void QueryManagerBase::markOperatorFinished(const dag_node_index index) {
query_exec_state_->setExecutionFinished(index);
for (const dag_node_index dependent_op_index : blocking_dependents_[index]) {
blocking_dependencies_[dependent_op_index].erase(index);
}
for (const dag_node_index dependent_op_index : output_consumers_[index]) {
non_blocking_dependencies_[dependent_op_index].erase(index);
}
RelationalOperator *op = query_dag_->getNodePayloadMutable(index);
op->updateCatalogOnCompletion();
const relation_id output_rel = op->getOutputRelationID();
for (const pair<dag_node_index, bool> &dependent_link : query_dag_->getDependents(index)) {
const dag_node_index dependent_op_index = dependent_link.first;
if (output_rel >= 0) {
// Signal dependent operator that current operator is done feeding input blocks.
query_dag_->getNodePayloadMutable(dependent_op_index)->doneFeedingInputBlocks(output_rel);
}
if (checkAllBlockingDependenciesMet(dependent_op_index)) {
// Process the dependent operator (of the operator whose WorkOrder
// was just executed) for which all the dependencies have been met.
if (!fetchNormalWorkOrders(dependent_op_index) &&
non_blocking_dependencies_[dependent_op_index].empty() &&
checkNormalExecutionOver(dependent_op_index) &&
(!checkRebuildRequired(dependent_op_index) || initiateRebuild(dependent_op_index))) {
markOperatorFinished(dependent_op_index);
}
}
}
}
/**
* @brief Check if the execution of the given operator is over.
*
* @param index The index of the given operator in the DAG.
*
* @return True if the execution of the given operator is over, false
* otherwise.
**/
inline bool checkOperatorExecutionOver(const dag_node_index index) const {
if (checkRebuildRequired(index)) {
return (checkNormalExecutionOver(index) && checkRebuildOver(index));
} else {
return checkNormalExecutionOver(index);
}
}
bool Foreman::initiateRebuild(dag_node_index index) {
DEBUG_ASSERT(!workorders_container_->hasRebuildWorkOrder(index));
DEBUG_ASSERT(checkRebuildRequired(index));
DEBUG_ASSERT(!checkRebuildInitiated(index));
getRebuildWorkOrders(index, workorders_container_.get());
rebuild_status_[index] = std::make_pair(
true, workorders_container_->getNumRebuildWorkOrders(index));
return (rebuild_status_[index].second == 0);
}
bool QueryManager::initiateRebuild(const dag_node_index index) {
DCHECK(!workorders_container_->hasRebuildWorkOrder(index));
DCHECK(checkRebuildRequired(index));
DCHECK(!checkRebuildInitiated(index));
getRebuildWorkOrders(index, workorders_container_.get());
query_exec_state_->setRebuildStatus(
index, workorders_container_->getNumRebuildWorkOrders(index), true);
return (query_exec_state_->getNumRebuildWorkOrders(index) == 0);
}
void Foreman::processOperator(RelationalOperator *op,
dag_node_index index,
bool recursively_check_dependents) {
if (fetchNormalWorkOrders(op, index)) {
// Fetched work orders. Return to wait for the generated work orders to
// execute, and skip the execution-finished checks.
return;
}
if (checkNormalExecutionOver(index)) {
if (checkRebuildRequired(index)) {
if (!checkRebuildInitiated(index)) {
// Rebuild hasn't started, initiate it.
if (initiateRebuild(index)) {
// Rebuild initiated and completed right away.
markOperatorFinished(index);
} else {
// Rebuild WorkOrders have been generated.
return;
}
} else if (checkRebuildOver(index)) {
// Rebuild had been initiated and it is over.
markOperatorFinished(index);
}
} else {
// Rebuild is not required and normal execution over, mark finished.
markOperatorFinished(index);
}
// If we reach here, that means the operator has been marked as finished.
if (recursively_check_dependents) {
for (pair<dag_node_index, bool> dependent_link :
query_dag_->getDependents(index)) {
if (checkAllBlockingDependenciesMet(dependent_link.first)) {
processOperator(
query_dag_->getNodePayloadMutable(dependent_link.first),
dependent_link.first, true);
}
}
}
}
}
void QueryManagerBase::processWorkOrderCompleteMessage(
const dag_node_index op_index,
const partition_id part_id) {
query_exec_state_->decrementNumQueuedWorkOrders(op_index);
if (!checkNormalExecutionOver(op_index)) {
// Normal execution under progress for this operator.
return;
}
if (checkRebuildRequired(op_index)) {
DCHECK(!checkRebuildInitiated(op_index));
if (!initiateRebuild(op_index)) {
// Rebuild under progress.
return;
}
// Rebuild initiated and completed right away.
}
markOperatorFinished(op_index);
}
void QueryManager::processWorkOrderCompleteMessage(
const dag_node_index op_index) {
query_exec_state_->decrementNumQueuedWorkOrders(op_index);
// Check if new work orders are available and fetch them if so.
fetchNormalWorkOrders(op_index);
if (checkRebuildRequired(op_index)) {
if (checkNormalExecutionOver(op_index)) {
if (!checkRebuildInitiated(op_index)) {
if (initiateRebuild(op_index)) {
// Rebuild initiated and completed right away.
markOperatorFinished(op_index);
} else {
// Rebuild under progress.
}
} else if (checkRebuildOver(op_index)) {
// Rebuild was under progress and now it is over.
markOperatorFinished(op_index);
}
} else {
// Normal execution under progress for this operator.
}
} else if (checkOperatorExecutionOver(op_index)) {
// Rebuild not required for this operator and its normal execution is
// complete.
markOperatorFinished(op_index);
}
for (const pair<dag_node_index, bool> &dependent_link :
query_dag_->getDependents(op_index)) {
const dag_node_index dependent_op_index = dependent_link.first;
if (checkAllBlockingDependenciesMet(dependent_op_index)) {
// Process the dependent operator (of the operator whose WorkOrder
// was just executed) for which all the dependencies have been met.
processOperator(dependent_op_index, true);
}
}
}
// TODO(harshad) - There is duplication in terms of functionality provided by
// TMB and ForemanMessage class with respect to determining the message types.
// Try to use TMB message types for infering the messsage types consistently.
bool Foreman::processMessage(const ForemanMessage &message) {
const dag_node_index dag_size = query_dag_->size();
// Get the relational operator that caused this message to be sent.
dag_node_index response_op_index = message.getRelationalOpIndex();
const int worker_id = message.getWorkerID();
switch (message.getType()) {
case ForemanMessage::kWorkOrderCompletion:
{
// Completion of a regular WorkOrder.
DEBUG_ASSERT(worker_id >= 0);
--queued_workorders_per_op_[response_op_index];
// As the given worker finished executing a WorkOrder, decrement its
// number of queued WorkOrders.
workers_->decrementNumQueuedWorkOrders(worker_id);
// Check if new work orders are available and fetch them if so.
fetchNormalWorkOrders(
query_dag_->getNodePayloadMutable(response_op_index),
response_op_index);
if (checkRebuildRequired(response_op_index)) {
if (checkNormalExecutionOver(response_op_index)) {
if (!checkRebuildInitiated(response_op_index)) {
if (initiateRebuild(response_op_index)) {
// Rebuild initiated and completed right away.
markOperatorFinished(response_op_index);
} else {
// Rebuild under progress.
}
} else if (checkRebuildOver(response_op_index)) {
// Rebuild was under progress and now it is over.
markOperatorFinished(response_op_index);
}
} else {
// Normal execution under progress for this operator.
}
} else if (checkOperatorExecutionOver(response_op_index)) {
// Rebuild not required for this operator and its normal execution is
// complete.
markOperatorFinished(response_op_index);
}
for (pair<dag_node_index, bool> dependent_link :
query_dag_->getDependents(response_op_index)) {
RelationalOperator *dependent_op =
query_dag_->getNodePayloadMutable(dependent_link.first);
if (checkAllBlockingDependenciesMet(dependent_link.first)) {
// Process the dependent operator (of the operator whose WorkOrder
// was just executed) for which all the dependencies have been met.
processOperator(dependent_op, dependent_link.first, true);
}
}
}
break;
case ForemanMessage::kRebuildCompletion:
{
DEBUG_ASSERT(worker_id >= 0);
// Completion of a rebuild WorkOrder.
--rebuild_status_[response_op_index].second;
workers_->decrementNumQueuedWorkOrders(worker_id);
if (checkRebuildOver(response_op_index)) {
markOperatorFinished(response_op_index);
for (pair<dag_node_index, bool> dependent_link :
query_dag_->getDependents(response_op_index)) {
RelationalOperator *dependent_op =
query_dag_->getNodePayloadMutable(dependent_link.first);
if (checkAllBlockingDependenciesMet(dependent_link.first)) {
processOperator(dependent_op, dependent_link.first, true);
}
}
}
}
break;
case ForemanMessage::kDataPipeline:
{
// Data streaming message. Possible senders of this message include
// InsertDestination and some operators which modify existing blocks.
for (dag_node_index consumer_index :
output_consumers_[response_op_index]) {
RelationalOperator *consumer_op =
query_dag_->getNodePayloadMutable(consumer_index);
// Feed the streamed block to the consumer. Note that
// output_consumers_ only contain those dependents of operator with
// index = response_op_index which are eligible to receive streamed
// input.
consumer_op->feedInputBlock(message.getOutputBlockID(),
message.getRelationID());
// Because of the streamed input just fed, check if there are any new
// WorkOrders available and if so, fetch them.
fetchNormalWorkOrders(consumer_op, consumer_index);
} // end for (feeding input to dependents)
}
break;
case ForemanMessage::kWorkOrdersAvailable: {
// Check if new work orders are available.
fetchNormalWorkOrders(
query_dag_->getNodePayloadMutable(response_op_index),
response_op_index);
break;
}
default:
FATAL_ERROR("Unknown ForemanMessage type");
}
// Dispatch the WorkerMessages to the workers. We prefer to start the search
// for the schedulable WorkOrders beginning from response_op_index. The first
// candidate worker to receive the next WorkOrder is the one that sent the
// response message to Foreman.
dispatchWorkerMessages(((worker_id >= 0) ? worker_id : 0), response_op_index);
return num_operators_finished_ == dag_size;
}
