Migrating from cloud VMs

Python version, every pinned dependency, and any system package your code relies on. Undocumented environments are what make migrations drag.

List where data is read from and written to today. Local filesystem paths need a replacement; object storage paths usually carry over unchanged.

Shell steps, cron entries, and compiled binaries in your pipeline need explicit handling rather than an assumption they will come along.

Install the same dependencies and confirm your existing code runs unchanged on a small input before touching anything else.

Replace local paths with object storage or workspace-persistent locations, so every worker can reach the same data.

Run the same input through both and compare outputs. Do this before distributing anything, because debugging correctness and parallelism at once is twice the work.

Get it running correctly on one worker first. Only then wrap the per-unit work in a distributed function.

Once the migrated pipeline has run clean a few times, shut the VMs down. Parallel infrastructure kept "just in case" is a standing bill.