use of org.apache.drill.exec.proto.ExecProtos.FragmentHandle in project drill by apache.
the class SplittingParallelizer method generateWorkUnits.
/**
* Split plan into multiple plans based on parallelization
* Ideally it is applicable only to plans with two major fragments: Screen and UnionExchange
* But there could be cases where we can remove even multiple exchanges like in case of "order by"
* End goal is to get single major fragment: Screen with chain that ends up with a single minor fragment
* from Leaf Exchange. This way each plan can run independently without any exchange involvement
* @param options
* @param foremanNode - not really applicable
* @param queryId
* @param reader
* @param rootNode
* @param planningSet
* @param session
* @param queryContextInfo
* @return
* @throws ExecutionSetupException
*/
private List<QueryWorkUnit> generateWorkUnits(OptionList options, DrillbitEndpoint foremanNode, QueryId queryId, PhysicalPlanReader reader, Fragment rootNode, PlanningSet planningSet, UserSession session, QueryContextInformation queryContextInfo) throws ExecutionSetupException {
// now we generate all the individual plan fragments and associated assignments. Note, we need all endpoints
// assigned before we can materialize, so we start a new loop here rather than utilizing the previous one.
List<QueryWorkUnit> workUnits = Lists.newArrayList();
int plansCount = 0;
DrillbitEndpoint[] endPoints = null;
long initialAllocation = 0;
long maxAllocation = 0;
final Iterator<Wrapper> iter = planningSet.iterator();
while (iter.hasNext()) {
Wrapper wrapper = iter.next();
Fragment node = wrapper.getNode();
boolean isLeafFragment = node.getReceivingExchangePairs().size() == 0;
final PhysicalOperator physicalOperatorRoot = node.getRoot();
// get all the needed info from leaf fragment
if ((physicalOperatorRoot instanceof Exchange) && isLeafFragment) {
// need to get info about
// number of minor fragments
// assignedEndPoints
// allocation
plansCount = wrapper.getWidth();
initialAllocation = (wrapper.getInitialAllocation() != 0) ? wrapper.getInitialAllocation() / plansCount : 0;
maxAllocation = (wrapper.getMaxAllocation() != 0) ? wrapper.getMaxAllocation() / plansCount : 0;
endPoints = new DrillbitEndpoint[plansCount];
for (int mfId = 0; mfId < plansCount; mfId++) {
endPoints[mfId] = wrapper.getAssignedEndpoint(mfId);
}
}
}
if (plansCount == 0) {
// no exchange, return list of single QueryWorkUnit
workUnits.add(generateWorkUnit(options, foremanNode, queryId, reader, rootNode, planningSet, session, queryContextInfo));
return workUnits;
}
for (Wrapper wrapper : planningSet) {
Fragment node = wrapper.getNode();
final PhysicalOperator physicalOperatorRoot = node.getRoot();
if (physicalOperatorRoot instanceof Exchange) {
// get to 0 MajorFragment
continue;
}
boolean isRootNode = rootNode == node;
if (isRootNode && wrapper.getWidth() != 1) {
throw new ForemanSetupException(String.format("Failure while trying to setup fragment. " + "The root fragment must always have parallelization one. In the current case, the width was set to %d.", wrapper.getWidth()));
}
// this fragment is always leaf, as we are removing all the exchanges
boolean isLeafFragment = true;
FragmentHandle handle = //
FragmentHandle.newBuilder().setMajorFragmentId(//
wrapper.getMajorFragmentId()).setMinorFragmentId(// minor fragment ID is going to be always 0, as plan will be split
0).setQueryId(//
queryId).build();
// Create a minorFragment for each major fragment.
for (int minorFragmentId = 0; minorFragmentId < plansCount; minorFragmentId++) {
// those fragments should be empty
List<PlanFragment> fragments = Lists.newArrayList();
PlanFragment rootFragment = null;
FragmentRoot rootOperator = null;
IndexedFragmentNode iNode = new IndexedFragmentNode(minorFragmentId, wrapper);
wrapper.resetAllocation();
// two visitors here
// 1. To remove exchange
// 2. To reset operator IDs as exchanges were removed
PhysicalOperator op = physicalOperatorRoot.accept(ExchangeRemoverMaterializer.INSTANCE, iNode).accept(OperatorIdVisitor.INSTANCE, 0);
Preconditions.checkArgument(op instanceof FragmentRoot);
FragmentRoot root = (FragmentRoot) op;
// get plan as JSON
String plan;
String optionsData;
try {
plan = reader.writeJson(root);
optionsData = reader.writeJson(options);
} catch (JsonProcessingException e) {
throw new ForemanSetupException("Failure while trying to convert fragment into json.", e);
}
PlanFragment fragment = //
PlanFragment.newBuilder().setForeman(//
endPoints[minorFragmentId]).setFragmentJson(//
plan).setHandle(//
handle).setAssignment(//
endPoints[minorFragmentId]).setLeafFragment(//
isLeafFragment).setContext(queryContextInfo).setMemInitial(//
initialAllocation).setMemMax(// TODO - for some reason OOM is using leaf fragment max allocation divided by width
wrapper.getMaxAllocation()).setOptionsJson(optionsData).setCredentials(session.getCredentials()).addAllCollector(CountRequiredFragments.getCollectors(root)).build();
if (isRootNode) {
if (logger.isDebugEnabled()) {
logger.debug("Root fragment:\n {}", DrillStringUtils.unescapeJava(fragment.toString()));
}
rootFragment = fragment;
rootOperator = root;
} else {
if (logger.isDebugEnabled()) {
logger.debug("Remote fragment:\n {}", DrillStringUtils.unescapeJava(fragment.toString()));
}
throw new ForemanSetupException(String.format("There should not be non-root/remote fragment present in plan split, but there is:", DrillStringUtils.unescapeJava(fragment.toString())));
}
// fragments should be always empty here
workUnits.add(new QueryWorkUnit(rootOperator, rootFragment, fragments));
}
}
return workUnits;
}
use of org.apache.drill.exec.proto.ExecProtos.FragmentHandle in project drill by apache.
the class QueryManager method unpauseExecutingFragments.
/**
* Sends a resume signal to all fragments, regardless of their state, since the fragment might have paused before
* sending any message. Resume all fragments through the control tunnel.
*/
void unpauseExecutingFragments(final DrillbitContext drillbitContext) {
final Controller controller = drillbitContext.getController();
for (final FragmentData data : fragmentDataSet) {
final DrillbitEndpoint endpoint = data.getEndpoint();
final FragmentHandle handle = data.getHandle();
controller.getTunnel(endpoint).unpauseFragment(new SignalListener(endpoint, handle, SignalListener.Signal.UNPAUSE), handle);
}
}
use of org.apache.drill.exec.proto.ExecProtos.FragmentHandle in project drill by apache.
the class QueryManager method addFragment.
private void addFragment(final FragmentData fragmentData) {
final FragmentHandle fragmentHandle = fragmentData.getHandle();
final int majorFragmentId = fragmentHandle.getMajorFragmentId();
final int minorFragmentId = fragmentHandle.getMinorFragmentId();
IntObjectHashMap<FragmentData> minorMap = fragmentDataMap.get(majorFragmentId);
if (minorMap == null) {
minorMap = new IntObjectHashMap<>();
fragmentDataMap.put(majorFragmentId, minorMap);
}
minorMap.put(minorFragmentId, fragmentData);
fragmentDataSet.add(fragmentData);
}
use of org.apache.drill.exec.proto.ExecProtos.FragmentHandle in project drill by apache.
the class QueryManager method cancelExecutingFragments.
/**
* Stop all fragments with currently *known* active status (active as in SENDING, AWAITING_ALLOCATION, RUNNING).
*
* For the actual cancel calls for intermediate and leaf fragments, see
* {@link org.apache.drill.exec.work.batch.ControlMessageHandler#cancelFragment}
* (1) Root fragment: pending or running, send the cancel signal through a tunnel.
* (2) Intermediate fragment: pending or running, send the cancel signal through a tunnel (for local and remote
* fragments). The actual cancel is done by delegating the cancel to the work bus.
* (3) Leaf fragment: running, send the cancel signal through a tunnel. The cancel is done directly.
*/
void cancelExecutingFragments(final DrillbitContext drillbitContext) {
final Controller controller = drillbitContext.getController();
for (final FragmentData data : fragmentDataSet) {
switch(data.getState()) {
case SENDING:
case AWAITING_ALLOCATION:
case RUNNING:
final FragmentHandle handle = data.getHandle();
final DrillbitEndpoint endpoint = data.getEndpoint();
// TODO is the CancelListener redundant? Does the FragmentStatusListener get notified of the same?
controller.getTunnel(endpoint).cancelFragment(new SignalListener(endpoint, handle, SignalListener.Signal.CANCEL), handle);
break;
case FINISHED:
case CANCELLATION_REQUESTED:
case CANCELLED:
case FAILED:
// nothing to do
break;
}
}
}
use of org.apache.drill.exec.proto.ExecProtos.FragmentHandle in project drill by apache.
the class QueryManager method updateFragmentStatus.
private boolean updateFragmentStatus(final FragmentStatus fragmentStatus) {
final FragmentHandle fragmentHandle = fragmentStatus.getHandle();
final int majorFragmentId = fragmentHandle.getMajorFragmentId();
final int minorFragmentId = fragmentHandle.getMinorFragmentId();
final FragmentData data = fragmentDataMap.get(majorFragmentId).get(minorFragmentId);
final FragmentState oldState = data.getState();
final boolean inTerminalState = isTerminal(oldState);
final FragmentState currentState = fragmentStatus.getProfile().getState();
if (inTerminalState || (oldState == FragmentState.CANCELLATION_REQUESTED && !isTerminal(currentState))) {
// Already in a terminal state, or invalid state transition from CANCELLATION_REQUESTED. This shouldn't happen.
logger.warn(String.format("Received status message for fragment %s after fragment was in state %s. New state was %s", QueryIdHelper.getQueryIdentifier(fragmentHandle), oldState, currentState));
return false;
}
data.setStatus(fragmentStatus);
return oldState != currentState;
}
Aggregations