Search in sources :

Example 1 with TestingSplit

use of io.trino.testing.TestingSplit in project trino by trinodb.

the class TestSourcePartitionedScheduler method testStageBalancedSplitAssignment.

@Test
public void testStageBalancedSplitAssignment() {
    // use private node manager so we can add a node later
    InMemoryNodeManager nodeManager = new InMemoryNodeManager();
    nodeManager.addNode(CONNECTOR_ID, new InternalNode("other1", URI.create("http://127.0.0.1:11"), NodeVersion.UNKNOWN, false), new InternalNode("other2", URI.create("http://127.0.0.1:12"), NodeVersion.UNKNOWN, false), new InternalNode("other3", URI.create("http://127.0.0.1:13"), NodeVersion.UNKNOWN, false));
    NodeTaskMap nodeTaskMap = new NodeTaskMap(finalizerService);
    // Schedule 15 splits - there are 3 nodes, each node should get 5 splits
    PlanFragment firstPlan = createFragment();
    StageExecution firstStage = createStageExecution(firstPlan, nodeTaskMap);
    QueuedSplitSource firstSplitSource = new QueuedSplitSource(TestingSplit::createRemoteSplit);
    StageScheduler firstScheduler = getSourcePartitionedScheduler(firstSplitSource, firstStage, nodeManager, nodeTaskMap, 200, STAGE);
    firstSplitSource.addSplits(15);
    ScheduleResult scheduleResult = firstScheduler.schedule();
    assertTrue(scheduleResult.getBlocked().isDone());
    assertEquals(scheduleResult.getNewTasks().size(), 3);
    assertEquals(firstStage.getAllTasks().size(), 3);
    for (RemoteTask remoteTask : firstStage.getAllTasks()) {
        PartitionedSplitsInfo splitsInfo = remoteTask.getPartitionedSplitsInfo();
        assertEquals(splitsInfo.getCount(), 5);
    }
    // Add new node
    InternalNode additionalNode = new InternalNode("other4", URI.create("http://127.0.0.1:14"), NodeVersion.UNKNOWN, false);
    nodeManager.addNode(CONNECTOR_ID, additionalNode);
    // Schedule 5 splits in first query. Since the new node does not have any splits, all 5 splits are assigned to the new node
    firstSplitSource.addSplits(5);
    firstSplitSource.close();
    scheduleResult = firstScheduler.schedule();
    assertEffectivelyFinished(scheduleResult, firstScheduler);
    assertTrue(scheduleResult.getBlocked().isDone());
    assertEquals(scheduleResult.getNewTasks().size(), 1);
    assertEquals(firstStage.getAllTasks().size(), 4);
    for (RemoteTask remoteTask : firstStage.getAllTasks()) {
        PartitionedSplitsInfo splitsInfo = remoteTask.getPartitionedSplitsInfo();
        assertEquals(splitsInfo.getCount(), 5);
    }
    // Add new node
    InternalNode anotherAdditionalNode = new InternalNode("other5", URI.create("http://127.0.0.1:15"), NodeVersion.UNKNOWN, false);
    nodeManager.addNode(CONNECTOR_ID, anotherAdditionalNode);
    // Schedule 5 splits in another query. New query should be balanced across all nodes
    PlanFragment secondPlan = createFragment();
    StageExecution secondStage = createStageExecution(secondPlan, nodeTaskMap);
    StageScheduler secondScheduler = getSourcePartitionedScheduler(createFixedSplitSource(5, TestingSplit::createRemoteSplit), secondStage, nodeManager, nodeTaskMap, 200, STAGE);
    scheduleResult = secondScheduler.schedule();
    assertEffectivelyFinished(scheduleResult, secondScheduler);
    assertEquals(secondStage.getAllTasks().size(), 5);
    for (RemoteTask remoteTask : secondStage.getAllTasks()) {
        PartitionedSplitsInfo splitsInfo = remoteTask.getPartitionedSplitsInfo();
        assertEquals(splitsInfo.getCount(), 1);
    }
    firstStage.abort();
    secondStage.abort();
}
Also used : NodeTaskMap(io.trino.execution.NodeTaskMap) PipelinedStageExecution.createPipelinedStageExecution(io.trino.execution.scheduler.PipelinedStageExecution.createPipelinedStageExecution) PartitionedSplitsInfo(io.trino.execution.PartitionedSplitsInfo) MockRemoteTask(io.trino.execution.MockRemoteTaskFactory.MockRemoteTask) RemoteTask(io.trino.execution.RemoteTask) InternalNode(io.trino.metadata.InternalNode) TestingSplit(io.trino.testing.TestingSplit) PlanFragment(io.trino.sql.planner.PlanFragment) InMemoryNodeManager(io.trino.metadata.InMemoryNodeManager) SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler(io.trino.execution.scheduler.SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler) Test(org.testng.annotations.Test)

Example 2 with TestingSplit

use of io.trino.testing.TestingSplit in project trino by trinodb.

the class TestSourcePartitionedScheduler method testNewTaskScheduledWhenChildStageBufferIsUnderutilized.

@Test
public void testNewTaskScheduledWhenChildStageBufferIsUnderutilized() {
    NodeTaskMap nodeTaskMap = new NodeTaskMap(finalizerService);
    // use private node manager so we can add a node later
    InMemoryNodeManager nodeManager = new InMemoryNodeManager();
    nodeManager.addNode(CONNECTOR_ID, new InternalNode("other1", URI.create("http://127.0.0.1:11"), NodeVersion.UNKNOWN, false), new InternalNode("other2", URI.create("http://127.0.0.1:12"), NodeVersion.UNKNOWN, false), new InternalNode("other3", URI.create("http://127.0.0.1:13"), NodeVersion.UNKNOWN, false));
    NodeScheduler nodeScheduler = new NodeScheduler(new UniformNodeSelectorFactory(nodeManager, new NodeSchedulerConfig().setIncludeCoordinator(false), nodeTaskMap, new Duration(0, SECONDS)));
    PlanFragment plan = createFragment();
    StageExecution stage = createStageExecution(plan, nodeTaskMap);
    // setting under utilized child output buffer
    StageScheduler scheduler = newSourcePartitionedSchedulerAsStageScheduler(stage, TABLE_SCAN_NODE_ID, new ConnectorAwareSplitSource(CONNECTOR_ID, createFixedSplitSource(500, TestingSplit::createRemoteSplit)), new DynamicSplitPlacementPolicy(nodeScheduler.createNodeSelector(session, Optional.of(CONNECTOR_ID)), stage::getAllTasks), 500, new DynamicFilterService(metadata, functionManager, typeOperators, new DynamicFilterConfig()), new TableExecuteContextManager(), () -> false);
    // the queues of 3 running nodes should be full
    ScheduleResult scheduleResult = scheduler.schedule();
    assertEquals(scheduleResult.getBlockedReason().get(), SPLIT_QUEUES_FULL);
    assertEquals(scheduleResult.getNewTasks().size(), 3);
    assertEquals(scheduleResult.getSplitsScheduled(), 300);
    for (RemoteTask remoteTask : scheduleResult.getNewTasks()) {
        PartitionedSplitsInfo splitsInfo = remoteTask.getPartitionedSplitsInfo();
        assertEquals(splitsInfo.getCount(), 100);
    }
    // new node added - the pending splits should go to it since the child tasks are not blocked
    nodeManager.addNode(CONNECTOR_ID, new InternalNode("other4", URI.create("http://127.0.0.4:14"), NodeVersion.UNKNOWN, false));
    scheduleResult = scheduler.schedule();
    // split queue is full but still the source task creation isn't blocked
    assertEquals(scheduleResult.getBlockedReason().get(), SPLIT_QUEUES_FULL);
    assertEquals(scheduleResult.getNewTasks().size(), 1);
    assertEquals(scheduleResult.getSplitsScheduled(), 100);
}
Also used : NodeTaskMap(io.trino.execution.NodeTaskMap) PipelinedStageExecution.createPipelinedStageExecution(io.trino.execution.scheduler.PipelinedStageExecution.createPipelinedStageExecution) PartitionedSplitsInfo(io.trino.execution.PartitionedSplitsInfo) MockRemoteTask(io.trino.execution.MockRemoteTaskFactory.MockRemoteTask) RemoteTask(io.trino.execution.RemoteTask) Duration(io.airlift.units.Duration) PlanFragment(io.trino.sql.planner.PlanFragment) ConnectorAwareSplitSource(io.trino.split.ConnectorAwareSplitSource) InMemoryNodeManager(io.trino.metadata.InMemoryNodeManager) SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler(io.trino.execution.scheduler.SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler) TableExecuteContextManager(io.trino.execution.TableExecuteContextManager) InternalNode(io.trino.metadata.InternalNode) DynamicFilterService(io.trino.server.DynamicFilterService) TestingSplit(io.trino.testing.TestingSplit) DynamicFilterConfig(io.trino.execution.DynamicFilterConfig) Test(org.testng.annotations.Test)

Example 3 with TestingSplit

use of io.trino.testing.TestingSplit in project trino by trinodb.

the class TestSourcePartitionedScheduler method testNoNewTaskScheduledWhenChildStageBufferIsOverutilized.

@Test
public void testNoNewTaskScheduledWhenChildStageBufferIsOverutilized() {
    NodeTaskMap nodeTaskMap = new NodeTaskMap(finalizerService);
    // use private node manager so we can add a node later
    InMemoryNodeManager nodeManager = new InMemoryNodeManager();
    nodeManager.addNode(CONNECTOR_ID, new InternalNode("other1", URI.create("http://127.0.0.1:11"), NodeVersion.UNKNOWN, false), new InternalNode("other2", URI.create("http://127.0.0.1:12"), NodeVersion.UNKNOWN, false), new InternalNode("other3", URI.create("http://127.0.0.1:13"), NodeVersion.UNKNOWN, false));
    NodeScheduler nodeScheduler = new NodeScheduler(new UniformNodeSelectorFactory(nodeManager, new NodeSchedulerConfig().setIncludeCoordinator(false), nodeTaskMap, new Duration(0, SECONDS)));
    PlanFragment plan = createFragment();
    StageExecution stage = createStageExecution(plan, nodeTaskMap);
    // setting over utilized child output buffer
    StageScheduler scheduler = newSourcePartitionedSchedulerAsStageScheduler(stage, TABLE_SCAN_NODE_ID, new ConnectorAwareSplitSource(CONNECTOR_ID, createFixedSplitSource(400, TestingSplit::createRemoteSplit)), new DynamicSplitPlacementPolicy(nodeScheduler.createNodeSelector(session, Optional.of(CONNECTOR_ID)), stage::getAllTasks), 400, new DynamicFilterService(metadata, functionManager, typeOperators, new DynamicFilterConfig()), new TableExecuteContextManager(), () -> true);
    // the queues of 3 running nodes should be full
    ScheduleResult scheduleResult = scheduler.schedule();
    assertEquals(scheduleResult.getBlockedReason().get(), SPLIT_QUEUES_FULL);
    assertEquals(scheduleResult.getNewTasks().size(), 3);
    assertEquals(scheduleResult.getSplitsScheduled(), 300);
    for (RemoteTask remoteTask : scheduleResult.getNewTasks()) {
        PartitionedSplitsInfo splitsInfo = remoteTask.getPartitionedSplitsInfo();
        assertEquals(splitsInfo.getCount(), 100);
    }
    // new node added but 1 child's output buffer is overutilized - so lockdown the tasks
    nodeManager.addNode(CONNECTOR_ID, new InternalNode("other4", URI.create("http://127.0.0.4:14"), NodeVersion.UNKNOWN, false));
    scheduleResult = scheduler.schedule();
    assertEquals(scheduleResult.getBlockedReason().get(), SPLIT_QUEUES_FULL);
    assertEquals(scheduleResult.getNewTasks().size(), 0);
    assertEquals(scheduleResult.getSplitsScheduled(), 0);
}
Also used : NodeTaskMap(io.trino.execution.NodeTaskMap) PipelinedStageExecution.createPipelinedStageExecution(io.trino.execution.scheduler.PipelinedStageExecution.createPipelinedStageExecution) PartitionedSplitsInfo(io.trino.execution.PartitionedSplitsInfo) MockRemoteTask(io.trino.execution.MockRemoteTaskFactory.MockRemoteTask) RemoteTask(io.trino.execution.RemoteTask) Duration(io.airlift.units.Duration) PlanFragment(io.trino.sql.planner.PlanFragment) ConnectorAwareSplitSource(io.trino.split.ConnectorAwareSplitSource) InMemoryNodeManager(io.trino.metadata.InMemoryNodeManager) SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler(io.trino.execution.scheduler.SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler) TableExecuteContextManager(io.trino.execution.TableExecuteContextManager) InternalNode(io.trino.metadata.InternalNode) DynamicFilterService(io.trino.server.DynamicFilterService) TestingSplit(io.trino.testing.TestingSplit) DynamicFilterConfig(io.trino.execution.DynamicFilterConfig) Test(org.testng.annotations.Test)

Example 4 with TestingSplit

use of io.trino.testing.TestingSplit in project trino by trinodb.

the class TestSourcePartitionedScheduler method testScheduleSlowSplitSource.

@Test
public void testScheduleSlowSplitSource() {
    QueuedSplitSource queuedSplitSource = new QueuedSplitSource(TestingSplit::createRemoteSplit);
    PlanFragment plan = createFragment();
    NodeTaskMap nodeTaskMap = new NodeTaskMap(finalizerService);
    StageExecution stage = createStageExecution(plan, nodeTaskMap);
    StageScheduler scheduler = getSourcePartitionedScheduler(queuedSplitSource, stage, nodeManager, nodeTaskMap, 1, STAGE);
    // schedule with no splits - will block
    ScheduleResult scheduleResult = scheduler.schedule();
    assertFalse(scheduleResult.isFinished());
    assertFalse(scheduleResult.getBlocked().isDone());
    assertEquals(scheduleResult.getNewTasks().size(), 0);
    assertEquals(stage.getAllTasks().size(), 0);
    queuedSplitSource.addSplits(1);
    assertTrue(scheduleResult.getBlocked().isDone());
}
Also used : NodeTaskMap(io.trino.execution.NodeTaskMap) PipelinedStageExecution.createPipelinedStageExecution(io.trino.execution.scheduler.PipelinedStageExecution.createPipelinedStageExecution) TestingSplit(io.trino.testing.TestingSplit) PlanFragment(io.trino.sql.planner.PlanFragment) SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler(io.trino.execution.scheduler.SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler) Test(org.testng.annotations.Test)

Aggregations

NodeTaskMap (io.trino.execution.NodeTaskMap)4 PipelinedStageExecution.createPipelinedStageExecution (io.trino.execution.scheduler.PipelinedStageExecution.createPipelinedStageExecution)4 SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler (io.trino.execution.scheduler.SourcePartitionedScheduler.newSourcePartitionedSchedulerAsStageScheduler)4 PlanFragment (io.trino.sql.planner.PlanFragment)4 TestingSplit (io.trino.testing.TestingSplit)4 Test (org.testng.annotations.Test)4 MockRemoteTask (io.trino.execution.MockRemoteTaskFactory.MockRemoteTask)3 PartitionedSplitsInfo (io.trino.execution.PartitionedSplitsInfo)3 RemoteTask (io.trino.execution.RemoteTask)3 InMemoryNodeManager (io.trino.metadata.InMemoryNodeManager)3 InternalNode (io.trino.metadata.InternalNode)3 Duration (io.airlift.units.Duration)2 DynamicFilterConfig (io.trino.execution.DynamicFilterConfig)2 TableExecuteContextManager (io.trino.execution.TableExecuteContextManager)2 DynamicFilterService (io.trino.server.DynamicFilterService)2 ConnectorAwareSplitSource (io.trino.split.ConnectorAwareSplitSource)2