AMBARI-25620 Ambari Server STOP command on a node might fail because …

[ramkrish86]

…of a decommissioned NM's behaviour

What changes were proposed in this pull request?

As part of our use case, before we STOP any set of components in a node, if we see a NodeManager and Datanode in that node, we first Decommission them and then issue STOP components to all the components in the node.

DataNodes once decommissioned are not STOPPED but they are alive as a process. Unless we stop them. Whereas RM stops the NM on decommission. (this is a known behaviour). But what happens is that the RecoveryManager in ambari keeps restarting the service thinking its DESIRED state (in the agent side) is STARTED. So the restart keeps happening. So the state changes between STARTED <-> INSTALLED on the agent side and once this happens we communicate the Component status to the server side. On receiving this update the server sets the STATE as STARTED/INSTALLED as the case may be. Now coming back to the actual STOP command request that we gave, as per design in ambari server once all the component updates are sent, it processes them in batch and tries to do the in-memory transition of STATES on the server side (not the cache but the FSM (state machine transition). Here the event is INSTALL/STOP event for NM that the server is expecting but instead of getting an INSTALLED state it gets STARTED state. The reason as highlighted above. So the entire STOP command gets aborted by the server thinking there is some problem in what it sees. /Multimap is analog of Map<Object, List<Object>> but allows to avoid nested loop ListMultimap<String, ServiceComponentHostEvent> eventMap = formEventMap(stage, commandsToStart); Map<ExecutionCommand, String> commandsToAbort = new HashMap<>(); if (!eventMap.isEmpty()) { LOG.debug("==> processing {} serviceComponentHostEvents...", eventMap.size()); Cluster cluster = clusters.getCluster(stage.getClusterName()); if (cluster != null) { Map<ServiceComponentHostEvent, String> failedEvents = cluster.processServiceComponentHostEvents(eventMap);

        if (failedEvents.size() > 0) {
          LOG.error("==> {} events failed.", failedEvents.size());
        }

        for (Iterator<ExecutionCommand> iterator = commandsToUpdate.iterator(); iterator.hasNext(); ) {
          ExecutionCommand cmd = iterator.next();
          for (ServiceComponentHostEvent event : failedEvents.keySet()) {
            if (StringUtils.equals(event.getHostName(), cmd.getHostname()) &&
              StringUtils.equals(event.getServiceComponentName(), cmd.getRole())) {
              iterator.remove();
              commandsToAbort.put(cmd, failedEvents.get(event));
              break;
            }
          }
        }

Check the processServiceComponentHostEvents() for the way the transition happens and what is the Invalid Transition that happens over there. The log msg would be like this org.apache.ambari.server.state.fsm.InvalidStateTransitionException: Invalid event: HOST_SVCCOMP_INSTALL at STARTED Since this entire set of STOP component is considered as a FAILURe, we issue ABORT command and hence all the STOP command issued to the agent are aborted. This makes the DN to stay in the STARTED state itself and hence the remaining DELETE HOST command keeps failing. The idea is to ensure that for NM if decommissioned and the current state is STARTED for a HOST_SVCCOMP_INSTALL then mark it as not a failure condition so that the commands are not aborted.

How was this patch tested?

(Please explain how this patch was tested. Ex: unit tests, manual tests) The patch was tested with a scale down feature of more than 20 nodes at one shot.

[surnaik]

LGTM +1. Required in case where we are scaling down, especially since Nodemanager is a special scenario. Good to get Andrew's review on this one. Ping @aonishuk

[ramkrish86]

@hapylestat and @aonishuk - Gentle ping.