Docker build not using layer cache

Question

Given: I want to build a Dockerfile to compile a Scala application. To speedup the build I want the dependency download to be cached.

Problem: The command ./sbt -sbt-dir ./sbt-dir -ivy ./ivy update is not cached for some reason.

FROM openjdk:8 as workspace

ARG BUILD_VERSION

WORKDIR /build

COPY ./sbt ./sbt
COPY ./sbt-dist ./sbt-dist
COPY ./build.sbt ./build.sbt
COPY ./project/build.properties ./project/build.properties
COPY ./project/plugins.sbt ./project/plugins.sbt

RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy update

COPY ./ ./

# Embedded postgres need to be run as non-root user
RUN useradd -ms /bin/bash runner
RUN chown -R runner /build
USER runner

RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy clean test
RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy docker:stage -Ddocker.image.version="${BUILD_VERSION}"

Because this build runs in a new VM all the time I push the workspace image and pull it on the next run to build the cache from it

docker build --rm=false --cache-from=workspace --build-arg BUILD_VERSION=1 -t workspace .

Here is a piece of the output

Step 2/22 : ARG BUILD_VERSION
 ---> Using cache
 ---> de98ffcfad8e
Step 3/22 : WORKDIR /build
 ---> Using cache
 ---> 253b71142240
Step 4/22 : COPY ./sbt ./sbt
 ---> Using cache
 ---> 3091fa1e1821
Step 5/22 : COPY ./sbt-dist ./sbt-dist
 ---> Using cache
 ---> f9c68659cd91
Step 6/22 : COPY ./build.sbt ./build.sbt
 ---> Using cache
 ---> d30058c451fc
Step 7/22 : COPY ./project/build.properties ./project/build.properties
 ---> Using cache
 ---> 7451eb63303f
Step 8/22 : COPY ./project/plugins.sbt ./project/plugins.sbt
 ---> Using cache
 ---> 79ac2d1e5ff5
Step 9/22 : RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy update
 ---> Running in 609104e7045e
Getting org.scala-sbt sbt 1.0.3 ...

Can anyone explain me why Docker is not using the cache here? A link to a explanation how the cache really decides when to use the cache would be ok as well. As far as I know here Docker should use the cache until the signature of the RUN command changes.

Thanks in Advance

Answer 1

I found the answer by the way :) The Dockerfile itself was ok except one problem I could then find via docker history Which shows the real shell command which are executed by docker.

The Problem was that ARG BUILD_VERSION lead to docker adds to every run command the environment variable like /bin/sh -c "ARG=123 ./sbt ...". This leads to a different call signature and a different hash every time the arg changes and therefore the run command is not applied from cache. To fix this issue just move the ARG down to the first RUN command which needs it.

FROM openjdk:8 as workspace

WORKDIR /build

COPY ./sbt ./sbt
COPY ./sbt-dist ./sbt-dist
COPY ./build.sbt ./build.sbt
COPY ./project/build.properties ./project/build.properties
COPY ./project/plugins.sbt ./project/plugins.sbt

RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy update

COPY ./ ./

# Embedded postgres need to be run as non-root user
RUN useradd -ms /bin/bash runner
RUN chown -R runner /build
USER runner

RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy clean test

ARG BUILD_VERSION
RUN ./sbt -sbt-dir ./sbt-dir -ivy ./ivy docker:stage -Ddocker.image.version="${BUILD_VERSION}"