Allow "initdb.d" scripts to be executed instead of sourced

[tianon]

Closes #450

cc @willglynn (is this similar to the workaround you've applied in your own scripts?)

[willglynn]

@tianon Well… given an executable /docker-entrypoint-initdb.d/restore.sh containing:

#!/bin/sh -e
gunzip < /something | pg_restore

My expectation was that this would run in its own shell with set -e. This wasn't (and isn't) the case. My issue in #450 was caused by the shared environment changing to include set +o pipefail.

My workaround was to move this script to restore.sh.real and make restore.sh say:

#!/bin/sh -e
/docker-entrypoint-initdb.d/restore.sh.real

This runs my script in a separate shell, as expected, allowing the entrypoint script to change however it wants without affecting my restore script's operation. I'd rather it work that way out of the box, of course.

Compatibility is important, so I don't think the postgres images can just stop .ing scripts and start trying to execute them instead. Checking the provided shell script with test -x might be a suitable compromise: anyone who sets +x shouldn't be surprised when their script gets executed, and if that's a problem, they could mark it -x to revert back to the dot behavior.

[tianon]

Fair point, seems like a sane compromise. PR updated.

(One caveat is that this will mean any files bind-mounted from Windows will be executed instead of sourced.)

[yosifkit]

This seems sane and I am working to update the docs. Do we want to add the same functionality to MariaDB, MySQL, and Percona? They have had pipefail since Jan 2016 (docker-library/mysql#134, MariaDB/mariadb-docker#37, docker-library/percona#12).

[tianon]

Sure, seems sane.

[yosifkit]

LGTM, docs PR created: docker-library/docs#1247