public int getNextIndex() {
int val;
if (pointer == getRecordCount()) {
if (spilledBatches == 0) {
return -1;
}
try {
currentContainer.zeroVectors();
getBatch();
} catch (IOException e) {
throw new RuntimeException(e);
}
pointer = 1;
return 0;
}
if (sv2 == null) {
val = pointer;
pointer++;
assert val < currentContainer.getRecordCount();
} else {
val = pointer;
pointer++;
assert val < currentContainer.getRecordCount();
val = sv2.getIndex(val);
}
return val;
}
/**
* Creates a copier that does a project for every Nth record from a VectorContainer incoming into
* VectorContainer outgoing. Each Ordering in orderings generates a column, and evaluation of the
* expression associated with each Ordering determines the value of each column. These records
* will later be sorted based on the values in each column, in the same order as the orderings.
*
* @param sv4
* @param incoming
* @param outgoing
* @param orderings
* @return
* @throws SchemaChangeException
*/
private SampleCopier getCopier(
SelectionVector4 sv4,
VectorContainer incoming,
VectorContainer outgoing,
List<Ordering> orderings,
List<ValueVector> localAllocationVectors)
throws SchemaChangeException {
final ErrorCollector collector = new ErrorCollectorImpl();
final ClassGenerator<SampleCopier> cg =
CodeGenerator.getRoot(SampleCopier.TEMPLATE_DEFINITION, context.getFunctionRegistry());
int i = 0;
for (Ordering od : orderings) {
final LogicalExpression expr =
ExpressionTreeMaterializer.materialize(
od.getExpr(), incoming, collector, context.getFunctionRegistry());
SchemaPath schemaPath = SchemaPath.getSimplePath("f" + i++);
TypeProtos.MajorType.Builder builder =
TypeProtos.MajorType.newBuilder()
.mergeFrom(expr.getMajorType())
.clearMode()
.setMode(TypeProtos.DataMode.REQUIRED);
TypeProtos.MajorType newType = builder.build();
MaterializedField outputField = MaterializedField.create(schemaPath, newType);
if (collector.hasErrors()) {
throw new SchemaChangeException(
String.format(
"Failure while trying to materialize incoming schema. Errors:\n %s.",
collector.toErrorString()));
}
ValueVector vector = TypeHelper.getNewVector(outputField, oContext.getAllocator());
localAllocationVectors.add(vector);
TypedFieldId fid = outgoing.add(vector);
ValueVectorWriteExpression write = new ValueVectorWriteExpression(fid, expr, true);
HoldingContainer hc = cg.addExpr(write);
cg.getEvalBlock()._if(hc.getValue().eq(JExpr.lit(0)))._then()._return(JExpr.FALSE);
}
cg.rotateBlock();
cg.getEvalBlock()._return(JExpr.TRUE);
outgoing.buildSchema(BatchSchema.SelectionVectorMode.NONE);
try {
SampleCopier sampleCopier = context.getImplementationClass(cg);
sampleCopier.setupCopier(context, sv4, incoming, outgoing);
return sampleCopier;
} catch (ClassTransformationException | IOException e) {
throw new SchemaChangeException(e);
}
}
public void addBatch(VectorContainer newContainer) throws IOException {
assert fs != null;
assert path != null;
if (outputStream == null) {
outputStream = fs.create(path);
}
int recordCount = newContainer.getRecordCount();
WritableBatch batch = WritableBatch.getBatchNoHVWrap(recordCount, newContainer, false);
VectorAccessibleSerializable outputBatch = new VectorAccessibleSerializable(batch, allocator);
Stopwatch watch = Stopwatch.createStarted();
outputBatch.writeToStream(outputStream);
newContainer.zeroVectors();
logger.debug(
"Took {} us to spill {} records", watch.elapsed(TimeUnit.MICROSECONDS), recordCount);
spilledBatches++;
}
private void createCopier(
VectorAccessible batch,
List<BatchGroup> batchGroupList,
VectorContainer outputContainer,
boolean spilling)
throws SchemaChangeException {
try {
if (copier == null) {
CodeGenerator<PriorityQueueCopier> cg =
CodeGenerator.get(
PriorityQueueCopier.TEMPLATE_DEFINITION,
context.getFunctionRegistry(),
context.getOptions());
ClassGenerator<PriorityQueueCopier> g = cg.getRoot();
generateComparisons(g, batch);
g.setMappingSet(COPIER_MAPPING_SET);
CopyUtil.generateCopies(g, batch, true);
g.setMappingSet(MAIN_MAPPING);
copier = context.getImplementationClass(cg);
} else {
copier.close();
}
BufferAllocator allocator = spilling ? copierAllocator : oAllocator;
for (VectorWrapper<?> i : batch) {
ValueVector v = TypeHelper.getNewVector(i.getField(), allocator);
outputContainer.add(v);
}
copier.setup(context, allocator, batch, batchGroupList, outputContainer);
} catch (ClassTransformationException | IOException e) {
throw new RuntimeException(e);
}
}
public void removeField(MaterializedField field) throws SchemaChangeException {
ValueVector vector = fieldVectorMap.remove(field);
if (vector == null)
throw new SchemaChangeException("Failure attempting to remove an unknown field.");
container.remove(vector);
vector.close();
}
@Override
public void removeAllFields() {
for (VectorWrapper<?> vw : container) {
vw.clear();
}
container.clear();
fieldVectorMap.clear();
}
@Override
public int getRecordCount() {
if (sv2 != null) {
return sv2.getCount();
} else {
return currentContainer.getRecordCount();
}
}
/**
* This method is called when the first batch comes in. Incoming batches are collected until a
* threshold is met. At that point, the records in the batches are sorted and sampled, and the
* sampled records are stored in the distributed cache. Once a sufficient fraction of the
* fragments have shared their samples, each fragment grabs all the samples, sorts all the
* records, builds a partition table, and attempts to push the partition table to the distributed
* cache. Whichever table gets pushed first becomes the table used by all fragments for
* partitioning.
*
* @return True is successful. False if failed.
*/
private boolean getPartitionVectors() {
try {
if (!saveSamples()) {
return false;
}
CachedVectorContainer finalTable = null;
long val = minorFragmentSampleCount.incrementAndGet();
logger.debug("Incremented mfsc, got {}", val);
final long fragmentsBeforeProceed =
(long) Math.ceil(sendingMajorFragmentWidth * completionFactor);
final String finalTableKey = mapKey + "final";
if (val == fragmentsBeforeProceed) { // we crossed the barrier, build table and get data.
buildTable();
finalTable = tableMap.get(finalTableKey);
} else {
// Wait until sufficient number of fragments have submitted samples, or proceed after xx ms
// passed
// TODO: this should be polling.
if (val < fragmentsBeforeProceed) Thread.sleep(10);
for (int i = 0; i < 100 && finalTable == null; i++) {
finalTable = tableMap.get(finalTableKey);
if (finalTable != null) {
break;
}
Thread.sleep(10);
}
if (finalTable == null) {
buildTable();
}
finalTable = tableMap.get(finalTableKey);
}
Preconditions.checkState(finalTable != null);
// Extract vectors from the wrapper, and add to partition vectors. These vectors will be used
// for partitioning in
// the rest of this operator
for (VectorWrapper<?> w : finalTable.get()) {
partitionVectors.add(w.getValueVector());
}
} catch (ClassTransformationException
| IOException
| SchemaChangeException
| InterruptedException ex) {
kill(false);
logger.error("Failure while building final partition table.", ex);
context.fail(ex);
return false;
}
return true;
}
private VectorContainer constructHyperBatch(List<BatchGroup> batchGroupList) {
VectorContainer cont = new VectorContainer();
for (MaterializedField field : schema) {
ValueVector[] vectors = new ValueVector[batchGroupList.size()];
int i = 0;
for (BatchGroup group : batchGroupList) {
vectors[i++] =
group
.getValueAccessorById(
field.getValueClass(),
group.getValueVectorId(SchemaPath.getSimplePath(field.getPath())).getFieldIds())
.getValueVector();
}
cont.add(vectors);
}
cont.buildSchema(BatchSchema.SelectionVectorMode.FOUR_BYTE);
return cont;
}
private VectorContainer getBatch() throws IOException {
assert fs != null;
assert path != null;
if (inputStream == null) {
inputStream = fs.open(path);
}
VectorAccessibleSerializable vas = new VectorAccessibleSerializable(allocator);
Stopwatch watch = Stopwatch.createStarted();
vas.readFromStream(inputStream);
VectorContainer c = vas.get();
if (schema != null) {
c = SchemaUtil.coerceContainer(c, schema, context);
}
// logger.debug("Took {} us to read {} records", watch.elapsed(TimeUnit.MICROSECONDS),
// c.getRecordCount());
spilledBatches--;
currentContainer.zeroVectors();
Iterator<VectorWrapper<?>> wrapperIterator = c.iterator();
for (VectorWrapper w : currentContainer) {
TransferPair pair =
wrapperIterator.next().getValueVector().makeTransferPair(w.getValueVector());
pair.transfer();
}
currentContainer.setRecordCount(c.getRecordCount());
c.zeroVectors();
return c;
}
@Override
public void close() throws IOException {
currentContainer.zeroVectors();
if (sv2 != null) {
sv2.clear();
}
if (outputStream != null) {
outputStream.close();
}
if (inputStream != null) {
inputStream.close();
}
if (fs != null && fs.exists(path)) {
fs.delete(path, false);
}
}
@Override
public BatchSchema getSchema() {
return currentContainer.getSchema();
}
public void addField(ValueVector vector) {
container.add(vector);
fieldVectorMap.put(vector.getField(), vector);
}
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();
}
@Override
public IterOutcome innerNext() {
container.zeroVectors();
// if we got IterOutcome.NONE while getting partition vectors, and there are no batches on the
// queue, then we are
// done
if (upstreamNone && (batchQueue == null || batchQueue.size() == 0)) return IterOutcome.NONE;
// if there are batches on the queue, process them first, rather than calling incoming.next()
if (batchQueue != null && batchQueue.size() > 0) {
VectorContainer vc = batchQueue.poll();
recordCount = vc.getRecordCount();
try {
// Must set up a new schema each time, because ValueVectors are not reused between
// containers in queue
setupNewSchema(vc);
} catch (SchemaChangeException ex) {
kill(false);
logger.error("Failure during query", ex);
context.fail(ex);
return IterOutcome.STOP;
}
doWork(vc);
vc.zeroVectors();
return IterOutcome.OK_NEW_SCHEMA;
}
// Reaching this point, either this is the first iteration, or there are no batches left on the
// queue and there are
// more incoming
IterOutcome upstream = next(incoming);
if (this.first && upstream == IterOutcome.OK) {
throw new RuntimeException("Invalid state: First batch should have OK_NEW_SCHEMA");
}
// If this is the first iteration, we need to generate the partition vectors before we can
// proceed
if (this.first && upstream == IterOutcome.OK_NEW_SCHEMA) {
if (!getPartitionVectors()) {
cleanup();
return IterOutcome.STOP;
}
batchQueue = new LinkedBlockingQueue<>(this.sampledIncomingBatches);
first = false;
// Now that we have the partition vectors, we immediately process the first batch on the queue
VectorContainer vc = batchQueue.poll();
try {
setupNewSchema(vc);
} catch (SchemaChangeException ex) {
kill(false);
logger.error("Failure during query", ex);
context.fail(ex);
return IterOutcome.STOP;
}
doWork(vc);
vc.zeroVectors();
recordCount = vc.getRecordCount();
return IterOutcome.OK_NEW_SCHEMA;
}
// if this now that all the batches on the queue are processed, we begin processing the incoming
// batches. For the
// first one
// we need to generate a new schema, even if the outcome is IterOutcome.OK After that we can
// reuse the schema.
if (this.startedUnsampledBatches == false) {
this.startedUnsampledBatches = true;
if (upstream == IterOutcome.OK) upstream = IterOutcome.OK_NEW_SCHEMA;
}
switch (upstream) {
case NONE:
case NOT_YET:
case STOP:
cleanup();
recordCount = 0;
return upstream;
case OK_NEW_SCHEMA:
try {
setupNewSchema(incoming);
} catch (SchemaChangeException ex) {
kill(false);
logger.error("Failure during query", ex);
context.fail(ex);
return IterOutcome.STOP;
}
// fall through.
case OK:
doWork(incoming);
recordCount = incoming.getRecordCount();
return upstream; // change if upstream changed, otherwise normal.
default:
throw new UnsupportedOperationException();
}
}
public BatchGroup mergeAndSpill(LinkedList<BatchGroup> batchGroups) throws SchemaChangeException {
logger.debug("Copier allocator current allocation {}", copierAllocator.getAllocatedMemory());
logger.debug(
"mergeAndSpill: starting total size in memory = {}", oAllocator.getAllocatedMemory());
VectorContainer outputContainer = new VectorContainer();
List<BatchGroup> batchGroupList = Lists.newArrayList();
int batchCount = batchGroups.size();
for (int i = 0; i < batchCount / 2; i++) {
if (batchGroups.size() == 0) {
break;
}
BatchGroup batch = batchGroups.pollLast();
assert batch != null : "Encountered a null batch during merge and spill operation";
batchGroupList.add(batch);
}
if (batchGroupList.size() == 0) {
return null;
}
int estimatedRecordSize = 0;
for (VectorWrapper<?> w : batchGroupList.get(0)) {
try {
estimatedRecordSize += TypeHelper.getSize(w.getField().getType());
} catch (UnsupportedOperationException e) {
estimatedRecordSize += 50;
}
}
int targetRecordCount = Math.max(1, COPIER_BATCH_MEM_LIMIT / estimatedRecordSize);
VectorContainer hyperBatch = constructHyperBatch(batchGroupList);
createCopier(hyperBatch, batchGroupList, outputContainer, true);
int count = copier.next(targetRecordCount);
assert count > 0;
logger.debug(
"mergeAndSpill: estimated record size = {}, target record count = {}",
estimatedRecordSize,
targetRecordCount);
// 1 output container is kept in memory, so we want to hold on to it and transferClone
// allows keeping ownership
VectorContainer c1 = VectorContainer.getTransferClone(outputContainer, oContext);
c1.buildSchema(BatchSchema.SelectionVectorMode.NONE);
c1.setRecordCount(count);
String spillDir = dirs.next();
Path currSpillPath = new Path(Joiner.on("/").join(spillDir, fileName));
currSpillDirs.add(currSpillPath);
String outputFile = Joiner.on("/").join(currSpillPath, spillCount++);
try {
fs.deleteOnExit(currSpillPath);
} catch (IOException e) {
// since this is meant to be used in a batches's spilling, we don't propagate the exception
logger.warn("Unable to mark spill directory " + currSpillPath + " for deleting on exit", e);
}
stats.setLongStat(Metric.SPILL_COUNT, spillCount);
BatchGroup newGroup = new BatchGroup(c1, fs, outputFile, oContext);
try (AutoCloseable a = AutoCloseables.all(batchGroupList)) {
logger.info("Merging and spilling to {}", outputFile);
while ((count = copier.next(targetRecordCount)) > 0) {
outputContainer.buildSchema(BatchSchema.SelectionVectorMode.NONE);
outputContainer.setRecordCount(count);
// note that addBatch also clears the outputContainer
newGroup.addBatch(outputContainer);
}
injector.injectChecked(
context.getExecutionControls(), INTERRUPTION_WHILE_SPILLING, IOException.class);
newGroup.closeOutputStream();
} catch (Throwable e) {
// we only need to cleanup newGroup if spill failed
try {
AutoCloseables.close(e, newGroup);
} catch (Throwable t) {
/* close() may hit the same IO issue; just ignore */
}
throw UserException.resourceError(e)
.message("External Sort encountered an error while spilling to disk")
.addContext(e.getMessage() /* more detail */)
.build(logger);
} finally {
hyperBatch.clear();
}
logger.debug("mergeAndSpill: final total size in memory = {}", oAllocator.getAllocatedMemory());
logger.info("Completed spilling to {}", outputFile);
return newGroup;
}
@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);
}
}
@Override
public Iterator<VectorWrapper<?>> iterator() {
return currentContainer.iterator();
}
@Override
public void setNewSchema() throws SchemaChangeException {
container.buildSchema(SelectionVectorMode.NONE);
}
@Override
public TypedFieldId getValueVectorId(SchemaPath path) {
return currentContainer.getValueVectorId(path);
}
@Override
public VectorWrapper<?> getValueAccessorById(Class<?> clazz, int... ids) {
return currentContainer.getValueAccessorById(clazz, ids);
}
public Iterator<VectorWrapper<?>> iterator() {
return container.iterator();
}
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;
}