With insights from https://github.com/docker/docker/issues/29265
the behaviour is the same as the one from docker-compose
* connect to 1 network (at random) at start
* disconnect from that network
* reconnect to all the network from the previous configuration
Load authentication credentials for available credential stores in order of preference:
1. Environment variables REPO_USER, REPO_PASS
2. Docker config files
Request image pull with authentication header.
Wait until pull request is complete before exiting function.
No need to export this particular struct since we already have a public
Client interface available and a NewClient function which can be used to
instantiate the concrete struct.
If we receive an error while trying to shutdown/startup a particular
container we don't want to immediately terminate the current update
cycle. Instead we should continue processing the remaining containers
and simply log the error.
Under certain conditions when watchtower is monitoring a Docker Swarm
cluster there would be cases where an updated container could not be
started because the old hadn't yet been removed (name conflicts, mapped
port conflicts, etc). We suspect that this has something to do with the
async nature of swarm and even though we've asked the swarm master to
remove a container it may not be completely removed from the associated
node. The fix is to do some polling after the remove container call to
ensure that the container is truly gone before proceeding.
