Squash instructions in Dockerfile

[jackcaron]

Right now, the squash feature is experimental and also squash all layers into a single one.

My suggestion is to add instructions in a Dockerfile on what to squash.

Ex.: if I have a set of instructions like this:

RUN wget -qO- "https://mystorage.com/some_debs_files.tar.xz" | tar xJf - -C . \
  && apt-get -qq update && apt-get -qq install -y \
  libasound2 libatk-bridge2.0-0 libatk1.0-0 libatspi2.0-0 libcups2 libnspr4 \
  libnss3 libxcomposite1 libxdamage1 libxkbcommon0 libxrandr2 libxtst6 xdg-utils \
  && dpkg -i package1.deb \
  package2.deb \
  package3.deb \
  && rm ./package* \
  && rm -rf /var/lib/apt/lists/*

RUN more stuff

A situation like this can easy gets unreadable and harder to change.

But with possible squash instructions, it could look like this:

SQUASH start

RUN wget -qO- "https://mystorage.com/some_debs_files.tar.xz" | tar xJf - -C .

RUN apt-get -qq update && apt-get -qq install -y \
  libasound2 libatk-bridge2.0-0 libatk1.0-0 libatspi2.0-0 libcups2 libnspr4 \
  libnss3 libxcomposite1 libxdamage1 libxkbcommon0 libxrandr2 libxtst6 xdg-utils

RUN dpkg -i package1.deb \
  package2.deb \
  package3.deb

RUN m ./package* \
  && rm -rf /var/lib/apt/lists/*

SQUASH end

RUN more stuff

This is easier to read, debug, and modify. The SQUASH end would only be executed if it's reached.

Now I'm guessing that the one thing that could cause problems here is the fact that SQUASH start and SQUASH end wouldn't act as layers per se.

Possible behaviors:

• SQUASH start with no SQUASH end: squash all layers after SQUASH start
• SQUASH end with no SQUASH start: error
• two SQUASH start without any SQUASH end between them: error
• SQUASH start without a SQUASH end wouldn't impact other images using this one as a base
• If the SQUASH end is not reached because some instructions failed, the work after the SQUASH start is cached for faster rebuilds while debugging

[AkihiroSuda]

Isn't this enough?

FROM ubuntu
RUN ...
RUN ...
RUN ...

FROM scratch
COPY --from=0 / /

(Metadata like ENV, LABEL are lost though)