supabase/postgres

PGDATA forced to default location

Closed this issue · 6 comments

chevdor commented

Bug report

  • I confirm this is a bug with Supabase, not with my own application.
  • I confirm I have searched the Docs, GitHub Discussions, and Discord.

Describe the bug

The image forces PGDATA to be the default location: /var/lib/postgresql/data and fail it is overriden.
If the user defines PGDATA to another location, it is ignored and the container will fail starting.

Here is an output with PGDATA=/var/lib/postgresql/data/pgdata (NOTE the extra non standard /pgdata at the end):

 2025-05-27 09:00:49.140 UTC [1] FATAL:  "/var/lib/postgresql/data" is not a valid data directory

To Reproduce

You can test using docker and:

docker run --rm -it \
    -e PGDATA=/data \
    -e POSTGRES_PASSWORD=password \
     supabase/postgres:17.4.1.036

It will fail mentioning:

 2025-05-27 09:25:05.745 UTC [49] FATAL:  data directory "/var/lib/postgresql/data" does not exist

which should not be relevant since we want to use /data.

Expected behavior

The user can point to a custom PGDATA and the container starts.
Using the original postgres image works fine:

docker run --rm -it \
    -e PGDATA=/data \
    -e POSTGRES_PASSWORD=password \
     postgres:latest

System information

  • version: supabase-postgres:17.4.1.036
    but the issue is not specific to this version.
chevdor commented 
docker run --rm -it \
    -e PGDATA=/data \
    -e POSTGRES_PASSWORD=password \
     supabase/postgres:17.4.1.036

=> FAIL

docker run --rm -it \
    -e PGDATA=/data \
    -e POSTGRES_PASSWORD=password \
     supabase/postgres:17.4.1.036 postgres

=> WORKS

Provinding postgres effectively "removes" the surprising default CMD and its "-D", "/etc/postgresql".

Why would /etc/postgresql be passed as data folder ?

singh-inder commented

Just read through the Dockerfile, so I might be missing something.

All the configs are copied from ansible/files to /etc/postgresql. https://github.com/supabase/postgres/blob/develop/Dockerfile-15#L162

And in this postgresql.conf file which is copied to the image, data_directory is hard coded to /var/lib/postgresql/data.
In the image CMD ["postgres", "-D", "/etc/postgresql"], so even when you specify PGDATA env var postgres is still looking in /var/lib/postgresql/data

docker run --rm -it
-e PGDATA=/data
-e POSTGRES_PASSWORD=password
supabase/postgres:17.4.1.036 postgres
=> WORKS

I'm assuming that when you override the image args, you setup postgres container with default postgres config so it works but you lose the supabase customized config.

chevdor commented

The issue I see is that if a user provides PGDATA=/foo/bar it is not expected that postgres uses /etc/postgresql as datadir and I did not spot the reason why it reports missing /var/lib/postgresql/data.

But that's an interesting finding and that means the issue is how the postgres.conf is generated/pathed.

I feel it would make more sense to bring whatever config into the PGDATA specified by the user when the DB is initialized but I am surely missing something.

singh-inder commented

and I did not spot the reason why it reports missing /var/lib/postgresql/data

Let me clarify.

postgres entrypoint is docker-entrypoint.sh. db directory is created by calling docker_create_db_directories. This function is imported from the upstream postgres repo's docker-entrypoint.sh

If you look at docker_create_db_directories signature, it runs command mkdir -p $PGDATA. This will create directory as specified in PGDATA env. By default, this is /var/lib/postgresql/data. If you override it (e.g., with PGDATA: /data), then only /data will be created.

Now, in supabase postgres repo's Dockerfile the CMD is set to ["postgres", "-D", "/etc/postgresql"].
postgres loads config from /etc/postgresql/postgresql.conf. Inside that file, data_directory is still set to /var/lib/postgresql/data, which hasn’t been created - because PGDATA pointed elsewhere (/data), and that’s the only directory the entrypoint script created. Thus /var/lib/postgresql/data is missing.

The simplest solution is to modify the image’s entrypoint to update the data_directory inside postgresql.conf dynamically

Here is an example. I just used sed to modify the data_directory in postgresql.conf inside container and everything works fine. I had v15 image downloaded on system. Will work the same with v17

services:
  db:
    image: supabase/postgres:15.8.1.060
    ports:
      - 5432:5432
    entrypoint: /bin/bash -c "
        sed -i \"s|data_directory.*|data_directory = '/data'|\" /etc/postgresql/postgresql.conf && \
        exec docker-entrypoint.sh postgres -D /etc/postgresql
     "
    environment:
      POSTGRES_PASSWORD: password
      PGDATA: /data
    volumes:
      - ./data:/data
  • 👍1
samrose commented

Closing. PR's welcome if anyone has a proposition for desired changes.