LGTM, though we should probably add this to our tests for postgres and myqsl. 😉

Hi, this commit introduced a behaviour change that is biting me.

I used .sh files to change the database and thus create several ones with files like 001prod1.sh 001.prod1.sql 002.prod2.sh 002.prod2.sql, where 000.sh would set a MAIN_POSTGRES_USER=${POSTGRES_USER} etc, and 001.prod1.sh would be something like:

psql -v ON_ERROR_STOP=1 --username "$MAIN_POSTGRES_USER" <<-EOSQL
    CREATE USER ${PROD1_DB_USER} WITH PASSWORD '${PROD1_DB_PASSWORD}';
    CREATE DATABASE ${PROD1_DB};
    GRANT ALL PRIVILEGES ON DATABASE ${PROD1_DB} TO ${PROD1_DB_USER};
export POSTGRES_USER=${PROD1_DB_USER}
export POSTGRES_PASSWORD=${PROD1_DB_PASSWORD}
export POSTGRES_DB=${PROD1_DB}

Now you are evaluating the line

psql+=( --username "$POSTGRES_USER" --dbname "$POSTGRES_DB" )

out of the loop, and thus all DBs use the original user. I found the former behaviour more flexible and helped me to minimise containers while having separate DBs. Do you think it would make sense a patch putting this line back inside the loop, where it would restore the previous behaviour?

If you are already running psql commands in your shell scripts, why not rename your sql files (so that the entrypoint does not run them) and just import them yourself? psql --username ${PROD1_DB_USER} ... < 001-prod1-sql

That or just export on top of the psql variable that we use; though I am not sure if that may change like this has.