@Override
public void handleResponse(ClientResponse resp) {
if (resp == null) {
VoltDB.crashLocalVoltDB(
"Received a null response to a snapshot initiation request. "
+ "This should be impossible.",
true,
null);
} else if (resp.getStatus() != ClientResponse.SUCCESS) {
tmLog.info(
"Failed to complete partition detection snapshot, status: "
+ resp.getStatus()
+ ", reason: "
+ resp.getStatusString());
tmLog.info("Retrying partition detection snapshot...");
SnapshotUtil.requestSnapshot(
0L,
m_partSnapshotSchedule.getPath(),
m_partSnapshotSchedule.getPrefix() + System.currentTimeMillis(),
true,
SnapshotFormat.NATIVE,
null,
m_snapshotHandler,
true);
} else if (!SnapshotUtil.didSnapshotRequestSucceed(resp.getResults())) {
VoltDB.crashGlobalVoltDB(
"Unable to complete partition detection snapshot: " + resp.getResults()[0],
false,
null);
} else {
VoltDB.crashGlobalVoltDB(
"Partition detection snapshot completed. Shutting down.", false, null);
}
}
private void doPartitionDetectionActivities() {
// We should never re-enter here once we've decided we're partitioned and doomed
assert (!m_partitionDetected);
// After everything is resolved, write the new surviving set to ZK
List<Integer> currentNodes = null;
try {
currentNodes = m_hostMessenger.getLiveHostIds();
} catch (Exception e) {
}
Set<Integer> currentHosts = new HashSet<Integer>(currentNodes);
Set<Integer> previousHosts = readPriorKnownLiveNodes();
boolean partitionDetectionTriggered = makePPDDecision(previousHosts, currentHosts);
if (partitionDetectionTriggered) {
m_partitionDetected = true;
if (m_usingCommandLog) {
// Just shut down immediately
VoltDB.crashGlobalVoltDB(
"Use of command logging detected, no additional database snapshot will "
+ "be generated. Please use the 'recover' action to restore the database if necessary.",
false,
null);
} else {
SnapshotUtil.requestSnapshot(
0L,
m_partSnapshotSchedule.getPath(),
m_partSnapshotSchedule.getPrefix() + System.currentTimeMillis(),
true,
SnapshotFormat.NATIVE,
null,
m_snapshotHandler,
true);
}
}
// If the cluster host set has changed, then write the new set to ZK
// NOTE: we don't want to update the known live nodes if we've decided that our subcluster is
// dying, otherwise a poorly timed subsequent failure might reverse this decision. Any future
// promoted
// LeaderAppointer should make their partition detection decision based on the pre-partition
// cluster state.
else if (!currentHosts.equals(previousHosts)) {
writeKnownLiveNodes(currentNodes);
}
}
