How to Identify and Reduce an 18 GB Docker Image Layer During docker push

Problem background : After launching a container from an existing image and compiling a project inside it, the author tried to create a new image with docker commit and push it with docker push. Both commands were very slow, and the push displayed a layer of 18.1 GB.

Solution workflow

1. Locate the 18.1 GB layer

Get the writable layer directory (UpperDir) of the container named daft-test:

UPPER=$(docker inspect -f '{{index .GraphDriver.Data "UpperDir"}}' daft-test)
 echo "$UPPER"

List the size of top‑level directories inside that UpperDir:

du -xhd1 "$UPPER" 2>/dev/null | sort -h

The output (shown in the first screenshot) reveals that /tmp, /home and /usr each occupy several gigabytes.

Drill down into a specific directory, for example /tmp:

du -xhd2 "$UPPER/tmp" 2>/dev/null | sort -h | tail -30

The second screenshot shows many leftover Ray job files that consume a large amount of space; deleting them (and similarly cleaning other large directories) reduces the layer size.

2. Additional inspection commands

List the 50 largest files:

find "$UPPER" -xdev -type f -printf '%s\t%p
' 2>/dev/null |
  sort -nr |
  head -50 |
  numfmt --field=1 --to=iec

List files larger than 500 MB:

find "$UPPER" -xdev -type f -size +500M -printf '%s\t%p
' 2>/dev/null |
  sort -nr |
  numfmt --field=1 --to=iec

Show what files have changed in the container:

docker diff daft-test > /tmp/daft-test.diff
head -100 /tmp/daft-test.diff
wc -l /tmp/daft-test.diff

The diff output uses A for added files, C for modified files, and D for deleted files.

3. Common reasons for oversized layers

Rust target/
Maven ~/.m2/repository
Gradle ~/.gradle
pip, uv, conda caches
Downloaded source packages, wheels, tar.gz
Debug/release intermediate files from builds
Log or test data
Modifying a large directory in the base image, triggering overlay2 copy‑up

Especially the last case: if the original image contains a 5 GB directory and only part of it is changed, overlay2 may copy the whole directory into a new layer, creating another 5 GB layer.

By cleaning the identified large directories and files, the image size drops dramatically, and both docker commit and docker push run much faster.