com.bigdata.resources.SplitIndexPartitionTask Maven / Gradle / Ivy
package com.bigdata.resources;
import java.util.Arrays;
import java.util.UUID;
import java.util.concurrent.atomic.AtomicLong;
import com.bigdata.btree.BTree;
import com.bigdata.btree.ILocalBTreeView;
import com.bigdata.btree.ISimpleSplitHandler;
import com.bigdata.btree.IndexMetadata;
import com.bigdata.btree.IndexSegment;
import com.bigdata.journal.AbstractTask;
import com.bigdata.journal.ITx;
import com.bigdata.journal.TimestampUtility;
import com.bigdata.mdi.IResourceMetadata;
import com.bigdata.mdi.IndexPartitionCause;
import com.bigdata.mdi.LocalPartitionMetadata;
import com.bigdata.mdi.MetadataIndex;
import com.bigdata.mdi.PartitionLocator;
import com.bigdata.service.DataService;
import com.bigdata.service.Event;
import com.bigdata.service.EventResource;
import com.bigdata.service.Split;
/**
* Task splits an index partition which is a compact view (no more than one
* journal and one index segment) and should be invoked when the size of the
* index segment on the disk exceeds the nominal size of an index partition. The
* index partition is the result of a compacting merge, which could have been
* created by {@link IncrementalBuildTask} or {@link CompactingMergeTask}. The
* index partition is passed into this task because it is not yet part of the
* view. Based on the nominal size of the index partition and the size of the
* segment, N=segSize/nominalSize splits will be generated, requiring N-1
* separator keys.
*
* The task uses the linear list API to identify N-1 separator key which would
* split the index segment and assumes that the data is evenly distributed
* across the keys within the index segment. The buffered writes are ignored
* when determining the separator keys (most data will be on the index segment
* if the journal extent roughly the same as the nominal index segment extent
* and multiple index partitions are registered on the journal). Application
* constraints on the choice of the separator keys will be honored and can
* result in fewer splits being generated.
*
* Once the N-1 split points have been selected, N index segments are built -
* one from each of the N key ranges which those N-1 split points define. Once
* the index segment for each split has been built, an
* {@link AtomicUpdateSplitIndexPartitionTask} will atomically re-define the
* source index partition as N new index partition and copy the buffered writes
* into the appropriate index partition. During the atomic update the original
* index partition becomes un-defined and new index partitions are defined in
* its place which span the same total key range and have the same data.
*
* @see AtomicUpdateSplitIndexPartitionTask, which MUST be invoked in order to
* update the index partition definitions on the live journal and the
* {@link MetadataIndex} as an atomic operation.
*
* @author Bryan Thompson
* @version $Id: SplitIndexPartitionTask.java 2265 2009-10-26 12:51:06Z
* thompsonbry $
*/
public class SplitIndexPartitionTask extends
AbstractPrepareTask {
protected final ViewMetadata vmd;
protected final UUID[] moveTargets;
/**
* The adjusted nominal bytes on disk for a full shard after a compacting
* merge. Note that this MAY have been overridden in order to promote a
* split so you MUST use this instance and NOT
* {@link OverflowManager#nominalShardSize}.
*/
private final long adjustedNominalShardSize;
/**
* @param vmd
* @param moveTarget
* When non-null the new right-sibling (the tail)
* will be moved to the specified data service after the split.
*/
protected SplitIndexPartitionTask(final ViewMetadata vmd,
final UUID moveTarget) {
this(vmd, (moveTarget == null ? null : new UUID[] { moveTarget }));
}
/**
*
* @param vmd
* @param moveTargets
* When non-null the index partitions generated
* by the split will be moved to the identified data services. If
* this data service is included in the array, then an index
* partition will be left on this data service. If the array
* contains a single element, then only the rightSibling of the
* split will be moved.
*/
protected SplitIndexPartitionTask(final ViewMetadata vmd,
final UUID[] moveTargets) {
super(vmd.resourceManager, TimestampUtility
.asHistoricalRead(vmd.commitTime), vmd.name);
this.vmd = vmd;
if (vmd.pmd == null) {
throw new IllegalStateException("Not an index partition.");
}
if (vmd.pmd.getSourcePartitionId() != -1) {
throw new IllegalStateException(
"Split not allowed during move: sourcePartitionId="
+ vmd.pmd.getSourcePartitionId());
}
if (moveTargets != null) {
if (moveTargets.length == 0)
throw new IllegalArgumentException();
if (moveTargets.length == 1
&& resourceManager.getDataServiceUUID().equals(moveTargets[0])) {
// can't specify this data service as the sole target for a move.
throw new IllegalArgumentException();
}
for(UUID t : moveTargets) {
if(t == null)
throw new IllegalArgumentException();
}
}
this.moveTargets = moveTargets;
this.adjustedNominalShardSize = vmd.getAdjustedNominalShardSize();
}
@Override
protected void clearRefs() {
vmd.clearRef();
}
/**
* Decides how many index partitions should be generated (N) and builds N
* {@link IndexSegment}s from the source index partition. If N will be ONE
* (1) if a detailed inspection of the source index partition reveals that
* it SHOULD NOT be split.
*
* @return A {@link SplitResult} if the index partition was split into 2 or
* more index partitions -or- a {@link BuildResult} iff the index
* partition was not split.
*/
@Override
protected AbstractResult doTask() throws Exception {
final Event e = new Event(resourceManager.getFederation(),
new EventResource(vmd.indexMetadata), OverflowActionEnum.Split,
vmd.getParams()).addDetail("summary", OverflowActionEnum.Split
+ (moveTargets != null ? "+" + OverflowActionEnum.Move : "")
+ "(" + vmd.name + ")");
if (moveTargets != null) {
e.addDetail("moveTargets", Arrays.toString(moveTargets));
}
e.start();
SplitResult splitResult = null;
try {
try {
if (resourceManager.isOverflowAllowed())
throw new IllegalStateException();
final String name = vmd.name;
// Note: fused view for the source index partition.
final ILocalBTreeView src = vmd.getView();
/*
* Get the split points for the index. Each split point
* describes a new index partition. Together the split points
* MUST exactly span the source index partitions key range.
* There MUST NOT be any overlap in the key ranges for the
* splits.
*/
Split[] splits = null;
final ISimpleSplitHandler splitHandler = vmd.indexMetadata
.getSplitHandler();
try {
// FIXME This operation should verify that the shard is compact as a precondition.
if (vmd.compactView) {
/*
* Choose splits using the linear-list API based on the
* index segment data only.
*/
splits = SplitUtility.getSplits(resourceManager,
vmd.pmd, (IndexSegment) src.getSources()[1],
adjustedNominalShardSize,
splitHandler);
}
} catch (Throwable t) {
if (AsynchronousOverflowTask.isNormalShutdown(
resourceManager, t)) {
/*
* This looks like an exception arising from the normal
* shutdown of the data service so we return
* immediately. As of this point we have not had any
* side effects on the data service.
*/
log.warn("Normal shutdown? : " + t);
return null;
}
/*
* Note: this makes the asynchronous overflow more robust to
* a failure in the split handler. However, if the split
* handler never succeeds then the index will never get
* split and it will eventually dominate the data service on
* which it resides.
*/
log.error(
"Split handler failure - will do build instead: name="
+ name + " : " + t, t);
splits = null;
}
if (splits == null) {
final double overextension = ((double) vmd.sumSegBytes)
/ resourceManager.nominalShardSize;
if (overextension > resourceManager.shardOverextensionLimit
&& !resourceManager.isDisabledWrites(vmd.name)) {
/*
* The shard is overextended (it is at least two times
* its nominal maximum size) and is refusing a split.
* Continuing to do incremental builds here will mask
* the problem and cause the cost of a merge on the
* shard to increase over time and will drag down
* performance for this DS. In order to prevent this we
* MUST disallow further writes on the shard. The shard
* can be re-enabled for writes by an administrative
* action once the problem has been fixed.
*
* Note: The default split behavior should always find a
* separator key to split the shard. The mostly likely
* cause for a problem is an application defined split
* handler. Rather than allowing a poorly written split
* handler to foul up the works, we disallow further
* writes onto this shard until the application has
* fixed their split handler.
*/
log.error("Shard will not split - writes are disabled"
+ ": name="
+ vmd.name
+ ", size="
+ vmd.sumSegBytes
+ ", overextended="
+ (int) overextension
+ "x"
+ ", splitHandler="
+ (splitHandler == null ? "N/A" : splitHandler
.getClass().getName()));
// Disable writes on the index partition.
resourceManager.disableWrites(vmd.name);
}
/*
* No splits were chosen so the index will not be split at
* this time.
*/
if (moveTargets != null && moveTargets.length >= 1) {
// There is a move target, so move the index partition.
log.warn("No splits identified: will move: " + vmd);
return concurrencyManager.submit(
new MoveTask(vmd, moveTargets[0])).get();
} else if (vmd.mandatoryMerge) {
// Mandatory compacting merge.
log.warn("No splits identified: will merge: " + vmd);
return concurrencyManager.submit(
new CompactingMergeTask(vmd)).get();
} else {
// Incremental build.
log.warn("No splits identified: will build: " + vmd);
return concurrencyManager.submit(
new IncrementalBuildTask(vmd)).get();
}
}
// The #of splits.
final int nsplits = splits.length;
if (INFO)
log.info("Will build index segments for " + nsplits
+ " splits for " + name + " : "
+ Arrays.toString(splits));
// validate the splits before processing them.
SplitUtility.validateSplits(src, splits);
splitResult = SplitUtility.buildSplits(vmd, splits, e);
} finally {
/*
* We are done building index segments from the source index
* partition view so we clear our references for that view.
*/
clearRefs();
}
/*
* Do the atomic update
*/
doSplitAtomicUpdate(
resourceManager,
vmd,
splitResult,
OverflowActionEnum.Split,
resourceManager.overflowCounters.indexPartitionSplitCounter,
e);
if (moveTargets != null) {
/*
* Note: Unlike a normal move where there are writes on the old
* journal, all the historical data for the each of the index
* partitions is in an index segment that we just built (new
* writes MAY be buffered on the live journal, so we still have
* to deal with that). Therefore we use a different entry point
* into the MOVE operation, one which does not copy over the
* data from the old journal but will still copy over any
* buffered writes.
*/
if (moveTargets.length == 1) {
/*
* This handles the case where only one move target was
* specified. In this case it is NOT permitted for the
* move target to be this data service (this condition
* is checked by the ctor).
*/
/*
* Find the split whose newly built index partition has the
* smallest size.
*/
final int bestMoveIndex;
{
int indexOfMinLength = -1;
long minLength = Long.MAX_VALUE;
for (int i = 0; i < splitResult.buildResults.length; i++) {
final BuildResult r = splitResult.buildResults[i];
// #of bytes in that index segment.
final long length = r.builder.getCheckpoint().length;
if (length < minLength) {
indexOfMinLength = i;
minLength = length;
}
}
assert indexOfMinLength != -1 : splitResult.toString();
bestMoveIndex = indexOfMinLength;
if (INFO)
log.info("Best split to move: "
+ splitResult.splits[bestMoveIndex]);
}
/*
* Obtain a new partition identifier for the partition that
* will be created when we move the index partition to the
* target data service.
*/
final int newPartitionId = resourceManager
.nextPartitionId(vmd.indexMetadata.getName());
/*
* The name of the post-split index partition that is the
* source for the move operation.
*/
final String nameOfPartitionToMove = DataService
.getIndexPartitionName(vmd.indexMetadata.getName(),
splitResult.splits[bestMoveIndex].pmd
.getPartitionId());
/*
* Move.
*
* Note: We do not explicitly delete the source index
* segment for the source index partition after the move. It
* will be required for historical views of the that index
* partition in case any client gained access to the index
* partition after the split and before the move. It will
* eventually be released once the view of the source index
* partition becomes sufficiently aged that it falls off the
* head of the database history.
*/
MoveTask.doAtomicUpdate(resourceManager, nameOfPartitionToMove,
splitResult.buildResults[bestMoveIndex], moveTargets[0],
newPartitionId, e);
} else {
/*
* This handles the case where multiple move targets were
* specified. For this case, it is allowable for one of the
* move targets to be this data service, in which case we
* simply leaf the corresponding index partition in place.
*/
final int nsplits = splitResult.buildResults.length;
for (int i = 0; i < nsplits; i++) {
final UUID moveTarget = moveTargets[i
% moveTargets.length];
if (resourceManager.getDataServiceUUID().equals(moveTarget)) {
// ignore move to self.
if(INFO)
log.info("Ignoring move to self.");
continue;
}
/*
* Obtain a new partition identifier for the partition
* that will be created when we move the index partition
* to the target data service.
*/
final int newPartitionId = resourceManager
.nextPartitionId(vmd.indexMetadata.getName());
/*
* The name of the post-split index partition that is
* the source for the move operation.
*/
final String nameOfPartitionToMove = DataService
.getIndexPartitionName(vmd.indexMetadata
.getName(), splitResult.splits[i].pmd
.getPartitionId());
/*
* Move.
*
* Note: We do not explicitly delete the source index
* segment for the source index partition after the
* move. It will be required for historical views of the
* that index partition in case any client gained access
* to the index partition after the split and before the
* move. It will eventually be released once the view of
* the source index partition becomes sufficiently aged
* that it falls off the head of the database history.
*/
MoveTask.doAtomicUpdate(resourceManager,
nameOfPartitionToMove, splitResult.buildResults[i],
moveTarget, newPartitionId, e);
}
}
}
// Done.
return splitResult;
} finally {
if (splitResult != null) {
for (BuildResult buildResult : splitResult.buildResults) {
if (buildResult != null) {
/*
* At this point the index segment was either incorporated into
* the new view in a restart safe manner or there was an error.
* Either way, we now remove the index segment store's UUID from
* the retentionSet so it will be subject to the release policy
* of the StoreManager.
*/
resourceManager
.retentionSetRemove(buildResult.segmentMetadata
.getUUID());
}
}
}
e.end();
}
}
/**
*
* @param resourceManager
* @param vmd
* @param splits
* @param result
* @param action
* @param counter
* @param parentEvent
*/
static protected void doSplitAtomicUpdate(
final ResourceManager resourceManager, final ViewMetadata vmd,
final SplitResult result,
final OverflowActionEnum action,
final AtomicLong counter,
final Event parentEvent) {
try {
/*
* Form up the set of resources on which the atomic update task
* must have an exclusive lock before it can run. This includes
* both the source index partition and the name of each new
* index partition which will be generated by this split.
*
* Note: We MUST declare the resource locks for the indices that
* we are going to create in order to prevent tasks from
* accessing those indices until the atomic update task has
* committed. Note that the metadata index will be updated
* before the atomic update task commits, so it is possible (and
* does in fact happen) for clients to submit tasks that wind up
* directed to one of the new index partitions before the atomic
* update task commits.
*/
final Split[] splits = result.splits;
final String[] resources = new String[splits.length + 1];
{
resources[0] = result.name;
int i = 0;
for (final Split split : splits) {
final int partitionId = split.pmd.getPartitionId();
resources[i + 1] = DataService.getIndexPartitionName(
vmd.indexMetadata.getName(), partitionId);
i++;
}
}
/*
* Create task that will perform atomic update, converting the
* source index partition into N new index partitions.
*/
final AbstractTask task = new AtomicUpdateSplitIndexPartitionTask(
resourceManager, resources, action, vmd.indexMetadata
.getIndexUUID(), result, parentEvent.newSubEvent(
OverflowSubtaskEnum.AtomicUpdate).addDetail(
"summary",
action + "(" + vmd.name + "->"
+ Arrays.toString(resources)));
// submit atomic update task and wait for it to complete
resourceManager.getConcurrencyManager().submit(task).get();
// update the counter.
counter.incrementAndGet();
return;
} catch (Throwable t) {
/*
* Error handling - remove all generated files.
*
* @todo error handling should be in the atomic update task since it
* has greater visibility into when the resources are incorporated
* into a view and hence accessible to concurrent processes.
*/
for (BuildResult r : result.buildResults) {
if (r == null)
continue;
// make it releasable.
resourceManager.retentionSetRemove(r.segmentMetadata.getUUID());
// delete it.
resourceManager.deleteResource(r.segmentMetadata.getUUID(),
false/* isJournal */);
}
throw new RuntimeException(t);
}
}
/**
* An {@link ITx#UNISOLATED} operation that splits the live index using the
* same {@link Split} points, generating new index partitions with new
* partition identifiers. The old index partition is deleted as a
* post-condition. The new index partitions are registered as a
* post-condition. Any data that was accumulated in the live index on the
* live journal is copied into the appropriate new {@link BTree} for the new
* index partition on the live journal.
*
* @author Bryan Thompson
* @version $Id$
*/
static protected class AtomicUpdateSplitIndexPartitionTask extends
AbstractAtomicUpdateTask {
/**
* The expected UUID of the scale-out index.
*/
final protected UUID indexUUID;
/**
* Either a normal split or a tail split.
*/
protected final OverflowActionEnum action;
protected final SplitResult splitResult;
private final Event updateEvent;
/**
*
* @param resourceManager
* @param resource
* @param action
* @param indexUUID The UUID of the scale-out index.
* @param splitResult
*/
public AtomicUpdateSplitIndexPartitionTask(
final ResourceManager resourceManager, final String[] resource,
final OverflowActionEnum action, final UUID indexUUID,
final SplitResult splitResult, final Event updateEvent) {
super(resourceManager, ITx.UNISOLATED, resource);
if (action == null)
throw new IllegalArgumentException();
if (indexUUID == null)
throw new IllegalArgumentException();
if (splitResult == null)
throw new IllegalArgumentException();
if (updateEvent == null)
throw new IllegalArgumentException();
this.action = action;
this.indexUUID = indexUUID;
this.splitResult = splitResult;
this.updateEvent = updateEvent;
}
/**
* Atomic update.
*
* @return null.
*/
@Override
protected Void doTask() throws Exception {
updateEvent.start();
try {
if (resourceManager.isOverflowAllowed())
throw new IllegalStateException();
// The name of the scale-out index.
final String scaleOutIndexName = splitResult.indexMetadata
.getName();
// the name of the source index.
final String name = splitResult.name;
/*
* Note: the source index is the BTree on the live journal that
* has been absorbing writes since the last overflow (while the
* split was running asynchronously).
*
* This is NOT a fused view. All we are doing is re-distributing
* the buffered writes onto the B+Trees buffering writes for the
* new index partitions created by the split.
*/
final BTree src = ((ILocalBTreeView) getIndex(name))
.getMutableBTree();
assertSameIndex(indexUUID, src);
if (INFO) {
log.info("src=" + name + ", counter="
+ src.getCounter().get() + ", checkpoint="
+ src.getCheckpoint());
log.info("src=" + name + ", splitResult=" + splitResult);
}
// the value of the counter on the source BTree.
final long oldCounter = src.getCounter().get();
/*
* Locators for the new index partitions.
*/
final LocalPartitionMetadata oldpmd = (LocalPartitionMetadata) src
.getIndexMetadata().getPartitionMetadata();
if (oldpmd.getSourcePartitionId() != -1) {
throw new IllegalStateException(
"Split not allowed during move: sourcePartitionId="
+ oldpmd.getSourcePartitionId());
}
final Split[] splits = splitResult.splits;
final PartitionLocator[] locators = new PartitionLocator[splits.length];
for (int i = 0; i < splits.length; i++) {
// new metadata record (cloned).
final IndexMetadata md = src.getIndexMetadata().clone();
final LocalPartitionMetadata pmd = (LocalPartitionMetadata) splits[i].pmd;
assert pmd.getResources() == null : "Not expecting resources for index segment: "
+ pmd;
// the new partition identifier.
final int partitionId = pmd.getPartitionId();
// name of the new index partition.
final String name2 = DataService.getIndexPartitionName(
scaleOutIndexName, partitionId);
/*
* form locator for the new index partition for this split..
*/
final PartitionLocator locator = new PartitionLocator(pmd
.getPartitionId(),//
/*
* The (logical) data service.
*
* @todo The index partition data will be replicated
* at the byte image level for the live journal.
*
* @todo New index segment resources must be
* replicated as well.
*
* @todo Once the index partition data is fully
* replicated we update the metadata index.
*/
resourceManager.getDataServiceUUID(),//
pmd.getLeftSeparatorKey(),//
pmd.getRightSeparatorKey()//
);
locators[i] = locator;
final String summary = action + "(" + name + "->" + name2
+ ")";
/*
* Update the view definition.
*/
md
.setPartitionMetadata(new LocalPartitionMetadata(
pmd.getPartitionId(),//
-1, // Note: Split not allowed during move.
pmd.getLeftSeparatorKey(),//
pmd.getRightSeparatorKey(),//
new IResourceMetadata[] {//
/*
* Resources are (a) the new btree;
* and (b) the new index segment.
*/
resourceManager.getLiveJournal()
.getResourceMetadata(),
splitResult.buildResults[i].segmentMetadata },
IndexPartitionCause.split(resourceManager)
// /*
// * Note: history is record of the split.
// */
// , pmd.getHistory() + summary + " "//
));
/*
* create new btree.
*
* Note: the lower 32-bits of the counter will be zero. The
* high 32-bits will be the partition identifier assigned to
* the new index partition.
*/
final BTree btree = BTree.create(resourceManager
.getLiveJournal(), md);
// make sure the partition identifier was asserted.
assert partitionId == btree.getIndexMetadata()
.getPartitionMetadata().getPartitionId();
final long newCounter = btree.getCounter().get();
/*
* Note: this is true because partition identifiers always
* increase and the partition identifier is placed into the
* high word of the counter value for an index partition.
*/
assert newCounter > oldCounter : "newCounter=" + newCounter
+ " not GT oldCounter=" + oldCounter;
// lower bound (inclusive) for copy.
final byte[] fromKey = pmd.getLeftSeparatorKey();
// upper bound (exclusive) for copy.
final byte[] toKey = pmd.getRightSeparatorKey();
if (INFO)
log.info("Copying data to new btree: index="
+ scaleOutIndexName + ", pmd=" + pmd);
/*
* Copy all data in this split from the source index.
*
* Note: [overflow := false] since the btrees are on the
* same backing store.
*/
final long ncopied = btree.rangeCopy(src, fromKey, toKey,
false/* overflow */);
if (INFO)
log.info("Copied " + ncopied
+ " index entries from the live index " + name
+ " onto " + name2);
// register it on the live journal
if (INFO)
log.info("Registering index: " + name2);
getJournal().registerIndex(name2, btree);
}
// drop the source index (the old index partition)
if (INFO)
log.info("Dropping source index: " + name);
getJournal().dropIndex(name);
/*
* Notify the metadata service that the index partition has been
* split.
*/
resourceManager.getFederation().getMetadataService()
.splitIndexPartition(src.getIndexMetadata().getName(),//
new PartitionLocator(//
oldpmd.getPartitionId(), //
resourceManager.getDataServiceUUID(), //
oldpmd.getLeftSeparatorKey(),//
oldpmd.getRightSeparatorKey()//
), locators);
if (INFO)
log.info("Notified metadata service: name=" + name
+ " was split into " + Arrays.toString(locators));
// will notify tasks that index partition was split.
resourceManager.setIndexPartitionGone(name,
StaleLocatorReason.Split);
return null;
} finally {
updateEvent.end();
}
} // doTask()
} // class AtomicUpdate
}