Avoiding redundant container builds
The Docker cache largely works as intended, except on 1.9.0 where it sometimes isn’t invalidated correctly, however there are a few corner cases where it is not sufficient and builds end up doing a lot of redundant work.
Consider for instance the case of packaging hundreds of megabytes of assets in a container image. Even when the content of the Dockerfile is unchanged and every layer is a cache hit, the entire build context needs to be uploaded to the docker daemon, which can take a frustratingly long time.
Another situation in which the docker cache falls short is when a slightly non-standard build process is used, as for instance with gockerize, where the final image is built inside a container and the intermediate step of compiling the Go sources runs even if the output binary is unchanged.
In both cases, the age-old approach of comparing input and output timestamps to determine whether to perform the build step could safely shorten build times.
A rough solution
Dealing with timestamps on the command line is notoriously tricky, especially when attemting to maintain some semblance of portability. To keep things simple I figured it would be simpler to work with Unix timestamps.
Step 1: When was the build context last modified?
Surprisingly enough, there appears to be no simple standard way to obtain the modification time of a file as a Unix timestamp. The most reliable way I’ve found so far is to use
stat and customize the invocation based on the host OS.
if [[ $(uname -s) == "Darwin" ]] ; then stat_format="-f %m" elif [[ $(uname -s) == "Linux" ]] ; then stat_format="-c %Y" else echo "unsupported platform: always rebuild" 1>&2 return fi
With this out of the way, it is quite straightforward to derive the most recent modification timestamp in the build context:
changed=$(find "$1" -type f | xargs stat $stat_format | sort -nr | head -n 1)
Step 2: When was the docker image created?
Obtaining the creation timestamp of a docker image is easy enough but converting it to a Unix timestamp is somewhat more involved. After some experimentation, I settled on querying the docker remote API inside a container and extracting the timestamp with jq, which has the dual benefit of portability and not introducing a new dependency.
docker build -t test-newer - &>/dev/null <<EOF FROM alpine:3.3 RUN apk -U add curl jq EOF created=$(docker run --rm -v /var/run/docker.sock:/var/run/docker.sock \ test-newer sh -c "curl --fail --unix-socket /var/run/docker.sock \ http:/images/\$(echo $1 | jq -R -r @uri)/json 2>/dev/null \ | jq -r '.Created[0:19]+\"Z\" | fromdate'")
Using containers to avoid dealing with OS-specific variations in basic Unix tools, and outdated or missing dependencies is a neat trick that I’ve started using a lot lately.
Step 3: Comparing
Combining the output of the first two steps is pretty straightforward:
if [[ -n "$changed" ]] && [[ -n "$created" ]] && (( "$created" > "$changed" )) ; then echo "fresh" else echo "stale" fi
Although this first solution gets the job done, it lacks elegance and is not as efficient as it could be since it needs to sort the list of timestamps for the entire build context.
Looking at the man page for
find reveals the existence of the promising
-newermt filter. Instead of computing the newest timestamp in the build context one can directly test if any file is newer than a given timestamp, neatly avoiding the expensive sort and allowing the pipe to be closed early:
if [[ -z "$(find "$context" -newermt "$created" | head -n 1)" ]] ; then echo "fresh" else echo "stale" fi
The catch is that
find does not allow dates to be specified as Unix timestamps. There is GNU extension allowing it but that wouldn’t work on OSX, which leads us to revisit an early assumption. Specifically, dealing with Unix timestamps, which was expected to make things simple, turns out to create more issues than it solves.
Obtaining the creation timestamp of the docker image in a human-readable (and thus find-friendly) format results is a much simpler script:
created=$(docker inspect --format='' --type=image "$1" |\ cut -d. -f1 | sed 's/T/ /') if [[ -n "$created" ]] && [[ -z "$(TZ=UTC find "$2" -newermt "$created" | head -n 1)" ]] ; then echo "fresh" else echo "stale" fi
TZ=UTC above. Abandoning Unix timestamps means we have to be careful to compare date/times in the same time zone. The docker daemon sanely gives UTC timestamps so we need to make sure
find doesn’t mistakenly assume it to be in the local timezone.
A cleaned-up version of this script is available on github
Room for improvement
This approach is not quite optimal as docker will not update the creation timestamp of the image when a build hits the cache for all layers. This could lead to a situation where the timestamp check always flag the image as stale if a file in the build context sees its timestamp change in a way that does not invalidate the docker cache (unchanged content, .dockerignore, …)
A possible workaround, which may be acceptable if the build process is expensive but the final image small, as is the case for gockerize, is to disable the docker cache when building the final image.