How to clean GCR cache?

[tejal29]

A kaniko user asked via email ,

any suggestions on maintaining the cache in GCR? Storage is cheap, sure, but we make thousands of images a month for our CI/CD pipeline and that will add up quickly. Are there plans to include some sort of cleanup routines in the Kaniko image to manage the cache?

[raijinsetsu]

I think the following is true:
Every tagged image references the layers/images in the cache in their manifest (ie. the images we want to keep around). The repository prevents you from deleting those layers as they are "in use".

Ignoring HOW tagged images might be cleaned up (next step), I think it is safe to assume that we could enumerate all the images in the cache that were created more than some time period in the past and delete them, ignoring error responses that indicated it could not be deleted because it's "in use". This would remove any dangling cache images (ie. incomplete builds) as well as cache images that are no longer referenced by a tagged image.

I meant to add: if Kaniko could do this piece, then we would not have to copy the cache configuration to other tools or keep the two in-sync.

[tejal29]

This was also brought up in GoogleContainerTools/skaffold#3487

[tejal29]

I think the following is true:
Every tagged image references the layers/images in the cache in their manifest (ie. the images we want to keep around). The repository prevents you from deleting those layers as they are "in use".

Ignoring HOW tagged images might be cleaned up (next step), I think it is safe to assume that we could enumerate all the images in the cache that were created more than some time period in the past and delete them, ignoring error responses that indicated it could not be deleted because it's "in use". This would remove any dangling cache images (ie. incomplete builds) as well as cache images that are no longer referenced by a tagged image.

@raijinsetsu do you have a script to do that or a program ?

[raijinsetsu]

I have something... but there's a bug in the tagged-image deletion logic. So, just ignore that piece. I cannot send the entire script due to proprietary code, but here is the part of it that handles cache maintenance.

# transformListGcrImageTags
#       Reads "<digest> <tag/timestamp> [timestamp]" from stdin
#               and outputs "<digest> <timestamp> [tag]"
function transformListGcrImageTags() {

        local digest tag timestamp

        while read digest tag timestamp ; do

                if [[ "$digest" == "DIGEST" ]]; then
                        # ignore the header line
                        continue
                fi

                if [[ -z "$timestamp" ]]; then
                        # tag is actually the timestamp
                        echo $digest ${tag}Z
                else
                        echo $digest ${timestamp}Z $tag
                fi
        done
}

function listGcrImageTags() {

        if [[ -n "${2-}" ]]; then
                gcloud container images list-tags --filter "$2" "$1" | transformListGcrImageTags
        else
                gcloud container images list-tags "$1" | transformListGcrImageTags
        fi
}

# filterBranchImages <image> <git origin>
#       filters the images: if it has a remote branch, remove the tag
#       receives input from STDIN
#       example:
#               gcloud container images list-tags ${gcr_root}/rest-server/dev | filterBranchImages ${gcr_root}/rest-server/dev ${remote_origin}
function filterBranchImages() {

        local digest tag timestamp

        # read one line into fields
        while read digest timestamp tag ; do
        
                if [[ -n "$tag" ]]; then
                        # tagged image

                        # strip off the trailing commit hash from the tag to get the branch
                        local b=${tag%-*}
                        if ! (remoteBranchExists "branch" "$b" "$2" || remoteBranchExists "branch" "feat/$b" "$2" || remoteBranchExists "branch" "fix/$b" "$2" || remoteBranchExists "branch" "chore/$b" "$2" ) ; then
                                # the remote branch does not exist
                                echo "${1}:${tag}"
                        fi
                else
                        # untagged image - cannot determine remote branch
                        echo "${1}@sha256:${digest}"
                fi
        done
}

function filterCacheImages() {

        local digest tag timestamp

        while read digest timestamp tag ; do
                if [[ -n "$tag" ]]; then
                        echo "${1}:${tag}"
                else
                        echo "${1}@sha256:${digest}"
                fi
        done
}

function deleteTags() {

        # use xargs to batch up multiple deletions
        # also suppress stdout
        xargs -r gcloud container images delete > /dev/null
}

function cleanupGcrImages() {

        local expireTS=$((now - 1209600))
        local expire=$( date -Iseconds -d \@$expireTS )
        echo "Cache expiration: $expire"

        if [[ $TEST -eq 0 ]]; then
                listGcrImageTags ${gcr_root}/rest-server/dev | filterBranchImages ${gcr_root}/rest-server/dev ${remote_origin} | tee >(cat >&2) | deleteTags
                listGcrImageTags ${gcr_root}/rest-server/dev/cache "timestamp.datetime < $expire" | filterCacheImages ${gcr_root}/rest-server/dev/cache ${remote_origin} | tee >(cat >&2) | deleteTags
        else
                listGcrImageTags ${gcr_root}/rest-server/dev | filterBranchImages ${gcr_root}/rest-server/dev ${remote_origin}
                listGcrImageTags ${gcr_root}/rest-server/dev/cache "timestamp.datetime < $expire" | filterCacheImages ${gcr_root}/rest-server/dev/cache ${remote_origin}
        fi
}

It's supposed to conditionally delete the tagged image based on the presence of the corresponding branch in Git but we ran into an edge case where that is not true.

I think it would be great if the Cache cleanup piece were part of Kaniko but I understand if this is TOO repository specific.

[aaron-prindle]

Wanted to note that GCP's Artifact Registry has now added the concept of a "Cleanup Policy" which could help here:
https://cloud.google.com/artifact-registry/docs/repositories/cleanup-policy

[aaron-prindle]

Possible dupe of #998