/** Can the shard be allocated on at least one node based on the allocation deciders. */
private boolean canBeAllocatedToAtLeastOneNode(ShardRouting shard, RoutingAllocation allocation) {
for (ObjectCursor<DiscoveryNode> cursor : allocation.nodes().dataNodes().values()) {
RoutingNode node = allocation.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) {
return true;
}
}
return false;
}
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;
}
public boolean allocateUnassigned(RoutingAllocation allocation) {
boolean changed = false;
final RoutingNodes routingNodes = allocation.routingNodes();
final MetaData metaData = routingNodes.metaData();
final RoutingNodes.UnassignedShards.UnassignedIterator unassignedIterator =
routingNodes.unassigned().iterator();
while (unassignedIterator.hasNext()) {
ShardRouting shard = unassignedIterator.next();
if (shard.primary()) {
continue;
}
// 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 : allocation.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) {
logger.trace("{}: ignoring allocation, can't be allocated on any node", shard);
unassignedIterator.removeAndIgnore();
continue;
}
AsyncShardFetch.FetchResult<TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
shardStores = fetchData(shard, allocation);
if (shardStores.hasData() == false) {
logger.trace("{}: ignoring allocation, still fetching shard stores", shard);
unassignedIterator.removeAndIgnore();
continue; // still fetching
}
long lastSizeMatched = 0;
DiscoveryNode lastDiscoNodeMatched = null;
RoutingNode lastNodeMatched = null;
boolean hasReplicaData = false;
IndexMetaData indexMetaData = metaData.index(shard.getIndex());
for (Map.Entry<DiscoveryNode, TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
nodeStoreEntry : shardStores.getData().entrySet()) {
DiscoveryNode discoNode = nodeStoreEntry.getKey();
TransportNodesListShardStoreMetaData.StoreFilesMetaData storeFilesMetaData =
nodeStoreEntry.getValue().storeFilesMetaData();
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()) {
hasReplicaData |= storeFilesMetaData.iterator().hasNext();
ShardRouting primaryShard = routingNodes.activePrimary(shard);
if (primaryShard != null) {
assert primaryShard.active();
DiscoveryNode primaryNode = allocation.nodes().get(primaryShard.currentNodeId());
if (primaryNode != null) {
TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData primaryNodeFilesStore =
shardStores.getData().get(primaryNode);
if (primaryNodeFilesStore != null) {
TransportNodesListShardStoreMetaData.StoreFilesMetaData primaryNodeStore =
primaryNodeFilesStore.storeFilesMetaData();
if (primaryNodeStore != null && primaryNodeStore.allocated()) {
long sizeMatched = 0;
String primarySyncId = primaryNodeStore.syncId();
String replicaSyncId = storeFilesMetaData.syncId();
// see if we have a sync id we can make use of
if (replicaSyncId != null && replicaSyncId.equals(primarySyncId)) {
logger.trace(
"{}: node [{}] has same sync id {} as primary",
shard,
discoNode.name(),
replicaSyncId);
lastNodeMatched = node;
lastSizeMatched = Long.MAX_VALUE;
lastDiscoNodeMatched = discoNode;
} else {
for (StoreFileMetaData storeFileMetaData : storeFilesMetaData) {
String metaDataFileName = storeFileMetaData.name();
if (primaryNodeStore.fileExists(metaDataFileName)
&& primaryNodeStore.file(metaDataFileName).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.removeAndIgnore();
} 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;
unassignedIterator.initialize(lastNodeMatched.nodeId());
}
} else if (hasReplicaData == false) {
// if we didn't manage to find *any* data (regardless of matching sizes), check if the
// allocation
// of the replica shard needs to be delayed, and if so, add it to the ignore unassigned list
// note: we only care about replica in delayed allocation, since if we have an unassigned
// primary it
// will anyhow wait to find an existing copy of the shard to be allocated
// note: the other side of the equation is scheduling a reroute in a timely manner, which
// happens in the RoutingService
long delay =
shard
.unassignedInfo()
.getDelayAllocationExpirationIn(settings, indexMetaData.getSettings());
if (delay > 0) {
logger.debug(
"[{}][{}]: delaying allocation of [{}] for [{}]",
shard.index(),
shard.id(),
shard,
TimeValue.timeValueMillis(delay));
/**
* mark it as changed, since we want to kick a publishing to schedule future allocation,
* see {@link
* org.elasticsearch.cluster.routing.RoutingService#clusterChanged(ClusterChangedEvent)}).
*/
changed = true;
unassignedIterator.removeAndIgnore();
}
}
}
return changed;
}
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
final MetaData metaData = routingNodes.metaData();
RoutingNodes.UnassignedShards unassigned = routingNodes.unassigned();
unassigned.sort(
new PriorityComparator() {
@Override
protected Settings getIndexSettings(String index) {
IndexMetaData indexMetaData = metaData.index(index);
return indexMetaData.getSettings();
}
}); // sort for priority ordering
Iterator<ShardRouting> unassignedIterator = unassigned.iterator();
while (unassignedIterator.hasNext()) {
ShardRouting 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;
}
AsyncShardFetch<TransportNodesListGatewayStartedShards.NodeGatewayStartedShards> fetch =
asyncFetchStarted.get(shard.shardId());
if (fetch == null) {
fetch = new InternalAsyncFetch<>(logger, "shard_started", shard.shardId(), startedAction);
asyncFetchStarted.put(shard.shardId(), fetch);
}
AsyncShardFetch.FetchResult<TransportNodesListGatewayStartedShards.NodeGatewayStartedShards>
shardState = fetch.fetchData(nodes, metaData, allocation.getIgnoreNodes(shard.shardId()));
if (shardState.hasData() == false) {
logger.trace("{}: ignoring allocation, still fetching shard started state", shard);
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
continue;
}
shardState.processAllocation(allocation);
IndexMetaData indexMetaData = metaData.index(shard.getIndex());
/**
* Build a map of DiscoveryNodes to shard state number for the given shard. A state of -1
* means the shard does not exist on the node, where any shard state >= 0 is the state version
* of the shard on that node's disk.
*
* <p>A shard on shared storage will return at least shard state 0 for all nodes, indicating
* that the shard can be allocated to any node.
*/
ObjectLongHashMap<DiscoveryNode> nodesState = new ObjectLongHashMap<>();
for (TransportNodesListGatewayStartedShards.NodeGatewayStartedShards nodeShardState :
shardState.getData().values()) {
long version = nodeShardState.version();
// -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(),
version);
nodesState.put(nodeShardState.getNode(), version);
}
int numberOfAllocationsFound = 0;
long highestVersion = -1;
final Map<DiscoveryNode, Long> nodesWithVersion = Maps.newHashMap();
assert !nodesState.containsKey(null);
final Object[] keys = nodesState.keys;
final long[] values = nodesState.values;
Settings idxSettings = indexMetaData.settings();
for (int i = 0; i < keys.length; i++) {
if (keys[i] == null) {
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 (recoverOnAnyNode(idxSettings)) {
numberOfAllocationsFound++;
if (version > highestVersion) {
highestVersion = version;
}
// We always put the node without clearing the map
nodesWithVersion.put(node, version);
} else if (version != -1) {
numberOfAllocationsFound++;
// If we've found a new "best" candidate, clear the
// current candidates and add it
if (version > highestVersion) {
highestVersion = version;
nodesWithVersion.clear();
nodesWithVersion.put(node, version);
} else if (version == highestVersion) {
// If the candidate is the same, add it to the
// list, but keep the current candidate
nodesWithVersion.put(node, version);
}
}
}
// Now that we have a map of nodes to versions along with the
// number of allocations found (and not ignored), we need to sort
// it so the node with the highest version is at the beginning
List<DiscoveryNode> nodesWithHighestVersion = Lists.newArrayList();
nodesWithHighestVersion.addAll(nodesWithVersion.keySet());
CollectionUtil.timSort(
nodesWithHighestVersion,
new Comparator<DiscoveryNode>() {
@Override
public int compare(DiscoveryNode o1, DiscoveryNode o2) {
return Long.compare(nodesWithVersion.get(o2), nodesWithVersion.get(o1));
}
});
if (logger.isDebugEnabled()) {
logger.debug(
"[{}][{}] found {} allocations of {}, highest version: [{}]",
shard.index(),
shard.id(),
numberOfAllocationsFound,
shard,
highestVersion);
}
if (logger.isTraceEnabled()) {
StringBuilder sb = new StringBuilder("[");
for (DiscoveryNode n : nodesWithHighestVersion) {
sb.append("[");
sb.append(n.getName());
sb.append("]");
sb.append(" -> ");
sb.append(nodesWithVersion.get(n));
sb.append(", ");
}
sb.append("]");
logger.trace("{} candidates for allocation: {}", shard, sb.toString());
}
// 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 {
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
routingNodes.initialize(new ShardRouting(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
routingNodes.initialize(new ShardRouting(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 = unassigned.iterator();
while (unassignedIterator.hasNext()) {
ShardRouting shard = unassignedIterator.next();
if (shard.primary()) {
continue;
}
// 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) {
logger.trace("{}: ignoring allocation, can't be allocated on any node", shard);
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
continue;
}
AsyncShardFetch<TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData> fetch =
asyncFetchStore.get(shard.shardId());
if (fetch == null) {
fetch = new InternalAsyncFetch<>(logger, "shard_store", shard.shardId(), storeAction);
asyncFetchStore.put(shard.shardId(), fetch);
}
AsyncShardFetch.FetchResult<TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
shardStores =
fetch.fetchData(nodes, metaData, allocation.getIgnoreNodes(shard.shardId()));
if (shardStores.hasData() == false) {
logger.trace("{}: ignoring allocation, still fetching shard stores", shard);
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
continue; // still fetching
}
shardStores.processAllocation(allocation);
long lastSizeMatched = 0;
DiscoveryNode lastDiscoNodeMatched = null;
RoutingNode lastNodeMatched = null;
boolean hasReplicaData = false;
IndexMetaData indexMetaData = metaData.index(shard.getIndex());
for (Map.Entry<DiscoveryNode, TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
nodeStoreEntry : shardStores.getData().entrySet()) {
DiscoveryNode discoNode = nodeStoreEntry.getKey();
TransportNodesListShardStoreMetaData.StoreFilesMetaData storeFilesMetaData =
nodeStoreEntry.getValue().storeFilesMetaData();
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()) {
hasReplicaData |= storeFilesMetaData.iterator().hasNext();
ShardRouting primaryShard = routingNodes.activePrimary(shard);
if (primaryShard != null) {
assert primaryShard.active();
DiscoveryNode primaryNode = nodes.get(primaryShard.currentNodeId());
if (primaryNode != null) {
TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData primaryNodeFilesStore =
shardStores.getData().get(primaryNode);
if (primaryNodeFilesStore != null) {
TransportNodesListShardStoreMetaData.StoreFilesMetaData primaryNodeStore =
primaryNodeFilesStore.storeFilesMetaData();
if (primaryNodeStore != null && primaryNodeStore.allocated()) {
long sizeMatched = 0;
String primarySyncId = primaryNodeStore.syncId();
String replicaSyncId = storeFilesMetaData.syncId();
// see if we have a sync id we can make use of
if (replicaSyncId != null && replicaSyncId.equals(primarySyncId)) {
logger.trace(
"{}: node [{}] has same sync id {} as primary",
shard,
discoNode.name(),
replicaSyncId);
lastNodeMatched = node;
lastSizeMatched = Long.MAX_VALUE;
lastDiscoNodeMatched = discoNode;
} else {
for (StoreFileMetaData storeFileMetaData : storeFilesMetaData) {
String metaDataFileName = storeFileMetaData.name();
if (primaryNodeStore.fileExists(metaDataFileName)
&& primaryNodeStore.file(metaDataFileName).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;
routingNodes.initialize(shard, lastNodeMatched.nodeId());
unassignedIterator.remove();
}
} else if (hasReplicaData == false) {
// if we didn't manage to find *any* data (regardless of matching sizes), check if the
// allocation
// of the replica shard needs to be delayed, and if so, add it to the ignore unassigned list
// note: we only care about replica in delayed allocation, since if we have an unassigned
// primary it
// will anyhow wait to find an existing copy of the shard to be allocated
// note: the other side of the equation is scheduling a reroute in a timely manner, which
// happens in the RoutingService
long delay =
shard
.unassignedInfo()
.getDelayAllocationExpirationIn(settings, indexMetaData.getSettings());
if (delay > 0) {
logger.debug(
"[{}][{}]: delaying allocation of [{}] for [{}]",
shard.index(),
shard.id(),
shard,
TimeValue.timeValueMillis(delay));
/**
* mark it as changed, since we want to kick a publishing to schedule future allocation,
* see {@link
* org.elasticsearch.cluster.routing.RoutingService#clusterChanged(ClusterChangedEvent)}).
*/
changed = true;
unassignedIterator.remove();
routingNodes.ignoredUnassigned().add(shard);
}
}
}
return changed;
}
private MatchingNodes findMatchingNodes(
ShardRouting shard,
RoutingAllocation allocation,
TransportNodesListShardStoreMetaData.StoreFilesMetaData primaryStore,
AsyncShardFetch.FetchResult<TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
data) {
ObjectLongMap<DiscoveryNode> nodesToSize = new ObjectLongHashMap<>();
for (Map.Entry<DiscoveryNode, TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
nodeStoreEntry : data.getData().entrySet()) {
DiscoveryNode discoNode = nodeStoreEntry.getKey();
TransportNodesListShardStoreMetaData.StoreFilesMetaData storeFilesMetaData =
nodeStoreEntry.getValue().storeFilesMetaData();
if (storeFilesMetaData == null) {
// already allocated on that node...
continue;
}
RoutingNode node = allocation.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... (and it might be primary as well)
if (storeFilesMetaData.allocated()) {
continue;
}
// we don't have any files at all, it is an empty index
if (storeFilesMetaData.iterator().hasNext() == false) {
continue;
}
String primarySyncId = primaryStore.syncId();
String replicaSyncId = storeFilesMetaData.syncId();
// see if we have a sync id we can make use of
if (replicaSyncId != null && replicaSyncId.equals(primarySyncId)) {
logger.trace(
"{}: node [{}] has same sync id {} as primary", shard, discoNode.name(), replicaSyncId);
nodesToSize.put(discoNode, Long.MAX_VALUE);
} else {
long sizeMatched = 0;
for (StoreFileMetaData storeFileMetaData : storeFilesMetaData) {
String metaDataFileName = storeFileMetaData.name();
if (primaryStore.fileExists(metaDataFileName)
&& primaryStore.file(metaDataFileName).isSame(storeFileMetaData)) {
sizeMatched += storeFileMetaData.length();
}
}
logger.trace(
"{}: node [{}] has [{}/{}] bytes of re-usable data",
shard,
discoNode.name(),
new ByteSizeValue(sizeMatched),
sizeMatched);
nodesToSize.put(discoNode, sizeMatched);
}
}
return new MatchingNodes(nodesToSize);
}
public boolean allocateUnassigned(RoutingAllocation allocation) {
boolean changed = false;
final RoutingNodes routingNodes = allocation.routingNodes();
final RoutingNodes.UnassignedShards.UnassignedIterator unassignedIterator =
routingNodes.unassigned().iterator();
MetaData metaData = allocation.metaData();
while (unassignedIterator.hasNext()) {
ShardRouting shard = unassignedIterator.next();
if (shard.primary()) {
continue;
}
// if we are allocating a replica because of index creation, no need to go and find a copy,
// there isn't one...
IndexMetaData indexMetaData = metaData.index(shard.getIndexName());
if (shard.allocatedPostIndexCreate(indexMetaData) == false) {
continue;
}
// pre-check if it can be allocated to any node that currently exists, so we won't list the
// store for it for nothing
if (canBeAllocatedToAtLeastOneNode(shard, allocation) == false) {
logger.trace("{}: ignoring allocation, can't be allocated on any node", shard);
unassignedIterator.removeAndIgnore();
continue;
}
AsyncShardFetch.FetchResult<TransportNodesListShardStoreMetaData.NodeStoreFilesMetaData>
shardStores = fetchData(shard, allocation);
if (shardStores.hasData() == false) {
logger.trace("{}: ignoring allocation, still fetching shard stores", shard);
allocation.setHasPendingAsyncFetch();
unassignedIterator.removeAndIgnore();
continue; // still fetching
}
ShardRouting primaryShard = routingNodes.activePrimary(shard);
assert primaryShard != null
: "the replica shard can be allocated on at least one node, so there must be an active primary";
TransportNodesListShardStoreMetaData.StoreFilesMetaData primaryStore =
findStore(primaryShard, allocation, shardStores);
if (primaryStore == null || primaryStore.allocated() == false) {
// if we can't find the primary data, it is probably because the primary shard is corrupted
// (and listing failed)
// we want to let the replica be allocated in order to expose the actual problem with the
// primary that the replica
// will try and recover from
// Note, this is the existing behavior, as exposed in running
// CorruptFileTest#testNoPrimaryData
logger.trace(
"{}: no primary shard store found or allocated, letting actual allocation figure it out",
shard);
continue;
}
MatchingNodes matchingNodes = findMatchingNodes(shard, allocation, primaryStore, shardStores);
if (matchingNodes.getNodeWithHighestMatch() != null) {
RoutingNode nodeWithHighestMatch =
allocation.routingNodes().node(matchingNodes.getNodeWithHighestMatch().id());
// we only check on THROTTLE since we checked before before on NO
Decision decision =
allocation.deciders().canAllocate(shard, nodeWithHighestMatch, allocation);
if (decision.type() == Decision.Type.THROTTLE) {
logger.debug(
"[{}][{}]: throttling allocation [{}] to [{}] in order to reuse its unallocated persistent store",
shard.index(),
shard.id(),
shard,
nodeWithHighestMatch.node());
// we are throttling this, but we have enough to allocate to this node, ignore it for now
unassignedIterator.removeAndIgnore();
} else {
logger.debug(
"[{}][{}]: allocating [{}] to [{}] in order to reuse its unallocated persistent store",
shard.index(),
shard.id(),
shard,
nodeWithHighestMatch.node());
// we found a match
changed = true;
unassignedIterator.initialize(
nodeWithHighestMatch.nodeId(),
null,
allocation
.clusterInfo()
.getShardSize(shard, ShardRouting.UNAVAILABLE_EXPECTED_SHARD_SIZE));
}
} else if (matchingNodes.hasAnyData() == false) {
// if we didn't manage to find *any* data (regardless of matching sizes), check if the
// allocation of the replica shard needs to be delayed
changed |= ignoreUnassignedIfDelayed(unassignedIterator, shard);
}
}
return changed;
}
@Override
public RerouteExplanation execute(RoutingAllocation allocation, boolean explain) {
final DiscoveryNode discoNode = allocation.nodes().resolveNode(node);
final RoutingNodes routingNodes = allocation.routingNodes();
ShardRouting shardRouting = null;
for (ShardRouting routing : routingNodes.unassigned()) {
if (routing.shardId().equals(shardId)) {
// prefer primaries first to allocate
if (shardRouting == null || routing.primary()) {
shardRouting = routing;
}
}
}
if (shardRouting == null) {
if (explain) {
return new RerouteExplanation(
this,
allocation.decision(
Decision.NO,
"allocate_allocation_command",
"failed to find " + shardId + " on the list of unassigned shards"));
}
throw new IllegalArgumentException(
"[allocate] failed to find " + shardId + " on the list of unassigned shards");
}
if (shardRouting.primary() && !allowPrimary) {
if (explain) {
return new RerouteExplanation(
this,
allocation.decision(
Decision.NO,
"allocate_allocation_command",
"trying to allocate a primary shard " + shardId + ", which is disabled"));
}
throw new IllegalArgumentException(
"[allocate] trying to allocate a primary shard " + shardId + ", which is disabled");
}
RoutingNode routingNode = routingNodes.node(discoNode.id());
if (routingNode == null) {
if (!discoNode.dataNode()) {
if (explain) {
return new RerouteExplanation(
this,
allocation.decision(
Decision.NO,
"allocate_allocation_command",
"Allocation can only be done on data nodes, not [" + node + "]"));
}
throw new IllegalArgumentException(
"Allocation can only be done on data nodes, not [" + node + "]");
} else {
if (explain) {
return new RerouteExplanation(
this,
allocation.decision(
Decision.NO,
"allocate_allocation_command",
"Could not find [" + node + "] among the routing nodes"));
}
throw new IllegalStateException("Could not find [" + node + "] among the routing nodes");
}
}
Decision decision = allocation.deciders().canAllocate(shardRouting, routingNode, allocation);
if (decision.type() == Decision.Type.NO) {
if (explain) {
return new RerouteExplanation(this, decision);
}
throw new IllegalArgumentException(
"[allocate] allocation of "
+ shardId
+ " on node "
+ discoNode
+ " is not allowed, reason: "
+ decision);
}
// go over and remove it from the unassigned
for (RoutingNodes.UnassignedShards.UnassignedIterator it = routingNodes.unassigned().iterator();
it.hasNext(); ) {
if (it.next() != shardRouting) {
continue;
}
it.initialize(routingNode.nodeId());
if (shardRouting.primary()) {
// we need to clear the post allocation flag, since its an explicit allocation of the
// primary shard
// and we want to force allocate it (and create a new index for it)
routingNodes.addClearPostAllocationFlag(shardRouting.shardId());
}
break;
}
return new RerouteExplanation(this, decision);
}
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;
}