Error deploying datadelivery when ui version doesn't change

[johnbradley]

Error when deploying a new D4S2 version without changing datadelivery-ui:

TASK [build_docker_image : Clone the Git repo at the provided version] 
**************************************************************
...: FAILED! => {"changed": false, 
"cmd": "/bin/git clone --origin origin https://github.com/Duke-GCB/datadelivery-ui.git /tmp/build-docker-image-src/datadelivery-ui-v1.0.0", 
"msg": "fatal: destination path '/tmp/build-docker-image-src/datadelivery-ui-v1.0.0' already exists and is not an empty directory.", 
"rc": 128, 
"stderr": "fatal: destination path '/tmp/build-docker-image-src/datadelivery-ui-v1.0.0' already exists and is not an empty directory.\n", 
"stderr_lines": ["fatal: destination path '/tmp/build-docker-image-src/datadelivery-ui-v1.0.0' already exists and is not an empty directory."], "stdout": "", "stdout_lines": []}

The error occurs when building the ui image via https://github.com/Duke-GCB/gcb-ansible-roles/tree/master/build_docker_image.

[dleehr]

Good catch. Looks like this is the same underlying technical issue as #103 - when the requested version (tag or branch) is already checked out on the destination server, the git module step fails because it already exists. We probably need to remedy that in the role.

[dleehr]

The reason/requirement for this naming convention (git directory checked out to {{ local_repo_root }}/{{ image_name }}-{{ version }}) was primarily to make the checkouts (and therefore docker build contexts) unique based on the version (tag/branch/SHA).

gcb-ansible-roles/build_docker_image/tasks/main.yml

Lines 1 to 21 in 15107fa

	- name: Ensure directory exists for cloning git repo
	file:
	state: directory
	path: "{{ local_repo_root }}/{{ image_name }}-{{ version }}"
	# Note that this checks out the git repo to a specific local directory based on the
	# provided image_name and version. it does not infer a subdirectory name from the
	# remote git repo like the default git mode. We must use a unique name here so that
	# one host can support distinct image builds
	- name: Clone the Git repo at the provided version
	git:
	repo: "{{ repo_url }}"
	dest: "{{ local_repo_root }}/{{ image_name }}-{{ version }}"
	version: "{{ version }}"
	- name: Build a docker image
	docker_image:
	path: "{{ local_repo_root }}/{{ image_name }}-{{ version }}"
	name: "{{ image_name }}"
	tag: "{{ version }}"

While this works in simple cases, there are times where it falls down:

1. Deploying multiple services on a host where one does not change - the git clone step fails when the directory exists and is not empty
2. Attempting to deploy an update from a branch, where the HEAD changes but the name of the branch doesn't - the git clone step fails when the directory exists and is not empty.

One simple solution would be to always generate a truly unique temporary directory, clone there, build from there, and delete it at the end. That would sidestep the git errors, but there's still an oddity with the docker_image step. If the image:tag already exists locally, docker_image won't build a new one unless we set force to yes.

We don't have force turned, because I only wanted to build images that needed it (e.g. a new version number). We could get a little smarter about detecting changes to git, registering those, and then deciding what to do with docker, but I'm hesitant to complicate this further.

[dleehr]

I'm also sure there are established practices in various CI/build systems that address this.