@Override
public void close() {
try {
if (batchGroups != null) {
closeBatchGroups(batchGroups);
batchGroups = null;
}
if (spilledBatchGroups != null) {
closeBatchGroups(spilledBatchGroups);
spilledBatchGroups = null;
}
} finally {
if (builder != null) {
builder.clear();
builder.close();
}
if (sv4 != null) {
sv4.clear();
}
try {
if (copier != null) {
copier.close();
}
} catch (IOException e) {
throw new RuntimeException(e);
} finally {
copierAllocator.close();
super.close();
if (mSorter != null) {
mSorter.clear();
}
for (Iterator iter = this.currSpillDirs.iterator(); iter.hasNext(); iter.remove()) {
Path path = (Path) iter.next();
try {
if (fs != null && path != null && fs.exists(path)) {
if (fs.delete(path, true)) {
fs.cancelDeleteOnExit(path);
}
}
} catch (IOException e) {
// since this is meant to be used in a batches's cleanup, we don't propagate the
// exception
logger.warn("Unable to delete spill directory " + path, e);
}
}
}
}
}
private boolean hasMemoryForInMemorySort(int currentRecordCount) {
long currentlyAvailable = popConfig.getMaxAllocation() - oAllocator.getAllocatedMemory();
long neededForInMemorySort =
SortRecordBatchBuilder.memoryNeeded(currentRecordCount)
+ MSortTemplate.memoryNeeded(currentRecordCount);
return currentlyAvailable > neededForInMemorySort;
}
@Override
public IterOutcome innerNext() {
if (schema != null) {
if (spillCount == 0) {
return (getSelectionVector4().next()) ? IterOutcome.OK : IterOutcome.NONE;
} else {
Stopwatch w = Stopwatch.createStarted();
int count = copier.next(targetRecordCount);
if (count > 0) {
long t = w.elapsed(TimeUnit.MICROSECONDS);
logger.debug("Took {} us to merge {} records", t, count);
container.setRecordCount(count);
return IterOutcome.OK;
} else {
logger.debug("copier returned 0 records");
return IterOutcome.NONE;
}
}
}
int totalCount = 0;
int totalBatches = 0; // total number of batches received so far
try {
container.clear();
outer:
while (true) {
IterOutcome upstream;
if (first) {
upstream = IterOutcome.OK_NEW_SCHEMA;
} else {
upstream = next(incoming);
}
if (upstream == IterOutcome.OK && sorter == null) {
upstream = IterOutcome.OK_NEW_SCHEMA;
}
switch (upstream) {
case NONE:
if (first) {
return upstream;
}
break outer;
case NOT_YET:
throw new UnsupportedOperationException();
case STOP:
return upstream;
case OK_NEW_SCHEMA:
case OK:
VectorContainer convertedBatch;
// only change in the case that the schema truly changes. Artificial schema changes are
// ignored.
if (upstream == IterOutcome.OK_NEW_SCHEMA && !incoming.getSchema().equals(schema)) {
if (schema != null) {
if (unionTypeEnabled) {
this.schema = SchemaUtil.mergeSchemas(schema, incoming.getSchema());
} else {
throw new SchemaChangeException(
"Schema changes not supported in External Sort. Please enable Union type");
}
} else {
schema = incoming.getSchema();
}
convertedBatch = SchemaUtil.coerceContainer(incoming, schema, oContext);
for (BatchGroup b : batchGroups) {
b.setSchema(schema);
}
for (BatchGroup b : spilledBatchGroups) {
b.setSchema(schema);
}
this.sorter = createNewSorter(context, convertedBatch);
} else {
convertedBatch = SchemaUtil.coerceContainer(incoming, schema, oContext);
}
if (first) {
first = false;
}
if (convertedBatch.getRecordCount() == 0) {
for (VectorWrapper<?> w : convertedBatch) {
w.clear();
}
break;
}
SelectionVector2 sv2;
if (incoming.getSchema().getSelectionVectorMode()
== BatchSchema.SelectionVectorMode.TWO_BYTE) {
sv2 = incoming.getSelectionVector2().clone();
} else {
try {
sv2 = newSV2();
} catch (InterruptedException e) {
return IterOutcome.STOP;
} catch (OutOfMemoryException e) {
throw new OutOfMemoryException(e);
}
}
int count = sv2.getCount();
totalCount += count;
totalBatches++;
sorter.setup(context, sv2, convertedBatch);
sorter.sort(sv2);
RecordBatchData rbd = new RecordBatchData(convertedBatch, oAllocator);
boolean success = false;
try {
rbd.setSv2(sv2);
batchGroups.add(new BatchGroup(rbd.getContainer(), rbd.getSv2(), oContext));
if (peakNumBatches < batchGroups.size()) {
peakNumBatches = batchGroups.size();
stats.setLongStat(Metric.PEAK_BATCHES_IN_MEMORY, peakNumBatches);
}
batchesSinceLastSpill++;
if ( // If we haven't spilled so far, do we have enough memory for MSorter if this
// turns out to be the last incoming batch?
(spillCount == 0 && !hasMemoryForInMemorySort(totalCount))
||
// If we haven't spilled so far, make sure we don't exceed the maximum number of
// batches SV4 can address
(spillCount == 0 && totalBatches > Character.MAX_VALUE)
||
// TODO(DRILL-4438) - consider setting this threshold more intelligently,
// lowering caused a failing low memory condition (test in
// BasicPhysicalOpUnitTest)
// to complete successfully (although it caused perf decrease as there was more
// spilling)
// current memory used is more than 95% of memory usage limit of this operator
(oAllocator.getAllocatedMemory() > .95 * oAllocator.getLimit())
||
// Number of incoming batches (BatchGroups) exceed the limit and number of
// incoming batches accumulated
// since the last spill exceed the defined limit
(batchGroups.size() > SPILL_THRESHOLD
&& batchesSinceLastSpill >= SPILL_BATCH_GROUP_SIZE)) {
if (firstSpillBatchCount == 0) {
firstSpillBatchCount = batchGroups.size();
}
if (spilledBatchGroups.size() > firstSpillBatchCount / 2) {
logger.info("Merging spills");
final BatchGroup merged = mergeAndSpill(spilledBatchGroups);
if (merged != null) {
spilledBatchGroups.addFirst(merged);
}
}
final BatchGroup merged = mergeAndSpill(batchGroups);
if (merged != null) { // make sure we don't add null to spilledBatchGroups
spilledBatchGroups.add(merged);
batchesSinceLastSpill = 0;
}
}
success = true;
} finally {
if (!success) {
rbd.clear();
}
}
break;
case OUT_OF_MEMORY:
logger.debug("received OUT_OF_MEMORY, trying to spill");
if (batchesSinceLastSpill > 2) {
final BatchGroup merged = mergeAndSpill(batchGroups);
if (merged != null) {
spilledBatchGroups.add(merged);
batchesSinceLastSpill = 0;
}
} else {
logger.debug("not enough batches to spill, sending OUT_OF_MEMORY downstream");
return IterOutcome.OUT_OF_MEMORY;
}
break;
default:
throw new UnsupportedOperationException();
}
}
if (totalCount == 0) {
return IterOutcome.NONE;
}
if (spillCount == 0) {
if (builder != null) {
builder.clear();
builder.close();
}
builder = new SortRecordBatchBuilder(oAllocator);
for (BatchGroup group : batchGroups) {
RecordBatchData rbd = new RecordBatchData(group.getContainer(), oAllocator);
rbd.setSv2(group.getSv2());
builder.add(rbd);
}
builder.build(context, container);
sv4 = builder.getSv4();
mSorter = createNewMSorter();
mSorter.setup(context, oAllocator, getSelectionVector4(), this.container);
// For testing memory-leak purpose, inject exception after mSorter finishes setup
injector.injectUnchecked(context.getExecutionControls(), INTERRUPTION_AFTER_SETUP);
mSorter.sort(this.container);
// sort may have prematurely exited due to should continue returning false.
if (!context.shouldContinue()) {
return IterOutcome.STOP;
}
// For testing memory-leak purpose, inject exception after mSorter finishes sorting
injector.injectUnchecked(context.getExecutionControls(), INTERRUPTION_AFTER_SORT);
sv4 = mSorter.getSV4();
container.buildSchema(SelectionVectorMode.FOUR_BYTE);
} else { // some batches were spilled
final BatchGroup merged = mergeAndSpill(batchGroups);
if (merged != null) {
spilledBatchGroups.add(merged);
}
batchGroups.addAll(spilledBatchGroups);
spilledBatchGroups =
null; // no need to cleanup spilledBatchGroups, all it's batches are in batchGroups now
logger.warn(
"Starting to merge. {} batch groups. Current allocated memory: {}",
batchGroups.size(),
oAllocator.getAllocatedMemory());
VectorContainer hyperBatch = constructHyperBatch(batchGroups);
createCopier(hyperBatch, batchGroups, container, false);
int estimatedRecordSize = 0;
for (VectorWrapper<?> w : batchGroups.get(0)) {
try {
estimatedRecordSize += TypeHelper.getSize(w.getField().getType());
} catch (UnsupportedOperationException e) {
estimatedRecordSize += 50;
}
}
targetRecordCount =
Math.min(MAX_BATCH_SIZE, Math.max(1, COPIER_BATCH_MEM_LIMIT / estimatedRecordSize));
int count = copier.next(targetRecordCount);
container.buildSchema(SelectionVectorMode.NONE);
container.setRecordCount(count);
}
return IterOutcome.OK_NEW_SCHEMA;
} catch (SchemaChangeException ex) {
kill(false);
context.fail(
UserException.unsupportedError(ex)
.message("Sort doesn't currently support sorts with changing schemas")
.build(logger));
return IterOutcome.STOP;
} catch (ClassTransformationException | IOException ex) {
kill(false);
context.fail(ex);
return IterOutcome.STOP;
} catch (UnsupportedOperationException e) {
throw new RuntimeException(e);
}
}
private void buildTable()
throws SchemaChangeException, ClassTransformationException, IOException {
// Get all samples from distributed map
SortRecordBatchBuilder containerBuilder =
new SortRecordBatchBuilder(context.getAllocator(), MAX_SORT_BYTES);
for (CachedVectorContainer w : mmap.get(mapKey)) {
containerBuilder.add(w.get());
}
VectorContainer allSamplesContainer = new VectorContainer();
containerBuilder.build(context, allSamplesContainer);
List<Ordering> orderDefs = Lists.newArrayList();
int i = 0;
for (Ordering od : popConfig.getOrderings()) {
SchemaPath sp = SchemaPath.getSimplePath("f" + i++);
orderDefs.add(new Ordering(od.getDirection(), new FieldReference(sp)));
}
// sort the data incoming samples.
SelectionVector4 newSv4 = containerBuilder.getSv4();
Sorter sorter = SortBatch.createNewSorter(context, orderDefs, allSamplesContainer);
sorter.setup(context, newSv4, allSamplesContainer);
sorter.sort(newSv4, allSamplesContainer);
// Copy every Nth record from the samples into a candidate partition table, where N =
// totalSampledRecords/partitions
// Attempt to push this to the distributed map. Only the first candidate to get pushed will be
// used.
VectorContainer candidatePartitionTable = new VectorContainer();
SampleCopier copier = null;
List<ValueVector> localAllocationVectors = Lists.newArrayList();
copier =
getCopier(
newSv4,
allSamplesContainer,
candidatePartitionTable,
orderDefs,
localAllocationVectors);
int allocationSize = 50;
while (true) {
for (ValueVector vv : localAllocationVectors) {
AllocationHelper.allocate(vv, samplingFactor * partitions, allocationSize);
}
int skipRecords = containerBuilder.getSv4().getTotalCount() / partitions;
if (copier.copyRecords(skipRecords, skipRecords, partitions - 1)) {
assert copier.getOutputRecords() == partitions - 1
: String.format(
"output records: %d partitions: %d", copier.getOutputRecords(), partitions);
for (VectorWrapper<?> vw : candidatePartitionTable) {
vw.getValueVector().getMutator().setValueCount(copier.getOutputRecords());
}
break;
} else {
candidatePartitionTable.zeroVectors();
allocationSize *= 2;
}
}
candidatePartitionTable.setRecordCount(copier.getOutputRecords());
WritableBatch batch =
WritableBatch.getBatchNoHVWrap(
candidatePartitionTable.getRecordCount(), candidatePartitionTable, false);
CachedVectorContainer wrap =
new CachedVectorContainer(batch, context.getDrillbitContext().getAllocator());
tableMap.putIfAbsent(mapKey + "final", wrap, 1, TimeUnit.MINUTES);
candidatePartitionTable.clear();
allSamplesContainer.clear();
containerBuilder.clear();
wrap.clear();
}
private boolean saveSamples()
throws SchemaChangeException, ClassTransformationException, IOException {
recordsSampled = 0;
IterOutcome upstream;
// Start collecting batches until recordsToSample records have been collected
SortRecordBatchBuilder builder =
new SortRecordBatchBuilder(oContext.getAllocator(), MAX_SORT_BYTES);
builder.add(incoming);
recordsSampled += incoming.getRecordCount();
outer:
while (recordsSampled < recordsToSample) {
upstream = next(incoming);
switch (upstream) {
case NONE:
case NOT_YET:
case STOP:
upstreamNone = true;
break outer;
default:
// fall through
}
builder.add(incoming);
recordsSampled += incoming.getRecordCount();
if (upstream == IterOutcome.NONE) break;
}
VectorContainer sortedSamples = new VectorContainer();
builder.build(context, sortedSamples);
// Sort the records according the orderings given in the configuration
Sorter sorter = SortBatch.createNewSorter(context, popConfig.getOrderings(), sortedSamples);
SelectionVector4 sv4 = builder.getSv4();
sorter.setup(context, sv4, sortedSamples);
sorter.sort(sv4, sortedSamples);
// Project every Nth record to a new vector container, where N = recordsSampled/(samplingFactor
// * partitions).
// Uses the
// the expressions from the Orderings to populate each column. There is one column for each
// Ordering in
// popConfig.orderings.
VectorContainer containerToCache = new VectorContainer();
List<ValueVector> localAllocationVectors = Lists.newArrayList();
SampleCopier copier =
getCopier(
sv4, sortedSamples, containerToCache, popConfig.getOrderings(), localAllocationVectors);
int allocationSize = 50;
while (true) {
for (ValueVector vv : localAllocationVectors) {
AllocationHelper.allocate(vv, samplingFactor * partitions, allocationSize);
}
if (copier.copyRecords(
recordsSampled / (samplingFactor * partitions), 0, samplingFactor * partitions)) {
break;
} else {
containerToCache.zeroVectors();
allocationSize *= 2;
}
}
for (VectorWrapper<?> vw : containerToCache) {
vw.getValueVector().getMutator().setValueCount(copier.getOutputRecords());
}
containerToCache.setRecordCount(copier.getOutputRecords());
// Get a distributed multimap handle from the distributed cache, and put the vectors from the
// new vector container
// into a serializable wrapper object, and then add to distributed map
WritableBatch batch =
WritableBatch.getBatchNoHVWrap(containerToCache.getRecordCount(), containerToCache, false);
CachedVectorContainer sampleToSave = new CachedVectorContainer(batch, context.getAllocator());
mmap.put(mapKey, sampleToSave);
this.sampledIncomingBatches = builder.getHeldRecordBatches();
builder.clear();
batch.clear();
containerToCache.clear();
sampleToSave.clear();
return true;
}
