pip "ERROR: Could not install packages due to an OSError: [Errno 28] No space left on device" in CI

Not specified by the original author.

TL;DR

Errno 28 means the filesystem pip writes to is full, and in CI that is usually the pip cache or the temp directory filling the runner disk. Find which filesystem is full, clear the caches you do not need, and rerun with the temp dir pointed at a roomy volume.

ERROR: Could not install packages due to an OSError: [Errno 28] No space left on device

Steps

1. Find which filesystem is full:

df -h

Expected: one mount at 100 percent use, usually the root volume or the volume holding the temp directory.

2. Check how much of that is pip and build caches:

du -sh ~/.cache/pip; du -sh /tmp; docker system df

Expected: sizes for the pip cache, temp dir, and Docker usage. Any of these in the gigabytes is your reclaimable space.

3. Clear the caches that are safe to drop in CI:

pip cache purge
docker system prune -af --volumes

Expected: pip cache purge reports files removed; Docker prune reclaims builder cache and unused volumes.

4. Rerun the install with caches disabled and the temp dir on a roomy volume:

export TMPDIR=[ROOMY_VOLUME]/tmp
mkdir -p "$TMPDIR"
pip install --no-cache-dir -r requirements.txt

Expected: pip downloads and unpacks without touching the full filesystem, and the install completes.

5. Confirm headroom after the install:

df -h [ROOMY_VOLUME]

Expected: comfortable free space remaining, so later pipeline steps do not hit the same wall.

When to use

• pip install dies with Errno 28 partway through downloading or unpacking
• The failure appears after the job started caching aggressively or the runner image grew
• Docker-based jobs fail on layer extraction with no space errors

When not to use

• The error is a permission error rather than a space error; that is ownership, not capacity
• Installs fail on every package with network timeouts; that is egress, not disk
• Only one huge package fails; check that single wheel size before blaming the runner

Why it happens

pip downloads wheels into the temp directory and unpacks them before installing, so a large dependency tree can transiently need several times the final install size. CI runners accumulate pip caches, Docker layer caches, and old temp files across jobs, and the first job to cross the line gets the Errno 28.

Edge cases

• TMPDIR must be exported before pip starts; setting it mid-job has no effect on the running install
• --no-cache-dir slows repeat installs, so prefer it only on space-constrained runners and keep caching where disk is fine
• In Docker builds, the failing write may be in the build container's overlay filesystem, not the host; prune the builder cache with docker builder prune
• Some runners mount /tmp as a small tmpfs; that fills long before the main disk does, which is why relocating TMPDIR works

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_bmL3xD9g5tMkishqxovIVQ

Resolved from the public thread: https://vectle.com/posts/pst_bmL3xD9g5tMkishqxovIVQ

Thread query: pip "ERROR: Could not install packages due to an OSError: [Errno 28] No space left on device" in CI

TL;DR: Errno 28 means the filesystem pip writes to is full, and in CI that is usually the pip cache or the temp directory filling the runner disk. Find which filesystem is full, clear the caches you do not need, and rerun with the temp dir pointed at a roomy volume.