The map is only ever used in the loop to create and remove stacks, so it
doesn't need to be stored in the struct. This ensures that there can't
be any racy concurrent accesses to it.

Signed-off-by: Tom Wieczorek <[email protected]>

The only reason these channels get closed is if the watcher itself gets
closed. This happens only when the method returns, which in turn only
happens when the context is done. In this case, the loop has already
exited without a select on a potentially closed channel. So the branches
that checked for closed channels were effectively unreachable during
runtime.

Signed-off-by: Tom Wieczorek <[email protected]>

Rename cancelWatcher to stop and wait until the newly added stopped
channel is closed. Also, add a stopped channel to each stack to do the
same for each stack-specific goroutine.

Signed-off-by: Tom Wieczorek <[email protected]>

Cancel the contexts with a cause. Add this cause to the log statements
when exiting loops. Rename bundlePath to bundleDir to reflect the fact
that it is a directory, not a file.

Signed-off-by: Tom Wieczorek <[email protected]>

Exit the loop on error and restart it after a one-minute delay to allow
it to recover in a new run. Also replace the bespoke retry loop for
stacks with the Kubernetes client's wait package.

Signed-off-by: Tom Wieczorek <[email protected]>

Seems to be a remnant from the past.

Signed-off-by: Tom Wieczorek <[email protected]>

Signed-off-by: Ethan Mosbaugh <[email protected]>