dremio fragment 执行简单说明
dremio 的内部执行实际上与drill 是比较类似的，只是dremio 做了不少的优化处理
一个调用流程说明
来自官方文档 
- 参考调用链
实际执行管理的类
AttemptManager
@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(queryIdString + ":foreman");
try {
injector.injectChecked(queryContext.getExecutionControls(), INJECTOR_TRY_BEGINNING_ERROR,
ForemanException.class);
observer.beginState(AttemptObserver.toEvent(AttemptEvent.State.PENDING));
observer.queryStarted(queryRequest, queryContext.getSession().getCredentials().getUserName());
String ruleSetEngine = ruleBasedEngineSelector.resolveAndUpdateEngine(queryContext);
ResourceSchedulingProperties resourceSchedulingProperties = new ResourceSchedulingProperties();
resourceSchedulingProperties.setRoutingEngine(queryContext.getSession().getRoutingEngine());
resourceSchedulingProperties.setRuleSetEngine(ruleSetEngine);
final GroupResourceInformation groupResourceInformation =
maestroService.getGroupResourceInformation(queryContext.getOptions(), resourceSchedulingProperties);
queryContext.setGroupResourceInformation(groupResourceInformation);
// Checks for Run Query privileges for the selected Engine
checkRunQueryAccessPrivilege(groupResourceInformation);
// planning is done in the command pool
commandPool.submit(CommandPool.Priority.LOW, attemptId.toString() + ":foreman-planning",
(waitInMillis) -> {
observer.commandPoolWait(waitInMillis);
injector.injectPause(queryContext.getExecutionControls(), INJECTOR_PENDING_PAUSE, logger);
injector.injectChecked(queryContext.getExecutionControls(), INJECTOR_PENDING_ERROR,
ForemanException.class);
plan();
injector.injectPause(queryContext.getExecutionControls(), INJECTOR_PLAN_PAUSE, logger);
injector.injectChecked(queryContext.getExecutionControls(), INJECTOR_PLAN_ERROR,
ForemanException.class);
return null;
}, runInSameThread).get();
if (command.getCommandType() == CommandType.ASYNC_QUERY) {
AsyncCommand asyncCommand = (AsyncCommand) command;
committer = asyncCommand.getPhysicalPlan().getCommitter();
queryCleaner = asyncCommand.getPhysicalPlan().getCleaner();
moveToState(QueryState.STARTING, null);
// 通过maestroService 执行查询
maestroService.executeQuery(queryId, queryContext, asyncCommand.getPhysicalPlan(), runInSameThread,
new MaestroObserverWrapper(observer), new CompletionListenerImpl());
asyncCommand.executionStarted();
}
observer.beginState(AttemptObserver.toEvent(AttemptEvent.State.RUNNING));
moveToState(QueryState.RUNNING, null);
injector.injectChecked(queryContext.getExecutionControls(), INJECTOR_TRY_END_ERROR,
ForemanException.class);
} catch (ResourceUnavailableException e) {
// resource allocation failure is treated as a cancellation and not a failure
try {
// the caller (JobEventCollatingObserver) expects metadata event before a cancel/complete event.
observer.planCompleted(null);
} catch (Exception ignore) {
}
profileTracker.setCancelReason(e.getMessage());
moveToState(QueryState.CANCELED, null); // ENQUEUED/STARTING -> CANCELED transition
} catch (final UserException | ForemanException e) {
moveToState(QueryState.FAILED, e);
} catch (final OutOfMemoryError e) {
if (ErrorHelper.isDirectMemoryException(e)) {
moveToState(QueryState.FAILED, UserException.memoryError(e).build(logger));
} else {
/*
* FragmentExecutors use a NodeStatusListener to watch out for the death of their query's AttemptManager. 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.
*/
ProcessExit.exitHeap(e);
}
} catch (Throwable ex) {
UserCancellationException t = ErrorHelper.findWrappedCause(ex, UserCancellationException.class);
if (t != null) {
moveToState(QueryState.CANCELED, null);
} else {
moveToState(QueryState.FAILED,
new ForemanException("Unexpected exception during fragment initialization: " + ex.getMessage(), ex));
}
} 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 AttemptManager
* would wait on the cancelling thread to signal a resume and the cancelling thread would wait on the AttemptManager
* to accept events.
*/
try {
stateSwitch.start();
} catch (Exception e) {
moveToState(QueryState.FAILED, e);
}
// restore the thread's original name
currentThread.setName(originalName);
}
/*
* Note that despite the run() completing, the AttemptManager could continue to exist, and receives
* events about fragment completions. It won't go away until everything is completed, failed, or cancelled.
*/
}
MaestroServiceImpl
@Override
public void executeQuery(
QueryId queryId,
QueryContext context,
PhysicalPlan physicalPlan,
boolean runInSameThread,
MaestroObserver observer,
CompletionListener listener) throws ExecutionSetupException, ResourceAllocationException {
injector.injectChecked(context.getExecutionControls(), INJECTOR_EXECUTE_QUERY_BEGIN_ERROR,
ExecutionSetupException.class);
// Set up the active query.
QueryTracker queryTracker = new QueryTrackerImpl(queryId, context, physicalPlan, reader,
resourceAllocator.get(), executorSetService.get(), executorSelectionService.get(),
executorServiceClientFactory.get(), jobTelemetryClient.get(), observer,
listener,
() -> closeQuery(queryId), closeableSchedulerThreadPool);
Preconditions.checkState(activeQueryMap.putIfAbsent(queryId, queryTracker) == null,
"query already queued for execution " + QueryIdHelper.getQueryId(queryId));
// allocate execution resources on the calling thread, as this will most likely block
queryTracker.allocateResources();
try {
observer.beginState(AttemptObserver.toEvent(AttemptEvent.State.EXECUTION_PLANNING));
// do execution planning in the bound pool
commandPool.get().submit(CommandPool.Priority.MEDIUM,
QueryIdHelper.getQueryId(queryId) + ":execution-planning",
(waitInMillis) -> {
injector.injectChecked(context.getExecutionControls(),
INJECTOR_COMMAND_POOL_SUBMIT_ERROR, ExecutionSetupException.class);
observer.commandPoolWait(waitInMillis);
queryTracker.planExecution();
return null;
}, runInSameThread).get();
} catch (ExecutionException|InterruptedException e) {
throw new ExecutionSetupException("failure during execution planning", e);
}
observer.beginState(AttemptObserver.toEvent(AttemptEvent.State.STARTING));
// propagate the fragments.
queryTracker.startFragments();
injector.injectChecked(context.getExecutionControls(), INJECTOR_EXECUTE_QUERY_END_ERROR,
ExecutionSetupException.class);
}
执行计划
QueryTrackerImpl
@Override
public void planExecution() throws ExecutionSetupException {
executionPlanningResources = ExecutionPlanCreator.getParallelizationInfo(context, observer,
physicalPlan, executorSelectionService, resourceTracker.getResourceSchedulingDecisionInfo());
injector.injectChecked(context.getExecutionControls(),
INJECTOR_EXECUTION_PLANNING_ERROR, ExecutionSetupException.class);
injector.injectPause(context.getExecutionControls(),
INJECTOR_EXECUTION_PLANNING_PAUSE, logger);
executionPlan = ExecutionPlanCreator.getExecutionPlan(context, reader, observer, physicalPlan,
resourceTracker.getResources(),
executionPlanningResources.getPlanningSet(), executorSelectionService,
resourceTracker.getResourceSchedulingDecisionInfo(),
executionPlanningResources.getGroupResourceInformation());
observer.planCompleted(executionPlan);
physicalPlan = null; // no longer needed
}
Fragment 执行
QueryTrackerImpl
@Override
public void startFragments() throws ExecutionSetupException {
Preconditions.checkNotNull(executionPlan, "execution plan required");
// Populate fragments before sending the query fragments.
fragmentTracker.populate(executionPlan.getFragments(), resourceTracker.getResourceSchedulingDecisionInfo());
AbstractMaestroObserver fragmentActivateObserver = new AbstractMaestroObserver() {
@Override
public void activateFragmentFailed(Exception ex) {
fragmentTracker.sendOrActivateFragmentsFailed(ex);
}
};
injector.injectPause(context.getExecutionControls(),
INJECTOR_STARTING_PAUSE, logger);
try {
// FragmentStarter 此处会通过rpc 调用ExecutorService，里边包含了对于FragmentExecutors 服务的调用
FragmentStarter starter = new FragmentStarter(executorServiceClientFactory,
resourceTracker.getResourceSchedulingDecisionInfo(),
context.getExecutionControls(),context.getOptions());
starter.start(executionPlan, MaestroObservers.of(observer, fragmentActivateObserver));
executionPlan = null; // no longer needed
progressTracker = new ProgressTracker(queryId, jobTelemetryClient, observer);
} catch (Exception ex) {
fragmentTracker.sendOrActivateFragmentsFailed(ex);
throw ex;
}
}
说明
以上是一个简单的执行说明，实际上只包含了具体的查询调用，关于fragment 的数据聚合部分没有说明,后边进行详细说明下
参考资料
sabot/kernel/src/main/java/com/dremio/exec/planner/sql/handlers/query/NormalHandler.java
sabot/kernel/src/main/java/com/dremio/exec/planner/sql/handlers/commands/HandlerToPreparePlan.java
sabot/kernel/src/main/java/com/dremio/exec/planner/sql/handlers/commands/HandlerToPrepareArrowPlan.java
sabot/kernel/src/main/java/com/dremio/exec/planner/sql/handlers/commands/HandlerToExec.java
sabot/kernel/src/main/java/com/dremio/exec/work/foreman/AttemptManager.java
sabot/kernel/src/main/java/com/dremio/exec/maestro/MaestroService.java
sabot/kernel/src/main/java/com/dremio/exec/planner/fragment/SimpleParallelizer.java
https://drill.apache.org/docs/drill-query-execution/
