Containers and Environments for Reproducibility
Capturing the compiler, libraries, and system so a result runs the same on another machine years later.
The Environment Problem
A result depends not only on the code but on everything around it: the compiler, the numerical libraries, the operating system, the versions of every dependency. Change any of these and the result can shift or the code can fail to build. Capturing this environment is essential for reproducibility, because a reproduction attempt that uses a different library is not testing the science, it is testing the library.
Ways to Capture It
- Dependency manifests, listing exact versions of every library so the environment can be recreated.
- Environment managers, which build an isolated set of dependencies from a specification.
- Containers, which package the code together with its entire software environment into a single runnable image.
- Virtual machines, for capturing the operating system as well, at higher cost.
Containers
A container image bundles the code, its libraries, and system tools into a self-contained unit that runs identically wherever the container runtime is available. This solves the works-on-my-machine problem: the same image runs on a laptop, a cluster, and a colleague's workstation, giving the same environment each time. For reproducibility, an archived container image is a durable capsule of the computational environment.
Limits and Honesty
Containers pin software but not hardware. Different processors and instruction sets can still perturb floating-point results at the round-off level, so a container guarantees reproduction within tolerance more reliably than bit-for-bit reproduction. The container image itself must be archived and versioned, because a recipe that pulls the latest dependencies at build time is not reproducible; only a frozen, stored image is.
A cold-reproduction runbook, of the kind Kronos maintains, is strongest when paired with a pinned environment: a fresh machine plus the archived environment regenerates the result with no hidden state, which is the concrete meaning of a reproducible computational deposit.