shared_ptr<PlanExecutor> PipelineD::addCursorSource(const intrusive_ptr<Pipeline>& pipeline,
const intrusive_ptr<ExpressionContext>& expCtx,
shared_ptr<PlanExecutor> exec,
DepsTracker deps,
const BSONObj& queryObj,
const BSONObj& sortObj,
const BSONObj& projectionObj) {
// Get the full "namespace" name.
const string& fullName = expCtx->ns.ns();
// Put the PlanExecutor into a DocumentSourceCursor and add it to the front of the pipeline.
intrusive_ptr<DocumentSourceCursor> pSource =
DocumentSourceCursor::create(fullName, exec, expCtx);
// Note the query, sort, and projection for explain.
pSource->setQuery(queryObj);
pSource->setSort(sortObj);
if (deps.hasNoRequirements()) {
pSource->shouldProduceEmptyDocs();
}
if (!projectionObj.isEmpty()) {
pSource->setProjection(projectionObj, boost::none);
} else {
// There may be fewer dependencies now if the sort was covered.
if (!sortObj.isEmpty()) {
deps = pipeline->getDependencies(queryObj);
}
pSource->setProjection(deps.toProjection(), deps.toParsedDeps());
}
// Add the initial DocumentSourceCursor to the front of the pipeline. Then optimize again in
// case the new stage can be absorbed with the first stages of the pipeline.
pipeline->addInitialSource(pSource);
pipeline->optimizePipeline();
// DocumentSourceCursor expects a yielding PlanExecutor that has had its state saved. We
// deregister the PlanExecutor so that it can be registered with ClientCursor.
exec->deregisterExec();
exec->saveState();
return exec;
}
std::shared_ptr<PlanExecutor> PipelineD::prepareExecutor(
OperationContext* txn,
Collection* collection,
const NamespaceString& nss,
const intrusive_ptr<Pipeline>& pipeline,
const intrusive_ptr<ExpressionContext>& expCtx,
const intrusive_ptr<DocumentSourceSort>& sortStage,
const DepsTracker& deps,
const BSONObj& queryObj,
BSONObj* sortObj,
BSONObj* projectionObj) {
// The query system has the potential to use an index to provide a non-blocking sort and/or to
// use the projection to generate a covered plan. If this is possible, it is more efficient to
// let the query system handle those parts of the pipeline. If not, it is more efficient to use
// a $sort and/or a ParsedDeps object. Thus, we will determine whether the query system can
// provide a non-blocking sort or a covered projection before we commit to a PlanExecutor.
//
// To determine if the query system can provide a non-blocking sort, we pass the
// NO_BLOCKING_SORT planning option, meaning 'getExecutor' will not produce a PlanExecutor if
// the query system would use a blocking sort stage.
//
// To determine if the query system can provide a covered projection, we pass the
// NO_UNCOVERED_PROJECTS planning option, meaning 'getExecutor' will not produce a PlanExecutor
// if the query system would need to fetch the document to do the projection. The following
// logic uses the above strategies, with multiple calls to 'attemptToGetExecutor' to determine
// the most efficient way to handle the $sort and $project stages.
//
// LATER - We should attempt to determine if the results from the query are returned in some
// order so we can then apply other optimizations there are tickets for, such as SERVER-4507.
size_t plannerOpts = QueryPlannerParams::DEFAULT | QueryPlannerParams::NO_BLOCKING_SORT;
// If we are connecting directly to the shard rather than through a mongos, don't filter out
// orphaned documents.
if (ShardingState::get(txn)->needCollectionMetadata(txn, nss.ns())) {
plannerOpts |= QueryPlannerParams::INCLUDE_SHARD_FILTER;
}
if (deps.hasNoRequirements()) {
// If we don't need any fields from the input document, performing a count is faster, and
// will output empty documents, which is okay.
plannerOpts |= QueryPlannerParams::IS_COUNT;
}
// The only way to get a text score is to let the query system handle the projection. In all
// other cases, unless the query system can do an index-covered projection and avoid going to
// the raw record at all, it is faster to have ParsedDeps filter the fields we need.
if (!deps.needTextScore) {
plannerOpts |= QueryPlannerParams::NO_UNCOVERED_PROJECTIONS;
}
std::shared_ptr<PlanExecutor> exec;
BSONObj emptyProjection;
if (sortStage) {
// See if the query system can provide a non-blocking sort.
auto swExecutorSort = attemptToGetExecutor(
txn, collection, expCtx, queryObj, emptyProjection, *sortObj, plannerOpts);
if (swExecutorSort.isOK()) {
// Success! Now see if the query system can also cover the projection.
auto swExecutorSortAndProj = attemptToGetExecutor(
txn, collection, expCtx, queryObj, *projectionObj, *sortObj, plannerOpts);
if (swExecutorSortAndProj.isOK()) {
// Success! We have a non-blocking sort and a covered projection.
exec = std::move(swExecutorSortAndProj.getValue());
} else {
// The query system couldn't cover the projection.
*projectionObj = BSONObj();
exec = std::move(swExecutorSort.getValue());
}
// We know the sort is being handled by the query system, so remove the $sort stage.
pipeline->sources.pop_front();
if (sortStage->getLimitSrc()) {
// We need to reinsert the coalesced $limit after removing the $sort.
pipeline->sources.push_front(sortStage->getLimitSrc());
}
return exec;
}
// The query system can't provide a non-blocking sort.
*sortObj = BSONObj();
}
// Either there was no $sort stage, or the query system could not provide a non-blocking
// sort.
dassert(sortObj->isEmpty());
// See if the query system can cover the projection.
auto swExecutorProj = attemptToGetExecutor(
txn, collection, expCtx, queryObj, *projectionObj, *sortObj, plannerOpts);
if (swExecutorProj.isOK()) {
// Success! We have a covered projection.
return std::move(swExecutorProj.getValue());
}
// The query system couldn't provide a covered projection.
*projectionObj = BSONObj();
// If this doesn't work, nothing will.
return uassertStatusOK(attemptToGetExecutor(
txn, collection, expCtx, queryObj, *projectionObj, *sortObj, plannerOpts));
}