Details
Major Description
Symptom
========
We have a production system that involves a set of Brokers connected in a demand-forwarding Network-of-Brokers using HTTP-based bridges. Each Broker periodically scans its list of peer brokers by iterating over RegionBroker.getPeerBrokerInfos:
public synchronized BrokerInfo[] getPeerBrokerInfos()
{ BrokerInfo[] result = new BrokerInfo[brokerInfos.size()]; result = brokerInfos.toArray(result); return result; }This scanning code assumes that BrokerInfo.getBrokerId() is always non-null (since every broker should have an ID). However, we periodically noticed that BrokerInfo.getBrokerId() returned a NULL value, which was very unexpected.
Cause
======
We analyzed the DemandForwardingBridgeSupport and noticed that when the remote bridge/transport is started, it sends the local Broker's ID:
protected void startRemoteBridge() throws Exception
{ ... brokerInfo.setBrokerId(this.localBrokerId); remoteBroker.oneway(brokerInfo); }The local Broker's ID is not initially known until it is received from the local transport and processed by DemandForwardingBridge.serviceLocalBrokerInfo(...):
protected void serviceLocalBrokerInfo(Command command) throws InterruptedException {
synchronized (brokerInfoMutex) {
localBrokerId = ((BrokerInfo)command).getBrokerId();
localBrokerPath[0] = localBrokerId;
localBrokerIdKnownLatch.countDown();
The local Broker's ID is dispatched asynchronously when the local transport is started, as seen in TransportConnection.start():
public void start() throws Exception {
starting = true;
try {
synchronized (this) {
if (taskRunnerFactory != null)
else
{ taskRunner = null; } transport.start();
active = true;
dispatchAsync(connector.getBrokerInfo());
Because of the asynchronous dispatch, the remote bridge may be started before the local Broker's ID is known. This would be particularly evident when the local broker is under load processing a lot of tasks.
We've attached a unit test that demonstrates how a slow asynchronous dispatch on the local transport can cause the remote transport to transmit a null BrokerId.
Solution
======
DemandForwardingBridgeSupport already contains a localBrokerIdKnownLatch, so starting the remote transport should wait for this latch before accessing the local Broker's ID (see patch).