use of com.facebook.presto.operator.TaskContext in project presto by prestodb.
the class TestSpatialJoinOperator method testDistanceQuery.
@Test
public void testDistanceQuery() {
TaskContext taskContext = createTaskContext();
DriverContext driverContext = taskContext.addPipelineContext(0, true, true, false).addDriverContext();
RowPagesBuilder buildPages = rowPagesBuilder(ImmutableList.of(GEOMETRY, VARCHAR, DOUBLE)).row(stPoint(0, 0), "0_0", 1.5).row(null, "null", 1.5).row(stPoint(1, 0), "1_0", 1.5).pageBreak().row(stPoint(3, 0), "3_0", 1.5).pageBreak().row(stPoint(10, 0), "10_0", 1.5);
PagesSpatialIndexFactory pagesSpatialIndexFactory = buildIndex(driverContext, (build, probe, r) -> build.distance(probe) <= r.getAsDouble(), Optional.of(2), Optional.empty(), buildPages);
RowPagesBuilder probePages = rowPagesBuilder(ImmutableList.of(GEOMETRY, VARCHAR)).row(stPoint(0, 1), "0_1").row(null, "null").row(stPoint(1, 1), "1_1").pageBreak().row(stPoint(3, 1), "3_1").pageBreak().row(stPoint(10, 1), "10_1");
OperatorFactory joinOperatorFactory = new SpatialJoinOperatorFactory(2, new PlanNodeId("test"), INNER, probePages.getTypes(), Ints.asList(1), 0, Optional.empty(), pagesSpatialIndexFactory);
// Make sure that spatial index reference counting works with duplicate factories
joinOperatorFactory.duplicate().noMoreOperators();
MaterializedResult expected = resultBuilder(taskContext.getSession(), ImmutableList.of(VARCHAR, VARCHAR)).row("0_1", "0_0").row("0_1", "1_0").row("1_1", "0_0").row("1_1", "1_0").row("3_1", "3_0").row("10_1", "10_0").build();
assertOperatorEqualsIgnoreOrder(joinOperatorFactory, driverContext, probePages.build(), expected);
}
use of com.facebook.presto.operator.TaskContext in project presto by prestodb.
the class TestSqlTaskExecution method testSimple.
@Test(dataProvider = "executionStrategies", timeOut = 20_000)
public void testSimple(PipelineExecutionStrategy executionStrategy) throws Exception {
ScheduledExecutorService taskNotificationExecutor = newScheduledThreadPool(10, threadsNamed("task-notification-%s"));
ScheduledExecutorService driverYieldExecutor = newScheduledThreadPool(2, threadsNamed("driver-yield-%s"));
TaskExecutor taskExecutor = new TaskExecutor(5, 10, 3, 4, TASK_FAIR, Ticker.systemTicker());
taskExecutor.start();
try {
TaskStateMachine taskStateMachine = new TaskStateMachine(TASK_ID, taskNotificationExecutor);
PartitionedOutputBuffer outputBuffer = newTestingOutputBuffer(taskNotificationExecutor);
OutputBufferConsumer outputBufferConsumer = new OutputBufferConsumer(outputBuffer, OUTPUT_BUFFER_ID);
//
// test initialization: simple task with 1 pipeline
//
// pipeline 0 ... pipeline id
// partitioned ... partitioned/unpartitioned pipeline
// grouped ... execution strategy (in grouped test)
// ungrouped ... execution strategy (in ungrouped test)
//
// TaskOutput
// |
// Scan
//
// See #testComplex for all the bahaviors that are tested. Not all of them apply here.
TestingScanOperatorFactory testingScanOperatorFactory = new TestingScanOperatorFactory(0, TABLE_SCAN_NODE_ID, ImmutableList.of(VARCHAR));
TaskOutputOperatorFactory taskOutputOperatorFactory = new TaskOutputOperatorFactory(1, TABLE_SCAN_NODE_ID, outputBuffer, Function.identity(), new PagesSerdeFactory(new BlockEncodingManager(), false));
LocalExecutionPlan localExecutionPlan = new LocalExecutionPlan(ImmutableList.of(new DriverFactory(0, true, true, ImmutableList.of(testingScanOperatorFactory, taskOutputOperatorFactory), OptionalInt.empty(), executionStrategy, Optional.empty())), ImmutableList.of(TABLE_SCAN_NODE_ID), executionStrategy == GROUPED_EXECUTION ? StageExecutionDescriptor.fixedLifespanScheduleGroupedExecution(ImmutableList.of(TABLE_SCAN_NODE_ID), 8) : StageExecutionDescriptor.ungroupedExecution());
TaskContext taskContext = newTestingTaskContext(taskNotificationExecutor, driverYieldExecutor, taskStateMachine);
SqlTaskExecution sqlTaskExecution = SqlTaskExecution.createSqlTaskExecution(taskStateMachine, taskContext, outputBuffer, ImmutableList.of(), localExecutionPlan, taskExecutor, taskNotificationExecutor, createTestSplitMonitor());
//
// test body
assertEquals(taskStateMachine.getState(), TaskState.RUNNING);
switch(executionStrategy) {
case UNGROUPED_EXECUTION:
// add source for pipeline
sqlTaskExecution.addSources(ImmutableList.of(new TaskSource(TABLE_SCAN_NODE_ID, ImmutableSet.of(newScheduledSplit(0, TABLE_SCAN_NODE_ID, Lifespan.taskWide(), 100000, 123)), false)));
// assert that partial task result is produced
outputBufferConsumer.consume(123, ASSERT_WAIT_TIMEOUT);
// pause operator execution to make sure that
// * operatorFactory will be closed even though operator can't execute
// * completedDriverGroups will NOT include the newly scheduled driver group while pause is in place
testingScanOperatorFactory.getPauser().pause();
// add source for pipeline, mark as no more splits
sqlTaskExecution.addSources(ImmutableList.of(new TaskSource(TABLE_SCAN_NODE_ID, ImmutableSet.of(newScheduledSplit(1, TABLE_SCAN_NODE_ID, Lifespan.taskWide(), 200000, 300), newScheduledSplit(2, TABLE_SCAN_NODE_ID, Lifespan.taskWide(), 300000, 200)), true)));
// assert that pipeline will have no more drivers
waitUntilEquals(testingScanOperatorFactory::isOverallNoMoreOperators, true, ASSERT_WAIT_TIMEOUT);
// assert that no DriverGroup is fully completed
assertEquals(taskContext.getCompletedDriverGroups(), ImmutableSet.of());
// resume operator execution
testingScanOperatorFactory.getPauser().resume();
// assert that task result is produced
outputBufferConsumer.consume(300 + 200, ASSERT_WAIT_TIMEOUT);
outputBufferConsumer.assertBufferComplete(ASSERT_WAIT_TIMEOUT);
break;
case GROUPED_EXECUTION:
// add source for pipeline (driver group [1, 5]), mark driver group [1] as noMoreSplits
sqlTaskExecution.addSources(ImmutableList.of(new TaskSource(TABLE_SCAN_NODE_ID, ImmutableSet.of(newScheduledSplit(0, TABLE_SCAN_NODE_ID, Lifespan.driverGroup(1), 0, 1), newScheduledSplit(1, TABLE_SCAN_NODE_ID, Lifespan.driverGroup(5), 100000, 10)), ImmutableSet.of(Lifespan.driverGroup(1)), false)));
// assert that pipeline will have no more drivers for driver group [1]
waitUntilEquals(testingScanOperatorFactory::getDriverGroupsWithNoMoreOperators, ImmutableSet.of(Lifespan.driverGroup(1)), ASSERT_WAIT_TIMEOUT);
// assert that partial result is produced for both driver groups
outputBufferConsumer.consume(1 + 10, ASSERT_WAIT_TIMEOUT);
// assert that driver group [1] is fully completed
waitUntilEquals(taskContext::getCompletedDriverGroups, ImmutableSet.of(Lifespan.driverGroup(1)), ASSERT_WAIT_TIMEOUT);
// pause operator execution to make sure that
// * operatorFactory will be closed even though operator can't execute
// * completedDriverGroups will NOT include the newly scheduled driver group while pause is in place
testingScanOperatorFactory.getPauser().pause();
// add source for pipeline (driver group [5]), mark driver group [5] as noMoreSplits
sqlTaskExecution.addSources(ImmutableList.of(new TaskSource(TABLE_SCAN_NODE_ID, ImmutableSet.of(newScheduledSplit(2, TABLE_SCAN_NODE_ID, Lifespan.driverGroup(5), 200000, 300)), ImmutableSet.of(Lifespan.driverGroup(5)), false)));
// assert that pipeline will have no more drivers for driver group [1, 5]
waitUntilEquals(testingScanOperatorFactory::getDriverGroupsWithNoMoreOperators, ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5)), ASSERT_WAIT_TIMEOUT);
// assert that driver group [5] is NOT YET fully completed
assertEquals(taskContext.getCompletedDriverGroups(), ImmutableSet.of(Lifespan.driverGroup(1)));
// resume operator execution
testingScanOperatorFactory.getPauser().resume();
// assert that partial result is produced
outputBufferConsumer.consume(300, ASSERT_WAIT_TIMEOUT);
// assert that driver group [1, 5] is fully completed
waitUntilEquals(taskContext::getCompletedDriverGroups, ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5)), ASSERT_WAIT_TIMEOUT);
// pause operator execution to make sure that
// * operatorFactory will be closed even though operator can't execute
// * completedDriverGroups will NOT include the newly scheduled driver group while pause is in place
testingScanOperatorFactory.getPauser().pause();
// add source for pipeline (driver group [7]), mark pipeline as noMoreSplits without explicitly marking driver group 7
sqlTaskExecution.addSources(ImmutableList.of(new TaskSource(TABLE_SCAN_NODE_ID, ImmutableSet.of(newScheduledSplit(3, TABLE_SCAN_NODE_ID, Lifespan.driverGroup(7), 300000, 45), newScheduledSplit(4, TABLE_SCAN_NODE_ID, Lifespan.driverGroup(7), 400000, 54)), ImmutableSet.of(), true)));
// assert that pipeline will have no more drivers for driver group [1, 5, 7]
waitUntilEquals(testingScanOperatorFactory::getDriverGroupsWithNoMoreOperators, ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5), Lifespan.driverGroup(7)), ASSERT_WAIT_TIMEOUT);
// assert that pipeline will have no more drivers
waitUntilEquals(testingScanOperatorFactory::isOverallNoMoreOperators, true, ASSERT_WAIT_TIMEOUT);
// assert that driver group [1, 5] is fully completed
assertEquals(taskContext.getCompletedDriverGroups(), ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5)));
// resume operator execution
testingScanOperatorFactory.getPauser().resume();
// assert driver group [7] is not completed before output buffer is consumed
MILLISECONDS.sleep(1000);
assertEquals(taskContext.getCompletedDriverGroups(), ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5)));
// assert that result is produced
outputBufferConsumer.consume(45 + 54, ASSERT_WAIT_TIMEOUT);
outputBufferConsumer.assertBufferComplete(ASSERT_WAIT_TIMEOUT);
// assert that driver group [1, 5, 7] is fully completed
waitUntilEquals(taskContext::getCompletedDriverGroups, ImmutableSet.of(Lifespan.driverGroup(1), Lifespan.driverGroup(5), Lifespan.driverGroup(7)), ASSERT_WAIT_TIMEOUT);
break;
default:
throw new UnsupportedOperationException();
}
// complete the task by calling abort on it
outputBufferConsumer.abort();
TaskState taskState = taskStateMachine.getStateChange(TaskState.RUNNING).get(10, SECONDS);
assertEquals(taskState, TaskState.FINISHED);
} finally {
taskExecutor.stop();
taskNotificationExecutor.shutdownNow();
driverYieldExecutor.shutdown();
}
}
use of com.facebook.presto.operator.TaskContext in project presto by prestodb.
the class SqlTask method createTaskStatus.
private TaskStatus createTaskStatus(TaskHolder taskHolder) {
long taskStatusAgeInMilis = System.currentTimeMillis() - creationTimeInMillis;
// Always return a new TaskInfo with a larger version number;
// otherwise a client will not accept the update
long versionNumber = nextTaskInfoVersion.getAndIncrement();
TaskState state = taskStateMachine.getState();
List<ExecutionFailureInfo> failures = ImmutableList.of();
if (state == FAILED) {
failures = toFailures(taskStateMachine.getFailureCauses());
}
int queuedPartitionedDrivers = 0;
long queuedPartitionedSplitsWeight = 0L;
int runningPartitionedDrivers = 0;
long runningPartitionedSplitsWeight = 0L;
long physicalWrittenDataSizeInBytes = 0L;
long userMemoryReservationInBytes = 0L;
long systemMemoryReservationInBytes = 0L;
// TODO: add a mechanism to avoid sending the whole completedDriverGroups set over the wire for every task status reply
Set<Lifespan> completedDriverGroups = ImmutableSet.of();
long fullGcCount = 0;
long fullGcTimeInMillis = 0L;
long totalCpuTimeInNanos = 0L;
if (taskHolder.getFinalTaskInfo() != null) {
TaskStats taskStats = taskHolder.getFinalTaskInfo().getStats();
queuedPartitionedDrivers = taskStats.getQueuedPartitionedDrivers();
queuedPartitionedSplitsWeight = taskStats.getQueuedPartitionedSplitsWeight();
runningPartitionedDrivers = taskStats.getRunningPartitionedDrivers();
runningPartitionedSplitsWeight = taskStats.getRunningPartitionedSplitsWeight();
physicalWrittenDataSizeInBytes = taskStats.getPhysicalWrittenDataSizeInBytes();
userMemoryReservationInBytes = taskStats.getUserMemoryReservationInBytes();
systemMemoryReservationInBytes = taskStats.getSystemMemoryReservationInBytes();
fullGcCount = taskStats.getFullGcCount();
fullGcTimeInMillis = taskStats.getFullGcTimeInMillis();
totalCpuTimeInNanos = taskStats.getTotalCpuTimeInNanos();
} else if (taskHolder.getTaskExecution() != null) {
long physicalWrittenBytes = 0;
TaskContext taskContext = taskHolder.getTaskExecution().getTaskContext();
for (PipelineContext pipelineContext : taskContext.getPipelineContexts()) {
PipelineStatus pipelineStatus = pipelineContext.getPipelineStatus();
queuedPartitionedDrivers += pipelineStatus.getQueuedPartitionedDrivers();
queuedPartitionedSplitsWeight += pipelineStatus.getQueuedPartitionedSplitsWeight();
runningPartitionedDrivers += pipelineStatus.getRunningPartitionedDrivers();
runningPartitionedSplitsWeight += pipelineStatus.getRunningPartitionedSplitsWeight();
physicalWrittenBytes += pipelineContext.getPhysicalWrittenDataSize();
totalCpuTimeInNanos += pipelineContext.getPipelineStats().getTotalCpuTimeInNanos();
}
physicalWrittenDataSizeInBytes = physicalWrittenBytes;
userMemoryReservationInBytes = taskContext.getMemoryReservation().toBytes();
systemMemoryReservationInBytes = taskContext.getSystemMemoryReservation().toBytes();
completedDriverGroups = taskContext.getCompletedDriverGroups();
fullGcCount = taskContext.getFullGcCount();
fullGcTimeInMillis = taskContext.getFullGcTime().toMillis();
}
return new TaskStatus(taskInstanceId.getUuidLeastSignificantBits(), taskInstanceId.getUuidMostSignificantBits(), versionNumber, state, location, completedDriverGroups, failures, queuedPartitionedDrivers, runningPartitionedDrivers, outputBuffer.getUtilization(), isOutputBufferOverutilized(), physicalWrittenDataSizeInBytes, userMemoryReservationInBytes, systemMemoryReservationInBytes, queryContext.getPeakNodeTotalMemory(), fullGcCount, fullGcTimeInMillis, totalCpuTimeInNanos, taskStatusAgeInMilis, queuedPartitionedSplitsWeight, runningPartitionedSplitsWeight);
}
use of com.facebook.presto.operator.TaskContext in project presto by prestodb.
the class TestMemoryRevokingScheduler method testMemoryPoolRevoking.
@Test
public void testMemoryPoolRevoking() throws Exception {
QueryContext q1 = getOrCreateQueryContext(new QueryId("q1"), memoryPool);
QueryContext q2 = getOrCreateQueryContext(new QueryId("q2"), memoryPool);
SqlTask sqlTask1 = newSqlTask(q1.getQueryId(), memoryPool);
SqlTask sqlTask2 = newSqlTask(q2.getQueryId(), memoryPool);
TaskContext taskContext1 = getOrCreateTaskContext(sqlTask1);
PipelineContext pipelineContext11 = taskContext1.addPipelineContext(0, false, false, false);
DriverContext driverContext111 = pipelineContext11.addDriverContext();
OperatorContext operatorContext1 = driverContext111.addOperatorContext(1, new PlanNodeId("na"), "na");
OperatorContext operatorContext2 = driverContext111.addOperatorContext(2, new PlanNodeId("na"), "na");
DriverContext driverContext112 = pipelineContext11.addDriverContext();
OperatorContext operatorContext3 = driverContext112.addOperatorContext(3, new PlanNodeId("na"), "na");
TaskContext taskContext2 = getOrCreateTaskContext(sqlTask2);
PipelineContext pipelineContext21 = taskContext2.addPipelineContext(1, false, false, false);
DriverContext driverContext211 = pipelineContext21.addDriverContext();
OperatorContext operatorContext4 = driverContext211.addOperatorContext(4, new PlanNodeId("na"), "na");
OperatorContext operatorContext5 = driverContext211.addOperatorContext(5, new PlanNodeId("na"), "na");
List<SqlTask> tasks = ImmutableList.of(sqlTask1, sqlTask2);
MemoryRevokingScheduler scheduler = new MemoryRevokingScheduler(singletonList(memoryPool), () -> tasks, queryContexts::get, 1.0, 1.0, ORDER_BY_CREATE_TIME, false);
try {
scheduler.start();
allOperatorContexts = ImmutableSet.of(operatorContext1, operatorContext2, operatorContext3, operatorContext4, operatorContext5);
assertMemoryRevokingNotRequested();
assertEquals(10, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
assertMemoryRevokingNotRequested();
LocalMemoryContext revocableMemory1 = operatorContext1.localRevocableMemoryContext();
LocalMemoryContext revocableMemory3 = operatorContext3.localRevocableMemoryContext();
LocalMemoryContext revocableMemory4 = operatorContext4.localRevocableMemoryContext();
LocalMemoryContext revocableMemory5 = operatorContext5.localRevocableMemoryContext();
revocableMemory1.setBytes(3);
revocableMemory3.setBytes(6);
assertEquals(1, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// we are still good - no revoking needed
assertMemoryRevokingNotRequested();
revocableMemory4.setBytes(7);
assertEquals(-6, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// we need to revoke 3 and 6
assertMemoryRevokingRequestedFor(operatorContext1, operatorContext3);
// lets revoke some bytes
revocableMemory1.setBytes(0);
operatorContext1.resetMemoryRevokingRequested();
scheduler.awaitAsynchronousCallbacksRun();
assertMemoryRevokingRequestedFor(operatorContext3);
assertEquals(-3, memoryPool.getFreeBytes());
// and allocate some more
revocableMemory5.setBytes(3);
assertEquals(-6, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// we are still good with just OC3 in process of revoking
assertMemoryRevokingRequestedFor(operatorContext3);
// and allocate some more
revocableMemory5.setBytes(4);
assertEquals(-7, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// now we have to trigger revoking for OC4
assertMemoryRevokingRequestedFor(operatorContext3, operatorContext4);
} finally {
scheduler.stop();
}
}
use of com.facebook.presto.operator.TaskContext in project presto by prestodb.
the class TestMemoryRevokingScheduler method testRevokesPoolWhenFullBeforeQueryLimit.
@Test
public void testRevokesPoolWhenFullBeforeQueryLimit() throws Exception {
QueryContext q1 = getOrCreateQueryContext(new QueryId("q1"), memoryPool);
QueryContext q2 = getOrCreateQueryContext(new QueryId("q2"), memoryPool);
SqlTask sqlTask1 = newSqlTask(q1.getQueryId(), memoryPool);
SqlTask sqlTask2 = newSqlTask(q2.getQueryId(), memoryPool);
TaskContext taskContext1 = getOrCreateTaskContext(sqlTask1);
PipelineContext pipelineContext11 = taskContext1.addPipelineContext(0, false, false, false);
DriverContext driverContext111 = pipelineContext11.addDriverContext();
OperatorContext operatorContext1 = driverContext111.addOperatorContext(1, new PlanNodeId("na"), "na");
OperatorContext operatorContext2 = driverContext111.addOperatorContext(2, new PlanNodeId("na"), "na");
DriverContext driverContext112 = pipelineContext11.addDriverContext();
OperatorContext operatorContext3 = driverContext112.addOperatorContext(3, new PlanNodeId("na"), "na");
TaskContext taskContext2 = getOrCreateTaskContext(sqlTask2);
PipelineContext pipelineContext21 = taskContext2.addPipelineContext(1, false, false, false);
DriverContext driverContext211 = pipelineContext21.addDriverContext();
OperatorContext operatorContext4 = driverContext211.addOperatorContext(4, new PlanNodeId("na"), "na");
List<SqlTask> tasks = ImmutableList.of(sqlTask1, sqlTask2);
MemoryRevokingScheduler scheduler = new MemoryRevokingScheduler(singletonList(memoryPool), () -> tasks, queryContexts::get, 1.0, 1.0, ORDER_BY_CREATE_TIME, true);
try {
scheduler.start();
allOperatorContexts = ImmutableSet.of(operatorContext1, operatorContext2, operatorContext3, operatorContext4);
assertMemoryRevokingNotRequested();
assertEquals(10, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
assertMemoryRevokingNotRequested();
LocalMemoryContext revocableMemory1 = operatorContext1.localRevocableMemoryContext();
LocalMemoryContext revocableMemory3 = operatorContext3.localRevocableMemoryContext();
LocalMemoryContext revocableMemory4 = operatorContext4.localRevocableMemoryContext();
revocableMemory1.setBytes(3);
revocableMemory3.setBytes(6);
assertEquals(1, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// we are still good - no revoking needed
assertMemoryRevokingNotRequested();
revocableMemory4.setBytes(7);
assertEquals(-6, memoryPool.getFreeBytes());
scheduler.awaitAsynchronousCallbacksRun();
// we need to revoke 3 and 6
assertMemoryRevokingRequestedFor(operatorContext1, operatorContext3);
} finally {
scheduler.stop();
}
}
Aggregations