private Map<DiscoveryNode, TransportNodesListShardStoreMetaData.StoreFilesMetaData>
buildShardStores(DiscoveryNodes nodes, MutableShardRouting shard) {
Map<DiscoveryNode, TransportNodesListShardStoreMetaData.StoreFilesMetaData> shardStores =
cachedStores.get(shard.shardId());
ObjectOpenHashSet<String> nodesIds;
if (shardStores == null) {
shardStores = Maps.newHashMap();
cachedStores.put(shard.shardId(), shardStores);
nodesIds = ObjectOpenHashSet.from(nodes.dataNodes().keys());
} else {
nodesIds = ObjectOpenHashSet.newInstance();
// clean nodes that have failed
for (Iterator<DiscoveryNode> it = shardStores.keySet().iterator(); it.hasNext(); ) {
DiscoveryNode node = it.next();
if (!nodes.nodeExists(node.id())) {
it.remove();
}
}
for (ObjectCursor<DiscoveryNode> cursor : nodes.dataNodes().values()) {
DiscoveryNode node = cursor.value;
if (!shardStores.containsKey(node)) {
nodesIds.add(node.id());
}
}
}
if (!nodesIds.isEmpty()) {
String[] nodesIdsArray = nodesIds.toArray(String.class);
TransportNodesListShardStoreMetaData.NodesStoreFilesMetaData nodesStoreFilesMetaData =
listShardStoreMetaData
.list(shard.shardId(), false, nodesIdsArray, listTimeout)
.actionGet();
if (logger.isTraceEnabled()) {
if (nodesStoreFilesMetaData.failures().length > 0) {
StringBuilder sb =
new StringBuilder(shard + ": failures when trying to list stores on nodes:");
for (int i = 0; i < nodesStoreFilesMetaData.failures().length; i++) {
Throwable cause = ExceptionsHelper.unwrapCause(nodesStoreFilesMetaData.failures()[i]);
if (cause instanceof ConnectTransportException) {
continue;
}
sb.append("\n -> ")
.append(nodesStoreFilesMetaData.failures()[i].getDetailedMessage());
}
logger.trace(sb.toString());
}
}
for (TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData nodeStoreFilesMetaData :
nodesStoreFilesMetaData) {
if (nodeStoreFilesMetaData.storeFilesMetaData() != null) {
shardStores.put(
nodeStoreFilesMetaData.getNode(), nodeStoreFilesMetaData.storeFilesMetaData());
}
}
}
return shardStores;
}
@Override
public boolean canRebalance(ShardRouting shardRouting, RoutingAllocation allocation) {
if (clusterConcurrentRebalance == -1) {
return true;
}
int rebalance = 0;
for (RoutingNode node : allocation.routingNodes()) {
for (MutableShardRouting shard : node) {
if (shard.state() == ShardRoutingState.RELOCATING) {
rebalance++;
}
}
}
if (rebalance >= clusterConcurrentRebalance) {
return false;
}
return true;
}
private boolean applyStartedShards(
RoutingNodes routingNodes, Iterable<? extends ShardRouting> startedShardEntries) {
boolean dirty = false;
// apply shards might be called several times with the same shard, ignore it
for (ShardRouting startedShard : startedShardEntries) {
assert startedShard.state() == INITIALIZING;
// retrieve the relocating node id before calling startedShard().
String relocatingNodeId = null;
RoutingNodes.RoutingNodeIterator currentRoutingNode =
routingNodes.routingNodeIter(startedShard.currentNodeId());
if (currentRoutingNode != null) {
for (MutableShardRouting shard : currentRoutingNode) {
if (shard.shardId().equals(startedShard.shardId())) {
relocatingNodeId = shard.relocatingNodeId();
if (!shard.started()) {
dirty = true;
routingNodes.started(shard);
}
break;
}
}
}
// startedShard is the current state of the shard (post relocation for example)
// this means that after relocation, the state will be started and the currentNodeId will be
// the node we relocated to
if (relocatingNodeId == null) {
continue;
}
RoutingNodes.RoutingNodeIterator sourceRoutingNode =
routingNodes.routingNodeIter(relocatingNodeId);
if (sourceRoutingNode != null) {
while (sourceRoutingNode.hasNext()) {
MutableShardRouting shard = sourceRoutingNode.next();
if (shard.shardId().equals(startedShard.shardId())) {
if (shard.relocating()) {
dirty = true;
sourceRoutingNode.remove();
break;
}
}
}
}
}
return dirty;
}
private boolean moveShards(RoutingAllocation allocation) {
boolean changed = false;
// create a copy of the shards interleaving between nodes, and check if they can remain
List<MutableShardRouting> shards = new ArrayList<>();
int index = 0;
boolean found = true;
final RoutingNodes routingNodes = allocation.routingNodes();
while (found) {
found = false;
for (RoutingNode routingNode : routingNodes) {
if (index >= routingNode.size()) {
continue;
}
found = true;
shards.add(routingNode.get(index));
}
index++;
}
for (int i = 0; i < shards.size(); i++) {
MutableShardRouting shardRouting = shards.get(i);
// we can only move started shards...
if (!shardRouting.started()) {
continue;
}
final RoutingNode routingNode = routingNodes.node(shardRouting.currentNodeId());
Decision decision = allocation.deciders().canRemain(shardRouting, routingNode, allocation);
if (decision.type() == Decision.Type.NO) {
logger.debug(
"[{}][{}] allocated on [{}], but can no longer be allocated on it, moving...",
shardRouting.index(),
shardRouting.id(),
routingNode.node());
boolean moved = shardsAllocators.move(shardRouting, routingNode, allocation);
if (!moved) {
logger.debug("[{}][{}] can't move", shardRouting.index(), shardRouting.id());
} else {
changed = true;
}
}
}
return changed;
}
@Override
public void execute(RoutingAllocation allocation) throws ElasticSearchException {
DiscoveryNode discoNode = allocation.nodes().resolveNode(node);
MutableShardRouting shardRouting = null;
for (MutableShardRouting routing : allocation.routingNodes().unassigned()) {
if (routing.shardId().equals(shardId)) {
// prefer primaries first to allocate
if (shardRouting == null || routing.primary()) {
shardRouting = routing;
}
}
}
if (shardRouting == null) {
throw new ElasticSearchIllegalArgumentException(
"[allocate] failed to find " + shardId + " on the list of unassigned shards");
}
if (shardRouting.primary() && !allowPrimary) {
throw new ElasticSearchIllegalArgumentException(
"[allocate] trying to allocate a primary shard " + shardId + "], which is disabled");
}
RoutingNode routingNode = allocation.routingNodes().node(discoNode.id());
allocation.addIgnoreDisable(shardRouting.shardId(), routingNode.nodeId());
if (!allocation.deciders().canAllocate(shardRouting, routingNode, allocation).allowed()) {
throw new ElasticSearchIllegalArgumentException(
"[allocate] allocation of " + shardId + " on node " + discoNode + " is not allowed");
}
// go over and remove it from the unassigned
for (Iterator<MutableShardRouting> it = allocation.routingNodes().unassigned().iterator();
it.hasNext(); ) {
if (it.next() != shardRouting) {
continue;
}
it.remove();
routingNode.add(shardRouting);
break;
}
}
@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 sameHost() {
AllocationService strategy =
new AllocationService(
settingsBuilder().put(SameShardAllocationDecider.SAME_HOST_SETTING, true).build());
MetaData metaData =
newMetaDataBuilder()
.put(newIndexMetaDataBuilder("test").numberOfShards(2).numberOfReplicas(1))
.build();
RoutingTable routingTable = routingTable().addAsNew(metaData.index("test")).build();
ClusterState clusterState =
newClusterStateBuilder().metaData(metaData).routingTable(routingTable).build();
logger.info("--> adding two nodes with the same host");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(
newNodesBuilder()
.put(newNode("node1", new InetSocketTransportAddress("test1", 80)))
.put(newNode("node2", new InetSocketTransportAddress("test1", 80))))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState.readOnlyRoutingNodes().numberOfShardsOfType(ShardRoutingState.INITIALIZING),
equalTo(2));
logger.info(
"--> start all primary shards, no replica will be started since its on the same host");
routingTable =
strategy
.applyStartedShards(
clusterState, clusterState.readOnlyRoutingNodes().shardsWithState(INITIALIZING))
.routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState.readOnlyRoutingNodes().numberOfShardsOfType(ShardRoutingState.STARTED),
equalTo(2));
assertThat(
clusterState.readOnlyRoutingNodes().numberOfShardsOfType(ShardRoutingState.INITIALIZING),
equalTo(0));
logger.info("--> add another node, with a different host, replicas will be allocating");
clusterState =
newClusterStateBuilder()
.state(clusterState)
.nodes(
newNodesBuilder()
.putAll(clusterState.nodes())
.put(newNode("node3", new InetSocketTransportAddress("test2", 80))))
.build();
routingTable = strategy.reroute(clusterState).routingTable();
clusterState = newClusterStateBuilder().state(clusterState).routingTable(routingTable).build();
assertThat(
clusterState.readOnlyRoutingNodes().numberOfShardsOfType(ShardRoutingState.STARTED),
equalTo(2));
assertThat(
clusterState.readOnlyRoutingNodes().numberOfShardsOfType(ShardRoutingState.INITIALIZING),
equalTo(2));
for (MutableShardRouting shardRouting :
clusterState.readOnlyRoutingNodes().shardsWithState(INITIALIZING)) {
assertThat(shardRouting.currentNodeId(), equalTo("node3"));
}
}
@Test
public void testMultiIndexEvenDistribution() {
AllocationService strategy =
createAllocationService(
settingsBuilder()
.put("cluster.routing.allocation.concurrent_recoveries", 10)
.put("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).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().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 (MutableShardRouting 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));
}
@Override
public void execute(RoutingAllocation allocation) throws ElasticSearchException {
DiscoveryNode discoNode = allocation.nodes().resolveNode(node);
boolean found = false;
for (RoutingNodes.RoutingNodeIterator it =
allocation.routingNodes().routingNodeIter(discoNode.id());
it.hasNext(); ) {
MutableShardRouting shardRouting = it.next();
if (!shardRouting.shardId().equals(shardId)) {
continue;
}
found = true;
if (shardRouting.relocatingNodeId() != null) {
if (shardRouting.initializing()) {
// the shard is initializing and recovering from another node, simply cancel the recovery
it.remove();
// and cancel the relocating state from the shard its being relocated from
RoutingNode relocatingFromNode =
allocation.routingNodes().node(shardRouting.relocatingNodeId());
if (relocatingFromNode != null) {
for (MutableShardRouting fromShardRouting : relocatingFromNode) {
if (fromShardRouting.shardId().equals(shardRouting.shardId())
&& fromShardRouting.state() == RELOCATING) {
allocation.routingNodes().cancelRelocation(fromShardRouting);
break;
}
}
}
} else if (shardRouting.relocating()) {
// the shard is relocating to another node, cancel the recovery on the other node, and
// deallocate this one
if (!allowPrimary && shardRouting.primary()) {
// can't cancel a primary shard being initialized
throw new ElasticSearchIllegalArgumentException(
"[cancel_allocation] can't cancel "
+ shardId
+ " on node "
+ discoNode
+ ", shard is primary and initializing its state");
}
it.moveToUnassigned();
// now, go and find the shard that is initializing on the target node, and cancel it as
// well...
RoutingNodes.RoutingNodeIterator initializingNode =
allocation.routingNodes().routingNodeIter(shardRouting.relocatingNodeId());
if (initializingNode != null) {
while (initializingNode.hasNext()) {
MutableShardRouting initializingShardRouting = initializingNode.next();
if (initializingShardRouting.shardId().equals(shardRouting.shardId())
&& initializingShardRouting.state() == INITIALIZING) {
initializingNode.remove();
}
}
}
}
} else {
// the shard is not relocating, its either started, or initializing, just cancel it and move
// on...
if (!allowPrimary && shardRouting.primary()) {
// can't cancel a primary shard being initialized
throw new ElasticSearchIllegalArgumentException(
"[cancel_allocation] can't cancel "
+ shardId
+ " on node "
+ discoNode
+ ", shard is primary and started");
}
it.remove();
allocation
.routingNodes()
.unassigned()
.add(
new MutableShardRouting(
shardRouting.index(),
shardRouting.id(),
null,
shardRouting.primary(),
ShardRoutingState.UNASSIGNED,
shardRouting.version() + 1));
}
}
if (!found) {
throw new ElasticSearchIllegalArgumentException(
"[cancel_allocation] can't cancel "
+ shardId
+ ", failed to find it on node "
+ discoNode);
}
}
/**
* Applies the relevant logic to handle a failed shard. Returns <tt>true</tt> if changes happened
* that require relocation.
*/
private boolean applyFailedShard(
RoutingAllocation allocation, ShardRouting failedShard, boolean addToIgnoreList) {
// create a copy of the failed shard, since we assume we can change possible references to it
// without
// changing the state of failed shard
failedShard = new ImmutableShardRouting(failedShard);
IndexRoutingTable indexRoutingTable = allocation.routingTable().index(failedShard.index());
if (indexRoutingTable == null) {
return false;
}
RoutingNodes routingNodes = allocation.routingNodes();
boolean dirty = false;
if (failedShard.relocatingNodeId() != null) {
// the shard is relocating, either in initializing (recovery from another node) or relocating
// (moving to another node)
if (failedShard.state() == INITIALIZING) {
// the shard is initializing and recovering from another node
// first, we need to cancel the current node that is being initialized
RoutingNodes.RoutingNodeIterator initializingNode =
routingNodes.routingNodeIter(failedShard.currentNodeId());
if (initializingNode != null) {
while (initializingNode.hasNext()) {
MutableShardRouting shardRouting = initializingNode.next();
if (shardRouting.equals(failedShard)) {
dirty = true;
initializingNode.remove();
if (addToIgnoreList) {
// make sure we ignore this shard on the relevant node
allocation.addIgnoreShardForNode(
failedShard.shardId(), failedShard.currentNodeId());
}
break;
}
}
}
if (dirty) {
// now, find the node that we are relocating *from*, and cancel its relocation
RoutingNode relocatingFromNode = routingNodes.node(failedShard.relocatingNodeId());
if (relocatingFromNode != null) {
for (MutableShardRouting shardRouting : relocatingFromNode) {
if (shardRouting.shardId().equals(failedShard.shardId())
&& shardRouting.relocating()) {
dirty = true;
routingNodes.cancelRelocation(shardRouting);
break;
}
}
}
} else {
logger.debug("failed shard {} not found in routingNodes, ignoring it", failedShard);
}
return dirty;
} else if (failedShard.state() == RELOCATING) {
// the shard is relocating, meaning its the source the shard is relocating from
// first, we need to cancel the current relocation from the current node
// now, find the node that we are recovering from, cancel the relocation, remove it from the
// node
// and add it to the unassigned shards list...
RoutingNodes.RoutingNodeIterator relocatingFromNode =
routingNodes.routingNodeIter(failedShard.currentNodeId());
if (relocatingFromNode != null) {
while (relocatingFromNode.hasNext()) {
MutableShardRouting shardRouting = relocatingFromNode.next();
if (shardRouting.equals(failedShard)) {
dirty = true;
relocatingFromNode.remove();
if (addToIgnoreList) {
// make sure we ignore this shard on the relevant node
allocation.addIgnoreShardForNode(
failedShard.shardId(), failedShard.currentNodeId());
}
routingNodes
.unassigned()
.add(
new MutableShardRouting(
failedShard.index(),
failedShard.id(),
null,
failedShard.primary(),
ShardRoutingState.UNASSIGNED,
failedShard.version() + 1));
break;
}
}
}
if (dirty) {
// next, we need to find the target initializing shard that is recovering from, and remove
// it...
RoutingNodes.RoutingNodeIterator initializingNode =
routingNodes.routingNodeIter(failedShard.relocatingNodeId());
if (initializingNode != null) {
while (initializingNode.hasNext()) {
MutableShardRouting shardRouting = initializingNode.next();
if (shardRouting.shardId().equals(failedShard.shardId())
&& shardRouting.state() == INITIALIZING) {
dirty = true;
initializingNode.remove();
}
}
}
} else {
logger.debug("failed shard {} not found in routingNodes, ignoring it", failedShard);
}
} else {
throw new ElasticsearchIllegalStateException(
"illegal state for a failed shard, relocating node id is set, but state does not match: "
+ failedShard);
}
} else {
// the shard is not relocating, its either started, or initializing, just cancel it and move
// on...
RoutingNodes.RoutingNodeIterator node =
routingNodes.routingNodeIter(failedShard.currentNodeId());
if (node != null) {
while (node.hasNext()) {
MutableShardRouting shardRouting = node.next();
if (shardRouting.equals(failedShard)) {
dirty = true;
if (addToIgnoreList) {
// make sure we ignore this shard on the relevant node
allocation.addIgnoreShardForNode(failedShard.shardId(), failedShard.currentNodeId());
}
node.remove();
// move all the shards matching the failed shard to the end of the unassigned list
// so we give a chance for other allocations and won't create poison failed allocations
// that can keep other shards from being allocated (because of limits applied on how
// many
// shards we can start per node)
List<MutableShardRouting> shardsToMove = Lists.newArrayList();
for (Iterator<MutableShardRouting> unassignedIt = routingNodes.unassigned().iterator();
unassignedIt.hasNext(); ) {
MutableShardRouting unassignedShardRouting = unassignedIt.next();
if (unassignedShardRouting.shardId().equals(failedShard.shardId())) {
unassignedIt.remove();
shardsToMove.add(unassignedShardRouting);
}
}
if (!shardsToMove.isEmpty()) {
routingNodes.unassigned().addAll(shardsToMove);
}
routingNodes
.unassigned()
.add(
new MutableShardRouting(
failedShard.index(),
failedShard.id(),
null,
null,
failedShard.restoreSource(),
failedShard.primary(),
ShardRoutingState.UNASSIGNED,
failedShard.version() + 1));
break;
}
}
}
if (!dirty) {
logger.debug("failed shard {} not found in routingNodes, ignoring it", failedShard);
}
}
return dirty;
}
private boolean electPrimariesAndUnassignedDanglingReplicas(RoutingAllocation allocation) {
boolean changed = false;
RoutingNodes routingNodes = allocation.routingNodes();
if (!routingNodes.hasUnassignedPrimaries()) {
// move out if we don't have unassigned primaries
return changed;
}
// go over and remove dangling replicas that are initializing for primary shards
List<ShardRouting> shardsToFail = Lists.newArrayList();
for (MutableShardRouting shardEntry : routingNodes.unassigned()) {
if (shardEntry.primary()) {
for (MutableShardRouting routing : routingNodes.assignedShards(shardEntry)) {
if (!routing.primary() && routing.initializing()) {
shardsToFail.add(routing);
}
}
}
}
for (ShardRouting shardToFail : shardsToFail) {
changed |= applyFailedShard(allocation, shardToFail, false);
}
// now, go over and elect a new primary if possible, not, from this code block on, if one is
// elected,
// routingNodes.hasUnassignedPrimaries() will potentially be false
for (MutableShardRouting shardEntry : routingNodes.unassigned()) {
if (shardEntry.primary()) {
MutableShardRouting candidate = allocation.routingNodes().activeReplica(shardEntry);
if (candidate != null) {
routingNodes.swapPrimaryFlag(shardEntry, candidate);
if (candidate.relocatingNodeId() != null) {
changed = true;
// its also relocating, make sure to move the other routing to primary
RoutingNode node = routingNodes.node(candidate.relocatingNodeId());
if (node != null) {
for (MutableShardRouting shardRouting : node) {
if (shardRouting.shardId().equals(candidate.shardId()) && !shardRouting.primary()) {
routingNodes.swapPrimaryFlag(shardRouting);
break;
}
}
}
}
}
}
}
return changed;
}
private ObjectLongOpenHashMap<DiscoveryNode> buildShardStates(
final DiscoveryNodes nodes, MutableShardRouting shard) {
ObjectLongOpenHashMap<DiscoveryNode> shardStates = cachedShardsState.get(shard.shardId());
ObjectOpenHashSet<String> nodeIds;
if (shardStates == null) {
shardStates = new ObjectLongOpenHashMap<>();
cachedShardsState.put(shard.shardId(), shardStates);
nodeIds = ObjectOpenHashSet.from(nodes.dataNodes().keys());
} else {
// clean nodes that have failed
shardStates
.keys()
.removeAll(
new ObjectPredicate<DiscoveryNode>() {
@Override
public boolean apply(DiscoveryNode node) {
return !nodes.nodeExists(node.id());
}
});
nodeIds = ObjectOpenHashSet.newInstance();
// we have stored cached from before, see if the nodes changed, if they have, go fetch again
for (ObjectCursor<DiscoveryNode> cursor : nodes.dataNodes().values()) {
DiscoveryNode node = cursor.value;
if (!shardStates.containsKey(node)) {
nodeIds.add(node.id());
}
}
}
if (nodeIds.isEmpty()) {
return shardStates;
}
String[] nodesIdsArray = nodeIds.toArray(String.class);
TransportNodesListGatewayStartedShards.NodesGatewayStartedShards response =
listGatewayStartedShards.list(shard.shardId(), nodesIdsArray, listTimeout).actionGet();
if (logger.isDebugEnabled()) {
if (response.failures().length > 0) {
StringBuilder sb =
new StringBuilder(shard + ": failures when trying to list shards on nodes:");
for (int i = 0; i < response.failures().length; i++) {
Throwable cause = ExceptionsHelper.unwrapCause(response.failures()[i]);
if (cause instanceof ConnectTransportException) {
continue;
}
sb.append("\n -> ").append(response.failures()[i].getDetailedMessage());
}
logger.debug(sb.toString());
}
}
for (TransportNodesListGatewayStartedShards.NodeGatewayStartedShards nodeShardState :
response) {
// -1 version means it does not exists, which is what the API returns, and what we expect to
logger.trace(
"[{}] on node [{}] has version [{}] of shard",
shard,
nodeShardState.getNode(),
nodeShardState.version());
shardStates.put(nodeShardState.getNode(), nodeShardState.version());
}
return shardStates;
}
public boolean allocateUnassigned(RoutingAllocation allocation) {
boolean changed = false;
DiscoveryNodes nodes = allocation.nodes();
RoutingNodes routingNodes = allocation.routingNodes();
// First, handle primaries, they must find a place to be allocated on here
Iterator<MutableShardRouting> unassignedIterator = routingNodes.unassigned().iterator();
while (unassignedIterator.hasNext()) {
MutableShardRouting shard = unassignedIterator.next();
if (!shard.primary()) {
continue;
}
// this is an API allocation, ignore since we know there is no data...
if (!routingNodes
.routingTable()
.index(shard.index())
.shard(shard.id())
.primaryAllocatedPostApi()) {
continue;
}
ObjectLongOpenHashMap<DiscoveryNode> nodesState = buildShardStates(nodes, shard);
int numberOfAllocationsFound = 0;
long highestVersion = -1;
Set<DiscoveryNode> nodesWithHighestVersion = Sets.newHashSet();
final boolean[] states = nodesState.allocated;
final Object[] keys = nodesState.keys;
final long[] values = nodesState.values;
for (int i = 0; i < states.length; i++) {
if (!states[i]) {
continue;
}
DiscoveryNode node = (DiscoveryNode) keys[i];
long version = values[i];
// since we don't check in NO allocation, we need to double check here
if (allocation.shouldIgnoreShardForNode(shard.shardId(), node.id())) {
continue;
}
if (version != -1) {
numberOfAllocationsFound++;
if (highestVersion == -1) {
nodesWithHighestVersion.add(node);
highestVersion = version;
} else {
if (version > highestVersion) {
nodesWithHighestVersion.clear();
nodesWithHighestVersion.add(node);
highestVersion = version;
} else if (version == highestVersion) {
nodesWithHighestVersion.add(node);
}
}
}
}
// check if the counts meets the minimum set
int requiredAllocation = 1;
// if we restore from a repository one copy is more then enough
if (shard.restoreSource() == null) {
try {
IndexMetaData indexMetaData = routingNodes.metaData().index(shard.index());
String initialShards =
indexMetaData
.settings()
.get(
INDEX_RECOVERY_INITIAL_SHARDS,
settings.get(INDEX_RECOVERY_INITIAL_SHARDS, this.initialShards));
if ("quorum".equals(initialShards)) {
if (indexMetaData.numberOfReplicas() > 1) {
requiredAllocation = ((1 + indexMetaData.numberOfReplicas()) / 2) + 1;
}
} else if ("quorum-1".equals(initialShards) || "half".equals(initialShards)) {
if (indexMetaData.numberOfReplicas() > 2) {
requiredAllocation = ((1 + indexMetaData.numberOfReplicas()) / 2);
}
} else if ("one".equals(initialShards)) {
requiredAllocation = 1;
} else if ("full".equals(initialShards) || "all".equals(initialShards)) {
requiredAllocation = indexMetaData.numberOfReplicas() + 1;
} else if ("full-1".equals(initialShards) || "all-1".equals(initialShards)) {
if (indexMetaData.numberOfReplicas() > 1) {
requiredAllocation = indexMetaData.numberOfReplicas();
}
} else {
requiredAllocation = Integer.parseInt(initialShards);
}
} catch (Exception e) {
logger.warn(
"[{}][{}] failed to derived initial_shards from value {}, ignore allocation for {}",
shard.index(),
shard.id(),
initialShards,
shard);
}
}
// not enough found for this shard, continue...
if (numberOfAllocationsFound < requiredAllocation) {
// if we are restoring this shard we still can allocate
if (shard.restoreSource() == null) {
// we can't really allocate, so ignore it and continue
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: not allocating, number_of_allocated_shards_found [{}], required_number [{}]",
shard.index(),
shard.id(),
numberOfAllocationsFound,
requiredAllocation);
}
} else if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: missing local data, will restore from [{}]",
shard.index(),
shard.id(),
shard.restoreSource());
}
continue;
}
Set<DiscoveryNode> throttledNodes = Sets.newHashSet();
Set<DiscoveryNode> noNodes = Sets.newHashSet();
for (DiscoveryNode discoNode : nodesWithHighestVersion) {
RoutingNode node = routingNodes.node(discoNode.id());
if (node == null) {
continue;
}
Decision decision = allocation.deciders().canAllocate(shard, node, allocation);
if (decision.type() == Decision.Type.THROTTLE) {
throttledNodes.add(discoNode);
} else if (decision.type() == Decision.Type.NO) {
noNodes.add(discoNode);
} else {
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: allocating [{}] to [{}] on primary allocation",
shard.index(),
shard.id(),
shard,
discoNode);
}
// we found a match
changed = true;
// make sure we create one with the version from the recovered state
allocation
.routingNodes()
.assign(new MutableShardRouting(shard, highestVersion), node.nodeId());
unassignedIterator.remove();
// found a node, so no throttling, no "no", and break out of the loop
throttledNodes.clear();
noNodes.clear();
break;
}
}
if (throttledNodes.isEmpty()) {
// if we have a node that we "can't" allocate to, force allocation, since this is our master
// data!
if (!noNodes.isEmpty()) {
DiscoveryNode discoNode = noNodes.iterator().next();
RoutingNode node = routingNodes.node(discoNode.id());
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: forcing allocating [{}] to [{}] on primary allocation",
shard.index(),
shard.id(),
shard,
discoNode);
}
// we found a match
changed = true;
// make sure we create one with the version from the recovered state
allocation
.routingNodes()
.assign(new MutableShardRouting(shard, highestVersion), node.nodeId());
unassignedIterator.remove();
}
} else {
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: throttling allocation [{}] to [{}] on primary allocation",
shard.index(),
shard.id(),
shard,
throttledNodes);
}
// we are throttling this, but we have enough to allocate to this node, ignore it for now
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
}
}
if (!routingNodes.hasUnassigned()) {
return changed;
}
// Now, handle replicas, try to assign them to nodes that are similar to the one the primary was
// allocated on
unassignedIterator = routingNodes.unassigned().iterator();
while (unassignedIterator.hasNext()) {
MutableShardRouting shard = unassignedIterator.next();
// pre-check if it can be allocated to any node that currently exists, so we won't list the
// store for it for nothing
boolean canBeAllocatedToAtLeastOneNode = false;
for (ObjectCursor<DiscoveryNode> cursor : nodes.dataNodes().values()) {
RoutingNode node = routingNodes.node(cursor.value.id());
if (node == null) {
continue;
}
// if we can't allocate it on a node, ignore it, for example, this handles
// cases for only allocating a replica after a primary
Decision decision = allocation.deciders().canAllocate(shard, node, allocation);
if (decision.type() == Decision.Type.YES) {
canBeAllocatedToAtLeastOneNode = true;
break;
}
}
if (!canBeAllocatedToAtLeastOneNode) {
continue;
}
Map<DiscoveryNode, TransportNodesListShardStoreMetaData.StoreFilesMetaData> shardStores =
buildShardStores(nodes, shard);
long lastSizeMatched = 0;
DiscoveryNode lastDiscoNodeMatched = null;
RoutingNode lastNodeMatched = null;
for (Map.Entry<DiscoveryNode, TransportNodesListShardStoreMetaData.StoreFilesMetaData>
nodeStoreEntry : shardStores.entrySet()) {
DiscoveryNode discoNode = nodeStoreEntry.getKey();
TransportNodesListShardStoreMetaData.StoreFilesMetaData storeFilesMetaData =
nodeStoreEntry.getValue();
logger.trace("{}: checking node [{}]", shard, discoNode);
if (storeFilesMetaData == null) {
// already allocated on that node...
continue;
}
RoutingNode node = routingNodes.node(discoNode.id());
if (node == null) {
continue;
}
// check if we can allocate on that node...
// we only check for NO, since if this node is THROTTLING and it has enough "same data"
// then we will try and assign it next time
Decision decision = allocation.deciders().canAllocate(shard, node, allocation);
if (decision.type() == Decision.Type.NO) {
continue;
}
// if it is already allocated, we can't assign to it...
if (storeFilesMetaData.allocated()) {
continue;
}
if (!shard.primary()) {
MutableShardRouting primaryShard = routingNodes.activePrimary(shard);
if (primaryShard != null) {
assert primaryShard.active();
DiscoveryNode primaryNode = nodes.get(primaryShard.currentNodeId());
if (primaryNode != null) {
TransportNodesListShardStoreMetaData.StoreFilesMetaData primaryNodeStore =
shardStores.get(primaryNode);
if (primaryNodeStore != null && primaryNodeStore.allocated()) {
long sizeMatched = 0;
for (StoreFileMetaData storeFileMetaData : storeFilesMetaData) {
if (primaryNodeStore.fileExists(storeFileMetaData.name())
&& primaryNodeStore
.file(storeFileMetaData.name())
.isSame(storeFileMetaData)) {
sizeMatched += storeFileMetaData.length();
}
}
logger.trace(
"{}: node [{}] has [{}/{}] bytes of re-usable data",
shard,
discoNode.name(),
new ByteSizeValue(sizeMatched),
sizeMatched);
if (sizeMatched > lastSizeMatched) {
lastSizeMatched = sizeMatched;
lastDiscoNodeMatched = discoNode;
lastNodeMatched = node;
}
}
}
}
}
}
if (lastNodeMatched != null) {
// we only check on THROTTLE since we checked before before on NO
Decision decision = allocation.deciders().canAllocate(shard, lastNodeMatched, allocation);
if (decision.type() == Decision.Type.THROTTLE) {
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: throttling allocation [{}] to [{}] in order to reuse its unallocated persistent store with total_size [{}]",
shard.index(),
shard.id(),
shard,
lastDiscoNodeMatched,
new ByteSizeValue(lastSizeMatched));
}
// we are throttling this, but we have enough to allocate to this node, ignore it for now
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
} else {
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}]: allocating [{}] to [{}] in order to reuse its unallocated persistent store with total_size [{}]",
shard.index(),
shard.id(),
shard,
lastDiscoNodeMatched,
new ByteSizeValue(lastSizeMatched));
}
// we found a match
changed = true;
allocation.routingNodes().assign(shard, lastNodeMatched.nodeId());
unassignedIterator.remove();
}
}
}
return changed;
}
