A notebook works on one cluster but fails on another because dependency versions differ. Which practice improves repeatability?

1. Treat a table's owner as the library version
2. Depend only on a user's local package cache
3. Pin compatible runtime and library versions and validate the execution environment
4. Install arbitrary latest versions on every run without testing