private void sendPartitions() {
ClusterNode oldestNode = this.oldestNode.get();
try {
sendLocalPartitions(oldestNode, exchId);
} catch (ClusterTopologyCheckedException ignore) {
if (log.isDebugEnabled())
log.debug(
"Oldest node left during partition exchange [nodeId="
+ oldestNode.id()
+ ", exchId="
+ exchId
+ ']');
} catch (IgniteCheckedException e) {
scheduleRecheck();
U.error(
log,
"Failed to send local partitions to oldest node (will retry after timeout) [oldestNodeId="
+ oldestNode.id()
+ ", exchId="
+ exchId
+ ']',
e);
}
}
/**
* Sends query request.
*
* @param fut Distributed future.
* @param req Request.
* @param nodes Nodes.
* @throws IgniteCheckedException In case of error.
*/
@SuppressWarnings("unchecked")
private void sendRequest(
final GridCacheDistributedQueryFuture<?, ?, ?> fut,
final GridCacheQueryRequest req,
Collection<ClusterNode> nodes)
throws IgniteCheckedException {
assert fut != null;
assert req != null;
assert nodes != null;
final UUID locNodeId = cctx.localNodeId();
ClusterNode locNode = null;
Collection<ClusterNode> rmtNodes = null;
for (ClusterNode n : nodes) {
if (n.id().equals(locNodeId)) locNode = n;
else {
if (rmtNodes == null) rmtNodes = new ArrayList<>(nodes.size());
rmtNodes.add(n);
}
}
// Request should be sent to remote nodes before the query is processed on the local node.
// For example, a remote reducer has a state, we should not serialize and then send
// the reducer changed by the local node.
if (!F.isEmpty(rmtNodes)) {
cctx.io()
.safeSend(
rmtNodes,
req,
cctx.ioPolicy(),
new P1<ClusterNode>() {
@Override
public boolean apply(ClusterNode node) {
fut.onNodeLeft(node.id());
return !fut.isDone();
}
});
}
if (locNode != null) {
cctx.closures()
.callLocalSafe(
new Callable<Object>() {
@Override
public Object call() throws Exception {
req.beforeLocalExecution(cctx);
processQueryRequest(locNodeId, req);
return null;
}
});
}
}
/** @param nodeId Node to remove. */
private void removeNode(UUID nodeId) {
assert nodeId != null;
assert lock.writeLock().isHeldByCurrentThread();
ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);
assert oldest != null;
ClusterNode loc = cctx.localNode();
if (node2part != null) {
if (oldest.equals(loc) && !node2part.nodeId().equals(loc.id())) {
updateSeq.setIfGreater(node2part.updateSequence());
node2part =
new GridDhtPartitionFullMap(
loc.id(), loc.order(), updateSeq.incrementAndGet(), node2part, false);
} else node2part = new GridDhtPartitionFullMap(node2part, node2part.updateSequence());
part2node = new HashMap<>(part2node);
GridDhtPartitionMap parts = node2part.remove(nodeId);
if (parts != null) {
for (Integer p : parts.keySet()) {
Set<UUID> nodeIds = part2node.get(p);
if (nodeIds != null) {
nodeIds.remove(nodeId);
if (nodeIds.isEmpty()) part2node.remove(p);
}
}
}
consistencyCheck();
}
}
/** {@inheritDoc} */
@Override
public String toString() {
ClusterNode oldestNode = this.oldestNode.get();
return S.toString(
GridDhtPartitionsExchangeFuture.class,
this,
"oldest",
oldestNode == null ? "null" : oldestNode.id(),
"oldestOrder",
oldestNode == null ? "null" : oldestNode.order(),
"evtLatch",
evtLatch == null ? "null" : evtLatch.getCount(),
"remaining",
remaining(),
"super",
super.toString());
}
/**
* Send delete message to all meta cache nodes in the grid.
*
* @param msg Message to send.
*/
private void sendDeleteMessage(IgfsDeleteMessage msg) {
assert msg != null;
Collection<ClusterNode> nodes = meta.metaCacheNodes();
for (ClusterNode node : nodes) {
try {
igfsCtx.send(node, topic, msg, GridIoPolicy.SYSTEM_POOL);
} catch (IgniteCheckedException e) {
U.warn(
log,
"Failed to send IGFS delete message to node [nodeId="
+ node.id()
+ ", msg="
+ msg
+ ", err="
+ e.getMessage()
+ ']');
}
}
}
/** {@inheritDoc} */
@Override
public boolean own(GridDhtLocalPartition part) {
ClusterNode loc = cctx.localNode();
lock.writeLock().lock();
try {
if (part.own()) {
updateLocal(part.id(), loc.id(), part.state(), updateSeq.incrementAndGet());
consistencyCheck();
return true;
}
consistencyCheck();
return false;
} finally {
lock.writeLock().unlock();
}
}
/**
* @param node Node.
* @param id ID.
* @throws IgniteCheckedException If failed.
*/
private void sendLocalPartitions(ClusterNode node, @Nullable GridDhtPartitionExchangeId id)
throws IgniteCheckedException {
GridDhtPartitionsSingleMessage m =
new GridDhtPartitionsSingleMessage(
id, cctx.kernalContext().clientNode(), cctx.versions().last());
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (!cacheCtx.isLocal())
m.addLocalPartitionMap(cacheCtx.cacheId(), cacheCtx.topology().localPartitionMap());
}
if (log.isDebugEnabled())
log.debug(
"Sending local partitions [nodeId="
+ node.id()
+ ", exchId="
+ exchId
+ ", msg="
+ m
+ ']');
cctx.io().send(node, m, SYSTEM_POOL);
}
/** {@inheritDoc} */
@Override
public boolean afterExchange(GridDhtPartitionsExchangeFuture exchFut)
throws IgniteCheckedException {
boolean changed = waitForRent();
ClusterNode loc = cctx.localNode();
int num = cctx.affinity().partitions();
AffinityTopologyVersion topVer = exchFut.topologyVersion();
lock.writeLock().lock();
try {
if (stopping) return false;
assert topVer.equals(exchFut.topologyVersion())
: "Invalid topology version [topVer=" + topVer + ", exchId=" + exchFut.exchangeId() + ']';
if (log.isDebugEnabled())
log.debug(
"Partition map before afterExchange [exchId="
+ exchFut.exchangeId()
+ ", fullMap="
+ fullMapString()
+ ']');
long updateSeq = this.updateSeq.incrementAndGet();
for (int p = 0; p < num; p++) {
GridDhtLocalPartition locPart = localPartition(p, topVer, false, false);
if (cctx.affinity().localNode(p, topVer)) {
// This partition will be created during next topology event,
// which obviously has not happened at this point.
if (locPart == null) {
if (log.isDebugEnabled())
log.debug("Skipping local partition afterExchange (will not create): " + p);
continue;
}
GridDhtPartitionState state = locPart.state();
if (state == MOVING) {
if (cctx.rebalanceEnabled()) {
Collection<ClusterNode> owners = owners(p);
// If there are no other owners, then become an owner.
if (F.isEmpty(owners)) {
boolean owned = locPart.own();
assert owned
: "Failed to own partition [cacheName"
+ cctx.name()
+ ", locPart="
+ locPart
+ ']';
updateLocal(p, loc.id(), locPart.state(), updateSeq);
changed = true;
if (cctx.events().isRecordable(EVT_CACHE_REBALANCE_PART_DATA_LOST)) {
DiscoveryEvent discoEvt = exchFut.discoveryEvent();
cctx.events()
.addPreloadEvent(
p,
EVT_CACHE_REBALANCE_PART_DATA_LOST,
discoEvt.eventNode(),
discoEvt.type(),
discoEvt.timestamp());
}
if (log.isDebugEnabled()) log.debug("Owned partition: " + locPart);
} else if (log.isDebugEnabled())
log.debug(
"Will not own partition (there are owners to rebalance from) [locPart="
+ locPart
+ ", owners = "
+ owners
+ ']');
} else updateLocal(p, loc.id(), locPart.state(), updateSeq);
}
} else {
if (locPart != null) {
GridDhtPartitionState state = locPart.state();
if (state == MOVING) {
locPart.rent(false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug("Evicting moving partition (it does not belong to affinity): " + locPart);
}
}
}
}
consistencyCheck();
} finally {
lock.writeLock().unlock();
}
return changed;
}
/** {@inheritDoc} */
@Override
public void beforeExchange(GridDhtPartitionsExchangeFuture exchFut)
throws IgniteCheckedException {
waitForRent();
ClusterNode loc = cctx.localNode();
int num = cctx.affinity().partitions();
lock.writeLock().lock();
try {
GridDhtPartitionExchangeId exchId = exchFut.exchangeId();
if (stopping) return;
assert topVer.equals(exchId.topologyVersion())
: "Invalid topology version [topVer=" + topVer + ", exchId=" + exchId + ']';
if (exchId.isLeft()) removeNode(exchId.nodeId());
// In case if node joins, get topology at the time of joining node.
ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);
assert oldest != null;
if (log.isDebugEnabled())
log.debug(
"Partition map beforeExchange [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
long updateSeq = this.updateSeq.incrementAndGet();
// If this is the oldest node.
if (oldest.id().equals(loc.id())
|| exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion())) {
if (node2part == null) {
node2part = new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq);
if (log.isDebugEnabled())
log.debug(
"Created brand new full topology map on oldest node [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
} else if (!node2part.valid()) {
node2part =
new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);
if (log.isDebugEnabled())
log.debug(
"Created new full topology map on oldest node [exchId="
+ exchId
+ ", fullMap="
+ node2part
+ ']');
} else if (!node2part.nodeId().equals(loc.id())) {
node2part =
new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);
if (log.isDebugEnabled())
log.debug(
"Copied old map into new map on oldest node (previous oldest node left) [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
}
}
if (cctx.rebalanceEnabled()) {
for (int p = 0; p < num; p++) {
// If this is the first node in grid.
boolean added = exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion());
if ((oldest.id().equals(loc.id())
&& oldest.id().equals(exchId.nodeId())
&& exchId.isJoined())
|| added) {
assert exchId.isJoined() || added;
try {
GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);
assert locPart != null;
boolean owned = locPart.own();
assert owned
: "Failed to own partition for oldest node [cacheName"
+ cctx.name()
+ ", part="
+ locPart
+ ']';
if (log.isDebugEnabled()) log.debug("Owned partition for oldest node: " + locPart);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition on oldest node (no need to create a partition "
+ "if it no longer belongs to local node: "
+ e.partition());
}
}
// If this is not the first node in grid.
else {
if (node2part != null && node2part.valid()) {
if (cctx.affinity().localNode(p, topVer)) {
try {
// This will make sure that all non-existing partitions
// will be created in MOVING state.
GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition (no need to create a partition if it "
+ "no longer belongs to local node: "
+ e.partition());
}
}
}
// If this node's map is empty, we pre-create local partitions,
// so local map will be sent correctly during exchange.
else if (cctx.affinity().localNode(p, topVer)) {
try {
localPartition(p, topVer, true, false);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition (no need to pre-create a partition if it "
+ "no longer belongs to local node: "
+ e.partition());
}
}
}
}
} else {
// If preloader is disabled, then we simply clear out
// the partitions this node is not responsible for.
for (int p = 0; p < num; p++) {
GridDhtLocalPartition locPart = localPartition(p, topVer, false, false);
boolean belongs = cctx.affinity().localNode(p, topVer);
if (locPart != null) {
if (!belongs) {
GridDhtPartitionState state = locPart.state();
if (state.active()) {
locPart.rent(false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
if (log.isDebugEnabled())
log.debug(
"Evicting partition with rebalancing disabled "
+ "(it does not belong to affinity): "
+ locPart);
}
}
} else if (belongs) {
try {
// Pre-create partitions.
localPartition(p, topVer, true, false);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition with disabled rebalancer (no need to "
+ "pre-create a partition if it no longer belongs to local node: "
+ e.partition());
}
}
}
}
if (node2part != null && node2part.valid()) checkEvictions(updateSeq);
consistencyCheck();
if (log.isDebugEnabled())
log.debug(
"Partition map after beforeExchange [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
} finally {
lock.writeLock().unlock();
}
// Wait for evictions.
waitForRent();
}
/**
* Updates value for single partition.
*
* @param p Partition.
* @param nodeId Node ID.
* @param state State.
* @param updateSeq Update sequence.
*/
@SuppressWarnings({"MismatchedQueryAndUpdateOfCollection"})
private void updateLocal(int p, UUID nodeId, GridDhtPartitionState state, long updateSeq) {
assert lock.isWriteLockedByCurrentThread();
assert nodeId.equals(cctx.nodeId());
// In case if node joins, get topology at the time of joining node.
ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);
assert oldest != null;
// If this node became the oldest node.
if (oldest.id().equals(cctx.nodeId())) {
long seq = node2part.updateSequence();
if (seq != updateSeq) {
if (seq > updateSeq) {
if (this.updateSeq.get() < seq) {
// Update global counter if necessary.
boolean b = this.updateSeq.compareAndSet(this.updateSeq.get(), seq + 1);
assert b
: "Invalid update sequence [updateSeq="
+ updateSeq
+ ", seq="
+ seq
+ ", curUpdateSeq="
+ this.updateSeq.get()
+ ", node2part="
+ node2part.toFullString()
+ ']';
updateSeq = seq + 1;
} else updateSeq = seq;
}
node2part.updateSequence(updateSeq);
}
}
GridDhtPartitionMap map = node2part.get(nodeId);
if (map == null)
node2part.put(
nodeId,
map =
new GridDhtPartitionMap(
nodeId,
updateSeq,
Collections.<Integer, GridDhtPartitionState>emptyMap(),
false));
map.updateSequence(updateSeq);
map.put(p, state);
Set<UUID> ids = part2node.get(p);
if (ids == null) part2node.put(p, ids = U.newHashSet(3));
ids.add(nodeId);
}
/**
* @param updateSeq Update sequence.
* @return Checks if any of the local partitions need to be evicted.
*/
private boolean checkEvictions(long updateSeq) {
assert lock.isWriteLockedByCurrentThread();
boolean changed = false;
UUID locId = cctx.nodeId();
for (GridDhtLocalPartition part : locParts.values()) {
GridDhtPartitionState state = part.state();
if (state.active()) {
int p = part.id();
List<ClusterNode> affNodes = cctx.affinity().nodes(p, topVer);
if (!affNodes.contains(cctx.localNode())) {
Collection<UUID> nodeIds = F.nodeIds(nodes(p, topVer, OWNING));
// If all affinity nodes are owners, then evict partition from local node.
if (nodeIds.containsAll(F.nodeIds(affNodes))) {
part.rent(false);
updateLocal(part.id(), locId, part.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug("Evicted local partition (all affinity nodes are owners): " + part);
} else {
int ownerCnt = nodeIds.size();
int affCnt = affNodes.size();
if (ownerCnt > affCnt) {
List<ClusterNode> sorted = new ArrayList<>(cctx.discovery().nodes(nodeIds));
// Sort by node orders in ascending order.
Collections.sort(sorted, CU.nodeComparator(true));
int diff = sorted.size() - affCnt;
for (int i = 0; i < diff; i++) {
ClusterNode n = sorted.get(i);
if (locId.equals(n.id())) {
part.rent(false);
updateLocal(part.id(), locId, part.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug(
"Evicted local partition (this node is oldest non-affinity node): " + part);
break;
}
}
}
}
}
}
}
return changed;
}