acoustic-carpenter-78188
04/23/2026, 6:41 PMpreload stage in docker/devbox-bundled/Dockerfile pulls, tags, and saves the container images listed in images/manifest.txt one at a time. With 8 images in the manifest, cold builds spend most of the stage's wall time waiting on serial registry round-trips even though the operations are independent.
What changes were proposed in this pull request?
Rewrite docker/devbox-bundled/images/preload to run `podman pull`/`tag`/`save` for every manifest entry concurrently and then wait for them all:
• All entries launch in parallel as background jobs; the script exits non-zero if any job fails.
• Each image's stdout/stderr is buffered to a per-image .log file and only emitted on failure, so successful runs stay readable and failed ones clearly identify which image broke.
• Behavior and output artifacts (per-image .tar files under images/) are unchanged.
Net effect: preload runtime drops from roughly the sum of per-image pull times to roughly the max.
How was this patch tested?
• sh -n docker/devbox-bundled/images/preload — syntax check passes.
• Local make build in docker/devbox-bundled/ completes successfully with a visibly shorter preload stage.
Labels
changed
Check all the applicable boxes
• All commits are signed-off.
• main
• #6583
• perf(devbox): parallelize image preload in devbox-bundled build 👈
flyteorg/flyte