PostgreSQL in Openshift won't execute the entrypoint and can not start the database - postgresql

Question

We have a read-only Postgresql database that should run in Openshift cluster.
We are using RHEL as the undrlying operating system.
Our Dockerfile will install postgres software, create the database instance, loads the data to it than shuts the database down and save the image.
We are using only bash and sql scripts and deploy the database using flyway.
When starting the container the entrypoint script will simply startup the database instance using "pg_ctl" command then perform an endless loop to keep the container running.
The Dockerfile has as the last command USER 26, where 26 is the id of the postgres user. The entrypoint script can be started as the postgres user or by a sudo user.
Everything is working well in Docker.
In Openshift the container is started by a different user belonging to the root group, but not the root user nor the user 26. Actually Openshift ignores the USER 26 clause in the Dockerfile.
The user starting the container (we'll call it containeruser) has anyhow no rights to start the postgres instance , so when running the entrypoint it will get permission denied on the postgresql data folder.
I have tried different options, adding the containeruser user to the wheel group and modify the sudoers file to allow it using sudo and start the entrypoint as postgres user but with no success.
So I have my database image ready but can not start it in Openshift.
On the openshift configuration side we are not allowed to make changes like allowing sudo usage, or starting the container as root or postgres user.
Any idea or help to this problem?
I am not an Openshift expert.
Thank you!
Best regards,
rimetnac

Answer 1

You have two choices.
The preferred choice is to fix your image so that it can run as any user. For this, do not use the existing postgres user. Create a new user, where that user has group root. Then ensure that all directories/files that PostgreSQL needs to write to are owned by that user, but also have group root and writable by group. When the container is then started up, it will run as an assigned user ID, not in /etc/passwd, and so will fallback to using group root still. Because the directories/file are writable to group root, everything will still work. For more information see:
https://docs.openshift.org/latest/creating_images/guidelines.html
Specifically, section 'Support Arbitrary User IDs'.
The second option if you have admin control of the cluster, and your security team do not object that you are overriding the default security model, is to allow your image to be run as user ID it wants to.
First create a new service account:
oc create serviceaccount runasnonroot
Next grant that service account the ability to run as non root user ID of its choosing.
oc adm policy add-scc-to-user nonroot -z runasnonroot --as system:admin
Then patch the deployment config to use that service account.
oc patch dc/mydatabase --patch '{"spec":{"template":{"spec":{"serviceAccountName": "runasnonroot"}}}}'
Note that this still requires you use USER in the image with an integer user ID and not postgres. Otherwise it can't verify it will run as non root user. That is because if you use a user name instead of user ID, you could be maliciously mapping that to root.

Answer 2

I spent days figuring this out and found one good solution.
OpenShift Origin runs an image as a user created by it, as explained in this OpenShift blog post. This prevents programs from being able to access needed files and directories. To successfully run a program on OpenShift Origin, the blog post provides two solutions, however, the first will not work for PostgreSQL and the second has two disadvantages (explained in the notes):
Grant group write access to the directories used by the main program.
This will not solve the problem because, although the PostgreSQL files will be accessible by any program, they must be owned by the owner of the PostgreSQL process.
Ensure that when operating system libraries are used to look up a system user, one is returned for the ID of the user OpenShift Origin runs the image as. The following are two methods for doing this:
Use a package called nss_wrapper, "which intercepts any calls which look up details of a user and returns a valid entry."
Make the UNIX password database file (/etc/passwd) have global write permissions in the image build so that the OpenShift user can be added to it in the S2I run script.
Each option has a disadvantage: 1. install an extra package and 2. make user accounts insecure.
The best solution is to build the docker image to run as the user OpenShift Origin will run the image as. I built this instructional image with it.
One additional problem to note is that, as the owner of the PostgreSQL process must be the owner of the files and directories accessed by PostgreSQL, PostgreSQL must be set up (i.e. initdb, roles, databases, etc.) during the image build. This is because file ownership can only be changed during the image build and the ownership of the files must be changed after PostgreSQL has been set up for the reason explained in #2 below.
Here are the complete steps with notes for setting up PostgreSQL in the image build:
Manually create the PostgreSQL data directory and change its ownership to a non-root user that will be used to initialze PostgreSQL and set up the components (e.g. roles and databases) required to run the server on OpenShift Origin.
This is required because the "initdb" executable must be executed by a user other than root and will need access to the data directory. Additionally, this user cannot be the user OpenShift Origin will run the image as because it is not in the system.
Switch to the non-root user.
This is required because the initdb executable must be executed "as the user that will own the server process, because the server needs to have access to the files and directories that initdb creates" (PostgreSQL documentation) and because the PostgreSQL server will be started to set up components (e.g. roles and databases) required to run the server on OpenShift Origin.
Run the "initdb" executable.
Start the PostgreSQL server, set up the required components (roles, databases, etc.) and stop the PostgreSQL server.
Switch back to the root user.
Change the ownership of the PostgreSQL files and directories to the user OpenShift Origin will run the image as.
Edit (06/20/18): I have found that there is a solution to set up PostgreSQL after the image is built. The user OpenShift Origin will run the image as can be added to the system at the start of the build. This will allow PostgreSQL to be set up and the ownership of its files and directories to be changed after the image build.

Answer 3

After gathering the comments from all contributors I can asnwer my question as follows:
Option 1
When you create the postgres database during image build, you must configure openshift policies to allow starting your container as the user that created the database during image build. Use this option when the database must be filled with data and this operation takes much time making it inappropriate for a container start. the entrypoint will only start the already prepared database.
Option 2
Create your database when starting the container using the entrypoint script. Use this option when the database creation is fast enough to be done at container start.
Option 3
See the last comment from Adrian which seems to answer all the problems anyhow I didn't got the time to test it.
Thank you all for your contributions.