@Test
public void testSingleIndexShardFailed() {
AllocationService strategy =
createAllocationService(
settingsBuilder().put("cluster.routing.allocation.concurrent_recoveries", 10).build());
logger.info("Building initial routing table");
MetaData metaData =
MetaData.builder()
.put(IndexMetaData.builder("test").numberOfShards(1).numberOfReplicas(0))
.build();
RoutingTable routingTable = RoutingTable.builder().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
ClusterState.builder().metaData(metaData).routingTable(routingTable).build();
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(UNASSIGNED));
assertThat(routingTable.index("test").shard(0).shards().get(0).currentNodeId(), nullValue());
logger.info("Adding one node and rerouting");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder().put(newNode("node1")))
.build();
RoutingTable prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).unassigned(), equalTo(false));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(INITIALIZING));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node1"));
logger.info("Marking the shard as failed");
RoutingNodes routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyFailedShard(
clusterState, routingNodes.node("node1").shardsWithState(INITIALIZING).get(0))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(UNASSIGNED));
assertThat(routingTable.index("test").shard(0).shards().get(0).currentNodeId(), nullValue());
}
@Test
public void indexLevelShardsLimitAllocate() {
AllocationService strategy =
new AllocationService(
settingsBuilder().put("cluster.routing.allocation.concurrent_recoveries", 10).build());
logger.info("Building initial routing table");
MetaData metaData =
newMetaDataBuilder()
.put(
newIndexMetaDataBuilder("test")
.settings(
ImmutableSettings.settingsBuilder()
.put(IndexMetaData.SETTING_NUMBER_OF_SHARDS, 4)
.put(IndexMetaData.SETTING_NUMBER_OF_REPLICAS, 1)
.put(ShardsLimitAllocationDecider.INDEX_TOTAL_SHARDS_PER_NODE, 2)))
.build();
RoutingTable routingTable = routingTable().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
newClusterStateBuilder().metaData(metaData).routingTable(routingTable).build();
logger.info("Adding two nodes and performing rerouting");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(newNodesBuilder().put(newNode("node1")).put(newNode("node2")))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node1")
.numberOfShardsWithState(ShardRoutingState.INITIALIZING),
equalTo(2));
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node2")
.numberOfShardsWithState(ShardRoutingState.INITIALIZING),
equalTo(2));
logger.info("Start the primary shards");
RoutingNodes routingNodes = clusterState.routingNodes();
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node1")
.numberOfShardsWithState(ShardRoutingState.STARTED),
equalTo(2));
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node1")
.numberOfShardsWithState(ShardRoutingState.INITIALIZING),
equalTo(0));
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node2")
.numberOfShardsWithState(ShardRoutingState.STARTED),
equalTo(2));
assertThat(
clusterState
.readOnlyRoutingNodes()
.node("node2")
.numberOfShardsWithState(ShardRoutingState.INITIALIZING),
equalTo(0));
assertThat(clusterState.readOnlyRoutingNodes().unassigned().size(), equalTo(4));
logger.info("Do another reroute, make sure its still not allocated");
routingNodes = clusterState.routingNodes();
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
}
@Test
public void indexLevelShardsLimitRemain() {
AllocationService strategy =
new AllocationService(
settingsBuilder()
.put("cluster.routing.allocation.concurrent_recoveries", 10)
.put("cluster.routing.allocation.node_initial_primaries_recoveries", 10)
.put("cluster.routing.allocation.cluster_concurrent_rebalance", -1)
.put("cluster.routing.allocation.balance.index", 0.0f)
.put("cluster.routing.allocation.balance.replica", 1.0f)
.put("cluster.routing.allocation.balance.primary", 0.0f)
.build());
logger.info("Building initial routing table");
MetaData metaData =
newMetaDataBuilder()
.put(
newIndexMetaDataBuilder("test")
.settings(
ImmutableSettings.settingsBuilder()
.put(IndexMetaData.SETTING_NUMBER_OF_SHARDS, 5)
.put(IndexMetaData.SETTING_NUMBER_OF_REPLICAS, 0)))
.build();
RoutingTable routingTable = routingTable().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
newClusterStateBuilder().metaData(metaData).routingTable(routingTable).build();
logger.info("Adding one node and reroute");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(newNodesBuilder().put(newNode("node1")))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
logger.info("Start the primary shards");
RoutingNodes routingNodes = clusterState.routingNodes();
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(clusterState.readOnlyRoutingNodes().numberOfShardsOfType(STARTED), equalTo(5));
logger.info("add another index with 5 shards");
metaData =
newMetaDataBuilder()
.metaData(metaData)
.put(
newIndexMetaDataBuilder("test1")
.settings(
ImmutableSettings.settingsBuilder()
.put(IndexMetaData.SETTING_NUMBER_OF_SHARDS, 5)
.put(IndexMetaData.SETTING_NUMBER_OF_REPLICAS, 0)))
.build();
routingTable =
routingTable().routingTable(routingTable).addAsNew(metaData.index("test1")).build();
clusterState =
newClusterStateBuilder()
.state(clusterState)
.metaData(metaData)
.routingTable(routingTable)
.build();
logger.info("Add another one node and reroute");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(newNodesBuilder().putAll(clusterState.nodes()).put(newNode("node2")))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
routingNodes = clusterState.routingNodes();
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(clusterState.readOnlyRoutingNodes().numberOfShardsOfType(STARTED), equalTo(10));
for (MutableShardRouting shardRouting : clusterState.readOnlyRoutingNodes().node("node1")) {
assertThat(shardRouting.index(), equalTo("test"));
}
for (MutableShardRouting shardRouting : clusterState.readOnlyRoutingNodes().node("node2")) {
assertThat(shardRouting.index(), equalTo("test1"));
}
logger.info(
"update "
+ ShardsLimitAllocationDecider.INDEX_TOTAL_SHARDS_PER_NODE
+ " for test, see that things move");
metaData =
newMetaDataBuilder()
.metaData(metaData)
.put(
newIndexMetaDataBuilder("test")
.settings(
ImmutableSettings.settingsBuilder()
.put(IndexMetaData.SETTING_NUMBER_OF_SHARDS, 5)
.put(IndexMetaData.SETTING_NUMBER_OF_REPLICAS, 0)
.put(ShardsLimitAllocationDecider.INDEX_TOTAL_SHARDS_PER_NODE, 3)))
.build();
clusterState = newClusterStateBuilder().state(clusterState).metaData(metaData).build();
logger.info("reroute after setting");
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState.readOnlyRoutingNodes().node("node1").numberOfShardsWithState(STARTED),
equalTo(3));
assertThat(
clusterState.readOnlyRoutingNodes().node("node1").numberOfShardsWithState(RELOCATING),
equalTo(2));
assertThat(
clusterState.readOnlyRoutingNodes().node("node2").numberOfShardsWithState(RELOCATING),
equalTo(2));
assertThat(
clusterState.readOnlyRoutingNodes().node("node2").numberOfShardsWithState(STARTED),
equalTo(3));
// the first move will destroy the balance and the balancer will move 2 shards from node2 to
// node one right after
// moving the nodes to node2 since we consider INITIALIZING nodes during rebalance
routingNodes = clusterState.routingNodes();
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
// now we are done compared to EvenShardCountAllocator since the Balancer is not soely based on
// the average
assertThat(
clusterState.readOnlyRoutingNodes().node("node1").numberOfShardsWithState(STARTED),
equalTo(5));
assertThat(
clusterState.readOnlyRoutingNodes().node("node2").numberOfShardsWithState(STARTED),
equalTo(5));
}
@Test
public void testSingleIndexStartedShard() {
AllocationService strategy =
createAllocationService(
settingsBuilder().put("cluster.routing.allocation.concurrent_recoveries", 10).build());
logger.info("Building initial routing table");
MetaData metaData =
MetaData.builder()
.put(
IndexMetaData.builder("test")
.settings(settings(Version.CURRENT))
.numberOfShards(1)
.numberOfReplicas(0))
.build();
RoutingTable routingTable = RoutingTable.builder().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
ClusterState.builder(org.elasticsearch.cluster.ClusterName.DEFAULT)
.metaData(metaData)
.routingTable(routingTable)
.build();
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(UNASSIGNED));
assertThat(routingTable.index("test").shard(0).shards().get(0).currentNodeId(), nullValue());
logger.info("Adding one node and performing rerouting");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder().put(newNode("node1")))
.build();
RoutingTable prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(INITIALIZING));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node1"));
logger.info("Rerouting again, nothing should change");
prevRoutingTable = routingTable;
clusterState = ClusterState.builder(clusterState).build();
routingTable = strategy.reroute(clusterState).routingTable();
assertThat(routingTable == prevRoutingTable, equalTo(true));
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
logger.info("Marking the shard as started");
RoutingNodes routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(
clusterState, routingNodes.node("node1").shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable != prevRoutingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node1"));
logger.info("Starting another node and making sure nothing changed");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder(clusterState.nodes()).put(newNode("node2")))
.build();
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable == prevRoutingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node1"));
logger.info("Killing node1 where the shard is, checking the shard is relocated");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder(clusterState.nodes()).remove("node1"))
.build();
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable != prevRoutingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(INITIALIZING));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node2"));
logger.info(
"Start another node, make sure that things remain the same (shard is in node2 and initializing)");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder(clusterState.nodes()).put(newNode("node3")))
.build();
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable == prevRoutingTable, equalTo(true));
logger.info("Start the shard on node 2");
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(
clusterState, routingNodes.node("node2").shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(routingTable != prevRoutingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).shards().get(0).currentNodeId(), equalTo("node2"));
}
@Test
public void testMultiIndexUnevenNodes() {
AllocationService strategy =
createAllocationService(
settingsBuilder()
.put("cluster.routing.allocation.concurrent_recoveries", 10)
.put(
ClusterRebalanceAllocationDecider.CLUSTER_ROUTING_ALLOCATION_ALLOW_REBALANCE,
"always")
.put("cluster.routing.allocation.cluster_concurrent_rebalance", -1)
.build());
final int numberOfIndices = 10;
logger.info("Building initial routing table with " + numberOfIndices + " indices");
MetaData.Builder metaDataBuilder = MetaData.builder();
for (int i = 0; i < numberOfIndices; i++) {
metaDataBuilder.put(
IndexMetaData.builder("test" + i)
.settings(settings(Version.CURRENT))
.numberOfShards(1)
.numberOfReplicas(0));
}
MetaData metaData = metaDataBuilder.build();
RoutingTable.Builder routingTableBuilder = RoutingTable.builder();
for (int i = 0; i < numberOfIndices; i++) {
routingTableBuilder.addAsNew(metaData.index("test" + i));
}
RoutingTable routingTable = routingTableBuilder.build();
ClusterState clusterState =
ClusterState.builder(org.elasticsearch.cluster.ClusterName.DEFAULT)
.metaData(metaData)
.routingTable(routingTable)
.build();
assertThat(routingTable.indicesRouting().size(), equalTo(numberOfIndices));
logger.info("Starting 3 nodes and rerouting");
clusterState =
ClusterState.builder(clusterState)
.nodes(
DiscoveryNodes.builder()
.put(newNode("node1"))
.put(newNode("node2"))
.put(newNode("node3")))
.build();
RoutingTable prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(), equalTo(INITIALIZING));
}
RoutingNodes routingNodes = clusterState.routingNodes();
assertThat(numberOfShardsOfType(routingNodes, INITIALIZING), equalTo(numberOfIndices));
assertThat(
routingNodes.node("node1").numberOfShardsWithState(INITIALIZING),
anyOf(equalTo(3), equalTo(4)));
assertThat(
routingNodes.node("node2").numberOfShardsWithState(INITIALIZING),
anyOf(equalTo(3), equalTo(4)));
assertThat(
routingNodes.node("node2").numberOfShardsWithState(INITIALIZING),
anyOf(equalTo(3), equalTo(4)));
logger.info("Start two more nodes, things should remain the same");
clusterState =
ClusterState.builder(clusterState)
.nodes(
DiscoveryNodes.builder(clusterState.nodes())
.put(newNode("node4"))
.put(newNode("node5")))
.build();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
assertThat(prevRoutingTable == routingTable, equalTo(true));
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(),
anyOf(equalTo(RELOCATING), equalTo(STARTED)));
}
routingNodes = clusterState.routingNodes();
assertThat(
"4 source shard routing are relocating",
numberOfShardsOfType(routingNodes, RELOCATING),
equalTo(4));
assertThat(
"4 target shard routing are initializing",
numberOfShardsOfType(routingNodes, INITIALIZING),
equalTo(4));
logger.info("Now, mark the relocated as started");
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
// routingTable = strategy.reroute(new RoutingStrategyInfo(metaData, routingTable),
// nodes);
assertThat(prevRoutingTable != routingTable, equalTo(true));
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(),
anyOf(equalTo(RELOCATING), equalTo(STARTED)));
}
routingNodes = clusterState.routingNodes();
assertThat(numberOfShardsOfType(routingNodes, STARTED), equalTo(numberOfIndices));
for (RoutingNode routingNode : routingNodes) {
assertThat(routingNode.numberOfShardsWithState(STARTED), equalTo(2));
}
}
@Test
public void testMultiIndexEvenDistribution() {
AllocationService strategy =
createAllocationService(
settingsBuilder()
.put("cluster.routing.allocation.concurrent_recoveries", 10)
.put(
ClusterRebalanceAllocationDecider.CLUSTER_ROUTING_ALLOCATION_ALLOW_REBALANCE,
"always")
.put("cluster.routing.allocation.cluster_concurrent_rebalance", -1)
.build());
final int numberOfIndices = 50;
logger.info("Building initial routing table with " + numberOfIndices + " indices");
MetaData.Builder metaDataBuilder = MetaData.builder();
for (int i = 0; i < numberOfIndices; i++) {
metaDataBuilder.put(
IndexMetaData.builder("test" + i)
.settings(settings(Version.CURRENT))
.numberOfShards(1)
.numberOfReplicas(0));
}
MetaData metaData = metaDataBuilder.build();
RoutingTable.Builder routingTableBuilder = RoutingTable.builder();
for (int i = 0; i < numberOfIndices; i++) {
routingTableBuilder.addAsNew(metaData.index("test" + i));
}
RoutingTable routingTable = routingTableBuilder.build();
ClusterState clusterState =
ClusterState.builder(org.elasticsearch.cluster.ClusterName.DEFAULT)
.metaData(metaData)
.routingTable(routingTable)
.build();
assertThat(routingTable.indicesRouting().size(), equalTo(numberOfIndices));
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(), equalTo(UNASSIGNED));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).currentNodeId(), nullValue());
}
logger.info("Adding " + (numberOfIndices / 2) + " nodes");
DiscoveryNodes.Builder nodesBuilder = DiscoveryNodes.builder();
List<DiscoveryNode> nodes = newArrayList();
for (int i = 0; i < (numberOfIndices / 2); i++) {
nodesBuilder.put(newNode("node" + i));
}
RoutingTable prevRoutingTable = routingTable;
clusterState = ClusterState.builder(clusterState).nodes(nodesBuilder).build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).unassigned(), equalTo(false));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(), equalTo(INITIALIZING));
assertThat(routingTable.index("test" + i).shard(0).shards().get(0).primary(), equalTo(true));
// make sure we still have 2 shards initializing per node on the first 25 nodes
String nodeId = routingTable.index("test" + i).shard(0).shards().get(0).currentNodeId();
int nodeIndex = Integer.parseInt(nodeId.substring("node".length()));
assertThat(nodeIndex, lessThan(25));
}
RoutingNodes routingNodes = clusterState.routingNodes();
Set<String> encounteredIndices = newHashSet();
for (RoutingNode routingNode : routingNodes) {
assertThat(routingNode.numberOfShardsWithState(STARTED), equalTo(0));
assertThat(routingNode.size(), equalTo(2));
// make sure we still have 2 shards initializing per node on the only 25 nodes
int nodeIndex = Integer.parseInt(routingNode.nodeId().substring("node".length()));
assertThat(nodeIndex, lessThan(25));
// check that we don't have a shard associated with a node with the same index name (we have a
// single shard)
for (ShardRouting shardRoutingEntry : routingNode) {
assertThat(encounteredIndices, not(hasItem(shardRoutingEntry.index())));
encounteredIndices.add(shardRoutingEntry.index());
}
}
logger.info("Adding additional " + (numberOfIndices / 2) + " nodes, nothing should change");
nodesBuilder = DiscoveryNodes.builder(clusterState.nodes());
for (int i = (numberOfIndices / 2); i < numberOfIndices; i++) {
nodesBuilder.put(newNode("node" + i));
}
prevRoutingTable = routingTable;
clusterState = ClusterState.builder(clusterState).nodes(nodesBuilder).build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(false));
logger.info("Marking the shard as started");
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
int numberOfRelocatingShards = 0;
int numberOfStartedShards = 0;
for (int i = 0; i < numberOfIndices; i++) {
assertThat(routingTable.index("test" + i).shards().size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).size(), equalTo(1));
assertThat(routingTable.index("test" + i).shard(0).shards().size(), equalTo(1));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).unassigned(), equalTo(false));
assertThat(
routingTable.index("test" + i).shard(0).shards().get(0).state(),
anyOf(equalTo(STARTED), equalTo(RELOCATING)));
if (routingTable.index("test" + i).shard(0).shards().get(0).state() == STARTED) {
numberOfStartedShards++;
} else if (routingTable.index("test" + i).shard(0).shards().get(0).state() == RELOCATING) {
numberOfRelocatingShards++;
}
assertThat(routingTable.index("test" + i).shard(0).shards().get(0).primary(), equalTo(true));
// make sure we still have 2 shards either relocating or started on the first 25 nodes (still)
String nodeId = routingTable.index("test" + i).shard(0).shards().get(0).currentNodeId();
int nodeIndex = Integer.parseInt(nodeId.substring("node".length()));
assertThat(nodeIndex, lessThan(25));
}
assertThat(numberOfRelocatingShards, equalTo(25));
assertThat(numberOfStartedShards, equalTo(25));
}
@Test
public void testUpdateNumberOfReplicas() {
AllocationService strategy =
createAllocationService(
settingsBuilder().put("cluster.routing.allocation.concurrent_recoveries", 10).build());
logger.info("Building initial routing table");
MetaData metaData =
MetaData.builder()
.put(IndexMetaData.builder("test").numberOfShards(1).numberOfReplicas(1))
.build();
RoutingTable routingTable = RoutingTable.builder().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
ClusterState.builder().metaData(metaData).routingTable(routingTable).build();
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(2));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(2));
assertThat(routingTable.index("test").shard(0).shards().get(0).state(), equalTo(UNASSIGNED));
assertThat(routingTable.index("test").shard(0).shards().get(1).state(), equalTo(UNASSIGNED));
assertThat(routingTable.index("test").shard(0).shards().get(0).currentNodeId(), nullValue());
assertThat(routingTable.index("test").shard(0).shards().get(1).currentNodeId(), nullValue());
logger.info("Adding two nodes and performing rerouting");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder().put(newNode("node1")).put(newNode("node2")))
.build();
RoutingTable prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
logger.info("Start all the primary shards");
RoutingNodes routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
logger.info("Start all the replica shards");
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
final String nodeHoldingPrimary =
routingTable.index("test").shard(0).primaryShard().currentNodeId();
final String nodeHoldingReplica =
routingTable.index("test").shard(0).replicaShards().get(0).currentNodeId();
assertThat(nodeHoldingPrimary, not(equalTo(nodeHoldingReplica)));
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(2));
assertThat(routingTable.index("test").shard(0).shards().size(), equalTo(2));
assertThat(routingTable.index("test").shard(0).primaryShard().state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).primaryShard().currentNodeId(),
equalTo(nodeHoldingPrimary));
assertThat(routingTable.index("test").shard(0).replicaShards().size(), equalTo(1));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).currentNodeId(),
equalTo(nodeHoldingReplica));
logger.info("add another replica");
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable = RoutingTable.builder(routingTable).updateNumberOfReplicas(2).build();
metaData = MetaData.builder(clusterState.metaData()).updateNumberOfReplicas(2).build();
clusterState =
ClusterState.builder(clusterState).routingTable(routingTable).metaData(metaData).build();
assertThat(clusterState.metaData().index("test").numberOfReplicas(), equalTo(2));
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(3));
assertThat(routingTable.index("test").shard(0).primaryShard().state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).primaryShard().currentNodeId(),
equalTo(nodeHoldingPrimary));
assertThat(routingTable.index("test").shard(0).replicaShards().size(), equalTo(2));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).currentNodeId(),
equalTo(nodeHoldingReplica));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(1).state(), equalTo(UNASSIGNED));
logger.info("Add another node and start the added replica");
clusterState =
ClusterState.builder(clusterState)
.nodes(DiscoveryNodes.builder(clusterState.nodes()).put(newNode("node3")))
.build();
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(3));
assertThat(routingTable.index("test").shard(0).primaryShard().state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).primaryShard().currentNodeId(),
equalTo(nodeHoldingPrimary));
assertThat(routingTable.index("test").shard(0).replicaShards().size(), equalTo(2));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(STARTED).size(), equalTo(1));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(STARTED).get(0).currentNodeId(),
equalTo(nodeHoldingReplica));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(INITIALIZING).size(),
equalTo(1));
assertThat(
routingTable
.index("test")
.shard(0)
.replicaShardsWithState(INITIALIZING)
.get(0)
.currentNodeId(),
equalTo("node3"));
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable =
strategy
.applyStartedShards(clusterState, routingNodes.shardsWithState(INITIALIZING))
.routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(3));
assertThat(routingTable.index("test").shard(0).primaryShard().state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).primaryShard().currentNodeId(),
equalTo(nodeHoldingPrimary));
assertThat(routingTable.index("test").shard(0).replicaShards().size(), equalTo(2));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(STARTED).size(), equalTo(2));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(STARTED).get(0).currentNodeId(),
anyOf(equalTo(nodeHoldingReplica), equalTo("node3")));
assertThat(
routingTable.index("test").shard(0).replicaShardsWithState(STARTED).get(1).currentNodeId(),
anyOf(equalTo(nodeHoldingReplica), equalTo("node3")));
logger.info("now remove a replica");
routingNodes = clusterState.routingNodes();
prevRoutingTable = routingTable;
routingTable = RoutingTable.builder(routingTable).updateNumberOfReplicas(1).build();
metaData = MetaData.builder(clusterState.metaData()).updateNumberOfReplicas(1).build();
clusterState =
ClusterState.builder(clusterState).routingTable(routingTable).metaData(metaData).build();
assertThat(clusterState.metaData().index("test").numberOfReplicas(), equalTo(1));
assertThat(prevRoutingTable != routingTable, equalTo(true));
assertThat(routingTable.index("test").shards().size(), equalTo(1));
assertThat(routingTable.index("test").shard(0).size(), equalTo(2));
assertThat(routingTable.index("test").shard(0).primaryShard().state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).primaryShard().currentNodeId(),
equalTo(nodeHoldingPrimary));
assertThat(routingTable.index("test").shard(0).replicaShards().size(), equalTo(1));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).state(), equalTo(STARTED));
assertThat(
routingTable.index("test").shard(0).replicaShards().get(0).currentNodeId(),
anyOf(equalTo(nodeHoldingReplica), equalTo("node3")));
logger.info("do a reroute, should remain the same");
prevRoutingTable = routingTable;
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = ClusterState.builder(clusterState).routingTable(routingTable).build();
assertThat(prevRoutingTable != routingTable, equalTo(false));
}
@Test
public void testPrimaryRecoveryThrottling() {
AllocationService strategy =
new AllocationService(
settingsBuilder()
.put("cluster.routing.allocation.node_concurrent_recoveries", 3)
.put("cluster.routing.allocation.node_initial_primaries_recoveries", 3)
.build());
logger.info("Building initial routing table");
MetaData metaData =
newMetaDataBuilder()
.put(newIndexMetaDataBuilder("test").numberOfShards(10).numberOfReplicas(1))
.build();
RoutingTable routingTable = routingTable().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
newClusterStateBuilder().metaData(metaData).routingTable(routingTable).build();
logger.info("start one node, do reroute, only 3 should initialize");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(newNodesBuilder().put(newNode("node1")))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(routingTable.shardsWithState(STARTED).size(), equalTo(0));
assertThat(routingTable.shardsWithState(INITIALIZING).size(), equalTo(3));
assertThat(routingTable.shardsWithState(UNASSIGNED).size(), equalTo(17));
logger.info("start initializing, another 3 should initialize");
routingTable =
strategy
.applyStartedShards(clusterState, routingTable.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(routingTable.shardsWithState(STARTED).size(), equalTo(3));
assertThat(routingTable.shardsWithState(INITIALIZING).size(), equalTo(3));
assertThat(routingTable.shardsWithState(UNASSIGNED).size(), equalTo(14));
logger.info("start initializing, another 3 should initialize");
routingTable =
strategy
.applyStartedShards(clusterState, routingTable.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(routingTable.shardsWithState(STARTED).size(), equalTo(6));
assertThat(routingTable.shardsWithState(INITIALIZING).size(), equalTo(3));
assertThat(routingTable.shardsWithState(UNASSIGNED).size(), equalTo(11));
logger.info("start initializing, another 1 should initialize");
routingTable =
strategy
.applyStartedShards(clusterState, routingTable.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(routingTable.shardsWithState(STARTED).size(), equalTo(9));
assertThat(routingTable.shardsWithState(INITIALIZING).size(), equalTo(1));
assertThat(routingTable.shardsWithState(UNASSIGNED).size(), equalTo(10));
logger.info("start initializing, all primaries should be started");
routingTable =
strategy
.applyStartedShards(clusterState, routingTable.shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(routingTable.shardsWithState(STARTED).size(), equalTo(10));
assertThat(routingTable.shardsWithState(INITIALIZING).size(), equalTo(0));
assertThat(routingTable.shardsWithState(UNASSIGNED).size(), equalTo(10));
}