The container is regarded as a single unit; if a server inside the container fails the functions of that server are offloaded to another available server and it is taken out of service.
Once enough servers in a container are offline the entire unit has all computational load offloaded to another, identical container with sufficient capacity.
Then the now-offline unit is retrieved and serviced; probably a ground-up rebuild of all components.
… but I do like the idea of some dude in a wetsuit trying to replace a memory stick.