/**
* Called by execution pool to do query setup, and kick off remote execution.
*
* <p>Note that completion of this function is not the end of the Foreman's role in the query's
* lifecycle.
*/
@Override
public void run() {
// rename the thread we're using for debugging purposes
final Thread currentThread = Thread.currentThread();
final String originalName = currentThread.getName();
currentThread.setName(QueryIdHelper.getQueryId(queryId) + ":foreman");
// track how long the query takes
queryManager.markStartTime();
try {
injector.injectChecked(
queryContext.getExecutionControls(), "run-try-beginning", ForemanException.class);
queryText = queryRequest.getPlan();
// convert a run query request into action
switch (queryRequest.getType()) {
case LOGICAL:
parseAndRunLogicalPlan(queryRequest.getPlan());
break;
case PHYSICAL:
parseAndRunPhysicalPlan(queryRequest.getPlan());
break;
case SQL:
runSQL(queryRequest.getPlan());
break;
default:
throw new IllegalStateException();
}
injector.injectChecked(
queryContext.getExecutionControls(), "run-try-end", ForemanException.class);
} catch (final OutOfMemoryException | OutOfMemoryRuntimeException e) {
moveToState(QueryState.FAILED, UserException.memoryError(e).build(logger));
} catch (final ForemanException e) {
moveToState(QueryState.FAILED, e);
} catch (AssertionError | Exception ex) {
moveToState(
QueryState.FAILED,
new ForemanException(
"Unexpected exception during fragment initialization: " + ex.getMessage(), ex));
} catch (final OutOfMemoryError e) {
if ("Direct buffer memory".equals(e.getMessage())) {
moveToState(
QueryState.FAILED,
UserException.resourceError(e)
.message("One or more nodes ran out of memory while executing the query.")
.build(logger));
} else {
/*
* FragmentExecutors use a DrillbitStatusListener to watch out for the death of their query's Foreman. So, if we
* die here, they should get notified about that, and cancel themselves; we don't have to attempt to notify
* them, which might not work under these conditions.
*/
System.out.println("Node ran out of Heap memory, exiting.");
e.printStackTrace();
System.out.flush();
System.exit(-1);
}
} finally {
/*
* Begin accepting external events.
*
* Doing this here in the finally clause will guarantee that it occurs. Otherwise, if there
* is an exception anywhere during setup, it wouldn't occur, and any events that are generated
* as a result of any partial setup that was done (such as the FragmentSubmitListener,
* the ResponseSendListener, or an external call to cancel()), will hang the thread that makes the
* event delivery call.
*
* If we do throw an exception during setup, and have already moved to QueryState.FAILED, we just need to
* make sure that we can't make things any worse as those events are delivered, but allow
* any necessary remaining cleanup to proceed.
*
* Note that cancellations cannot be simulated before this point, i.e. pauses can be injected, because Foreman
* would wait on the cancelling thread to signal a resume and the cancelling thread would wait on the Foreman
* to accept events.
*/
acceptExternalEvents.countDown();
// If we received the resume signal before fragments are setup, the first call does not
// actually resume the
// fragments. Since setup is done, all fragments must have been delivered to remote nodes. Now
// we can resume.
if (resume) {
resume();
}
injector.injectPause(queryContext.getExecutionControls(), "foreman-ready", logger);
// restore the thread's original name
currentThread.setName(originalName);
}
/*
* Note that despite the run() completing, the Foreman continues to exist, and receives
* events (indirectly, through the QueryManager's use of stateListener), about fragment
* completions. It won't go away until everything is completed, failed, or cancelled.
*/
}
/**
* Set up the non-root fragments for execution. Some may be local, and some may be remote.
* Messages are sent immediately, so they may start returning data even before we complete this.
*
* @param fragments the fragments
* @throws ForemanException
*/
private void setupNonRootFragments(final Collection<PlanFragment> fragments)
throws ForemanException {
/*
* We will send a single message to each endpoint, regardless of how many fragments will be
* executed there. We need to start up the intermediate fragments first so that they will be
* ready once the leaf fragments start producing data. To satisfy both of these, we will
* make a pass through the fragments and put them into these two maps according to their
* leaf/intermediate state, as well as their target drillbit.
*/
final Multimap<DrillbitEndpoint, PlanFragment> leafFragmentMap = ArrayListMultimap.create();
final Multimap<DrillbitEndpoint, PlanFragment> intFragmentMap = ArrayListMultimap.create();
// record all fragments for status purposes.
for (final PlanFragment planFragment : fragments) {
logger.trace(
"Tracking intermediate remote node {} with data {}",
planFragment.getAssignment(),
planFragment.getFragmentJson());
queryManager.addFragmentStatusTracker(planFragment, false);
if (planFragment.getLeafFragment()) {
leafFragmentMap.put(planFragment.getAssignment(), planFragment);
} else {
intFragmentMap.put(planFragment.getAssignment(), planFragment);
}
}
/*
* We need to wait for the intermediates to be sent so that they'll be set up by the time
* the leaves start producing data. We'll use this latch to wait for the responses.
*
* However, in order not to hang the process if any of the RPC requests fails, we always
* count down (see FragmentSubmitFailures), but we count the number of failures so that we'll
* know if any submissions did fail.
*/
final int numIntFragments = intFragmentMap.keySet().size();
final ExtendedLatch endpointLatch = new ExtendedLatch(numIntFragments);
final FragmentSubmitFailures fragmentSubmitFailures = new FragmentSubmitFailures();
// send remote intermediate fragments
for (final DrillbitEndpoint ep : intFragmentMap.keySet()) {
sendRemoteFragments(ep, intFragmentMap.get(ep), endpointLatch, fragmentSubmitFailures);
}
final long timeout = RPC_WAIT_IN_MSECS_PER_FRAGMENT * numIntFragments;
if (numIntFragments > 0 && !endpointLatch.awaitUninterruptibly(timeout)) {
long numberRemaining = endpointLatch.getCount();
throw UserException.connectionError()
.message(
"Exceeded timeout (%d) while waiting send intermediate work fragments to remote nodes. "
+ "Sent %d and only heard response back from %d nodes.",
timeout, numIntFragments, numIntFragments - numberRemaining)
.build(logger);
}
// if any of the intermediate fragment submissions failed, fail the query
final List<FragmentSubmitFailures.SubmissionException> submissionExceptions =
fragmentSubmitFailures.submissionExceptions;
if (submissionExceptions.size() > 0) {
Set<DrillbitEndpoint> endpoints = Sets.newHashSet();
StringBuilder sb = new StringBuilder();
boolean first = true;
for (FragmentSubmitFailures.SubmissionException e :
fragmentSubmitFailures.submissionExceptions) {
DrillbitEndpoint endpoint = e.drillbitEndpoint;
if (endpoints.add(endpoint)) {
if (first) {
first = false;
} else {
sb.append(", ");
}
sb.append(endpoint.getAddress());
}
}
throw UserException.connectionError(submissionExceptions.get(0).rpcException)
.message("Error setting up remote intermediate fragment execution")
.addContext("Nodes with failures", sb.toString())
.build(logger);
}
injector.injectChecked(
queryContext.getExecutionControls(), "send-fragments", ForemanException.class);
/*
* Send the remote (leaf) fragments; we don't wait for these. Any problems will come in through
* the regular sendListener event delivery.
*/
for (final DrillbitEndpoint ep : leafFragmentMap.keySet()) {
sendRemoteFragments(ep, leafFragmentMap.get(ep), null, null);
}
}
@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);
}
}
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 void run() {
// if a cancel thread has already entered this executor, we have not reason to continue.
if (!hasCloseoutThread.compareAndSet(false, true)) {
return;
}
final Thread myThread = Thread.currentThread();
myThreadRef.set(myThread);
final String originalThreadName = myThread.getName();
final FragmentHandle fragmentHandle = fragmentContext.getHandle();
final DrillbitContext drillbitContext = fragmentContext.getDrillbitContext();
final ClusterCoordinator clusterCoordinator = drillbitContext.getClusterCoordinator();
final DrillbitStatusListener drillbitStatusListener = new FragmentDrillbitStatusListener();
final String newThreadName = QueryIdHelper.getExecutorThreadName(fragmentHandle);
try {
myThread.setName(newThreadName);
// if we didn't get the root operator when the executor was created, create it now.
final FragmentRoot rootOperator =
this.rootOperator != null
? this.rootOperator
: drillbitContext.getPlanReader().readFragmentOperator(fragment.getFragmentJson());
root = ImplCreator.getExec(fragmentContext, rootOperator);
if (root == null) {
return;
}
clusterCoordinator.addDrillbitStatusListener(drillbitStatusListener);
updateState(FragmentState.RUNNING);
acceptExternalEvents.countDown();
injector.injectPause(fragmentContext.getExecutionControls(), "fragment-running", logger);
final DrillbitEndpoint endpoint = drillbitContext.getEndpoint();
logger.debug(
"Starting fragment {}:{} on {}:{}",
fragmentHandle.getMajorFragmentId(),
fragmentHandle.getMinorFragmentId(),
endpoint.getAddress(),
endpoint.getUserPort());
final UserGroupInformation queryUserUgi =
fragmentContext.isImpersonationEnabled()
? ImpersonationUtil.createProxyUgi(fragmentContext.getQueryUserName())
: ImpersonationUtil.getProcessUserUGI();
queryUserUgi.doAs(
new PrivilegedExceptionAction<Void>() {
public Void run() throws Exception {
injector.injectChecked(
fragmentContext.getExecutionControls(), "fragment-execution", IOException.class);
/*
* Run the query until root.next returns false OR we no longer need to continue.
*/
while (shouldContinue() && root.next()) {
// loop
}
return null;
}
});
} catch (OutOfMemoryError | OutOfMemoryException e) {
if (!(e instanceof OutOfMemoryError) || "Direct buffer memory".equals(e.getMessage())) {
fail(UserException.memoryError(e).build(logger));
} else {
// we have a heap out of memory error. The JVM in unstable, exit.
CatastrophicFailure.exit(
e, "Unable to handle out of memory condition in FragmentExecutor.", -2);
}
} catch (AssertionError | Exception e) {
fail(e);
} finally {
// no longer allow this thread to be interrupted. We synchronize here to make sure that cancel
// can't set an
// interruption after we have moved beyond this block.
synchronized (myThreadRef) {
myThreadRef.set(null);
Thread.interrupted();
}
// We need to sure we countDown at least once. We'll do it here to guarantee that.
acceptExternalEvents.countDown();
// here we could be in FAILED, RUNNING, or CANCELLATION_REQUESTED
cleanup(FragmentState.FINISHED);
clusterCoordinator.removeDrillbitStatusListener(drillbitStatusListener);
myThread.setName(originalThreadName);
}
}
