fix socket shutdown in pod watcher

System IO raises an EOFError when reading from a closed socket,
which happens when we close a long running connection listening
for pod events. Celluloid transforms EOFError into MailboxShutdown
and terminates the actor quietly. We ned to rescue EOFError before
it gets to Celluloid.

plugins/kubernetes/app/models/watchers/cluster_pod_watcher.rb

@@ -10,15 +10,13 @@ class ClusterPodWatcher
     finalizer :stop_watching
 
     def initialize(client)
-      @started = false
       @client = client
       @watch_stream = nil
       async :start_watching
     end
 
     def start_watching
       info 'watcher started'
-      @started = true
       @watch_stream = @client.watch_pods
       @watch_stream.each do |notice|
         handle_notice notice
@@ -27,7 +25,6 @@ def start_watching
 
     def stop_watching
       info 'watcher stopped'
-      @started = false
       if @watch_stream
         @watch_stream.finish
         @watch_stream = nil

plugins/kubernetes/config/initializers/watchers.rb

@@ -17,7 +17,7 @@ def each(&block)
           fail KubeException.new(response.code, response.reason, response)
         end
         read_stream(response.body, &block)
-      rescue Errno::EBADF
+      rescue IOError
         raise unless @finished
       end