AI-written summary · about