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;
}
@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"));
}
}
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;
}
