Package org.apache.drill.yarn.zk

Class ZKRegistry

java.lang.Object
org.apache.drill.yarn.zk.ZKRegistry
All Implemented Interfaces:
DrillbitStatusListener, Pollable, TaskLifecycleListener
public class ZKRegistry extends Object implements TaskLifecycleListener, DrillbitStatusListener, Pollable
AM-specific implementation of a Drillbit registry backed by ZooKeeper. Listens to ZK events for registering a Drillbit and deregistering. Alerts the Cluster Controller of these events.

Locking strategy: Receives events from both ZK and the cluster controller, both of which must be synchronized. To prevent deadlocks, this class NEVER calls into the cluster controller while holding a lock. This prevents the following:

ClusterController --> ZKRegistry (OK)
ZK --> ZKRegistry (OK)
ZK --> ZKRegistry --> Cluster Controller (bad)

In the case of registration, ZK calls the registry which must alert the cluster controller. Cluster controller alerting is handled outside the ZK update critical section.

Because ZK events are occur relatively infrequently, any deadlock will occur once in a blue moon, which will make it very hard to reproduce. So, extreme caution is needed at design time to prevent the problem.

  • Field Details

  • Constructor Details

  • Method Details

    • start

      public void start(RegistryHandler controller)
      Called during AM startup to initialize ZK. Checks if any Drillbits are already running. These are "unmanaged" because the AM could not have started them (since they predate the AM.)

    • drillbitRegistered

      public void drillbitRegistered(Set<CoordinationProtos.DrillbitEndpoint> registeredDrillbits)
      Callback from ZK to indicate that one or more drillbits have become registered. We handle registrations in a critical section, then alert the cluster controller outside the critical section.
      Specified by:
      drillbitRegistered in interface DrillbitStatusListener
      Parameters:
      registeredDrillbits - the set of newly registered drillbits. Note: the complete set of currently registered bits could be different.

    • drillbitUnregistered

      public void drillbitUnregistered(Set<CoordinationProtos.DrillbitEndpoint> unregisteredDrillbits)
      Callback from ZK to indicate that one or more drillbits have become deregistered from ZK. We handle the deregistrations in a critical section, but updates to the cluster controller outside of a critical section.
      Specified by:
      drillbitUnregistered in interface DrillbitStatusListener
      Parameters:
      unregisteredDrillbits - the set of newly unregistered drillbits.

    • stateChange

      public void stateChange(TaskLifecycleListener.Event event, EventContext context)
      Listen for selected YARN task state changes. Called from within the cluster controller's critical section.
      Specified by:
      stateChange in interface TaskLifecycleListener

    • isRegistered

      public boolean isRegistered(Task task)
      Report whether the given task is still registered in ZK. Called while waiting for a deregistration event to catch possible cases where the messages is lost. The message should never be lost, but we've seen cases where tasks hang in this state. This is a potential work-around.
      Parameters:
      task -
      Returns:
      True if the given task is regestered. False otherwise.

    • tick

      public void tick(long curTime)
      Periodically check ZK status. If the ZK connection has timed out, something is very seriously wrong. Shut the whole Drill cluster down since Drill cannot operate without ZooKeeper.

      This method should not be synchronized. It checks only the ZK state, not internal state. Further, if we do reconnect to ZK, then a ZK thread may attempt to update this registry, which will acquire a synchronization lock.

      Specified by:
      tick in interface Pollable

    • finish

      public void finish(RegistryHandler handler)

    • listUnmanagedDrillits

      public List<String> listUnmanagedDrillits()

    • getRegistryForTesting

      protected Map<String,ZKRegistry.DrillbitTracker> getRegistryForTesting()
      Get the current registry for testing. Why for testing? Because this is unsynchronized. In production code, the map may change out from under you.
      Returns:
      The current registry.