Erlang supervisor with one critical child

Question

We are in the process of re-organizing our applications supervision tree to make it more robustly handle failures and re-starts. However, we have a scenario where we have one parent supervisor that starts four child supervisors. The problem we have is that the first child supervisor starts several children gen_servers that must be started and initialized prior to the second child supervisor starting or it will fail.

So, I need a startup like the following:

test_app.erl -> super_supervisor -> [config_supervisor, auth_supervisor, rest_supervisor]

The trick I'm having trouble with is that config_supervisor must complete all initialization prior to auth_supervisor or rest_supervisor being started. With the rest_for_one startup strategy I get, essentially, this behavior but only by allowing auth_supervisor to fail because the needed config is not there. I would prefer to just request that config_supervisor is completed with it's initialization (which includes starting several gen_servers) prior to moving-on to auth_supervisor.

This seems like a common scenario that would have been conquered previously but, I am having a hard time "googling" a solution. Does anybody have advice or sample code that might exist to handle this scenario?

Answer 1

Did you look at the rest_for_one restart strategy? It seems that it should be covenient in this case, the middle supervisor starts the gen_servers in a defined order and last the leaf supervisor who in turn start the critical process.

Answer 2

Supervisors do a synchronous start of their children, starting each one in turn before starting the next in the order they occur in the childspeclist. So your super_supervisor should start its children in the right order, first config_supervisor, then auth_supervisor and finally rest_supervisor by having them in that order. A supervisor must (successfully) start all its children before it is considered to be started. So if config_supervisor has all the necessary processes which must be started during the initialization as its children then super_supervisor will not start the other supervisors until the config_supervisor is done.

In this case you would not need rest_for_one to ensure starting in the right order if the children are in the right order in the childspeclist.

For a worker process, gen_server/gen_fsm/gen_event, they are considered started when their init callback returns.

Have I understood your description and question correctly?

Answer 3

You may try to move config_supervisor into its own application and set the application as a requirement for the main one, in this case the config application will be started first and then the main supervisor with auth_supervisor, etc will start their initialisation.