boost::shared_ptr<Runner> PipelineD::prepareCursorSource(
Collection* collection,
const intrusive_ptr<Pipeline>& pPipeline,
const intrusive_ptr<ExpressionContext>& pExpCtx) {
// get the full "namespace" name
const string& fullName = pExpCtx->ns.ns();
pExpCtx->opCtx->lockState()->assertAtLeastReadLocked(fullName);
// We will be modifying the source vector as we go
Pipeline::SourceContainer& sources = pPipeline->sources;
// Inject a MongodImplementation to sources that need them.
for (size_t i = 0; i < sources.size(); i++) {
DocumentSourceNeedsMongod* needsMongod =
dynamic_cast<DocumentSourceNeedsMongod*>(sources[i].get());
if (needsMongod) {
needsMongod->injectMongodInterface(
boost::make_shared<MongodImplementation>(pExpCtx));
}
}
if (!sources.empty() && sources.front()->isValidInitialSource()) {
if (dynamic_cast<DocumentSourceMergeCursors*>(sources.front().get())) {
// Enable the hooks for setting up authentication on the subsequent internal
// connections we are going to create. This would normally have been done
// when SetShardVersion was called, but since SetShardVersion is never called
// on secondaries, this is needed.
ShardedConnectionInfo::addHook();
}
return boost::shared_ptr<Runner>(); // don't need a cursor
}
// Look for an initial match. This works whether we got an initial query or not.
// If not, it results in a "{}" query, which will be what we want in that case.
const BSONObj queryObj = pPipeline->getInitialQuery();
if (!queryObj.isEmpty()) {
// This will get built in to the Cursor we'll create, so
// remove the match from the pipeline
sources.pop_front();
}
// Find the set of fields in the source documents depended on by this pipeline.
const DepsTracker deps = pPipeline->getDependencies(queryObj);
// Passing query an empty projection since it is faster to use ParsedDeps::extractFields().
// This will need to change to support covering indexes (SERVER-12015). There is an
// exception for textScore since that can only be retrieved by a query projection.
const BSONObj projectionForQuery = deps.needTextScore ? deps.toProjection() : BSONObj();
/*
Look for an initial sort; we'll try to add this to the
Cursor we create. If we're successful in doing that (further down),
we'll remove the $sort from the pipeline, because the documents
will already come sorted in the specified order as a result of the
index scan.
*/
intrusive_ptr<DocumentSourceSort> sortStage;
BSONObj sortObj;
if (!sources.empty()) {
sortStage = dynamic_cast<DocumentSourceSort*>(sources.front().get());
if (sortStage) {
// build the sort key
sortObj = sortStage->serializeSortKey(/*explain*/false).toBson();
}
}
// Create the Runner.
//
// If we try to create a Runner that includes both the match and the
// sort, and the two are incompatible wrt the available indexes, then
// we don't get a Runner back.
//
// So we try to use both first. If that fails, try again, without the
// sort.
//
// If we don't have a sort, jump straight to just creating a Runner
// without the sort.
//
// If we are able to incorporate the sort into the Runner, remove it
// from the head of the pipeline.
//
// LATER - we should be able to find this out before we create the
// cursor. Either way, we can then apply other optimizations there
// are tickets for, such as SERVER-4507.
const size_t runnerOptions = QueryPlannerParams::DEFAULT
| QueryPlannerParams::INCLUDE_SHARD_FILTER
| QueryPlannerParams::NO_BLOCKING_SORT
;
boost::shared_ptr<Runner> runner;
bool sortInRunner = false;
const WhereCallbackReal whereCallback(pExpCtx->ns.db());
if (sortStage) {
CanonicalQuery* cq;
Status status =
CanonicalQuery::canonicalize(pExpCtx->ns,
queryObj,
sortObj,
projectionForQuery,
&cq,
whereCallback);
Runner* rawRunner;
if (status.isOK() && getRunner(collection, cq, &rawRunner, runnerOptions).isOK()) {
// success: The Runner will handle sorting for us using an index.
runner.reset(rawRunner);
sortInRunner = true;
sources.pop_front();
if (sortStage->getLimitSrc()) {
// need to reinsert coalesced $limit after removing $sort
sources.push_front(sortStage->getLimitSrc());
}
}
}
if (!runner.get()) {
const BSONObj noSort;
CanonicalQuery* cq;
uassertStatusOK(
CanonicalQuery::canonicalize(pExpCtx->ns,
queryObj,
noSort,
projectionForQuery,
&cq,
whereCallback));
Runner* rawRunner;
uassertStatusOK(getRunner(collection, cq, &rawRunner, runnerOptions));
runner.reset(rawRunner);
}
// DocumentSourceCursor expects a yielding Runner that has had its state saved.
runner->saveState();
// Put the Runner into a DocumentSourceCursor and add it to the front of the pipeline.
intrusive_ptr<DocumentSourceCursor> pSource =
DocumentSourceCursor::create(fullName, runner, pExpCtx);
// Note the query, sort, and projection for explain.
pSource->setQuery(queryObj);
if (sortInRunner)
pSource->setSort(sortObj);
pSource->setProjection(deps.toProjection(), deps.toParsedDeps());
while (!sources.empty() && pSource->coalesce(sources.front())) {
sources.pop_front();
}
pPipeline->addInitialSource(pSource);
return runner;
}
shared_ptr<PlanExecutor> PipelineD::prepareCursorSource(
OperationContext* txn,
Collection* collection,
const NamespaceString& nss,
const intrusive_ptr<Pipeline>& pPipeline,
const intrusive_ptr<ExpressionContext>& pExpCtx) {
// We will be modifying the source vector as we go.
Pipeline::SourceContainer& sources = pPipeline->sources;
// Inject a MongodImplementation to sources that need them.
for (auto&& source : sources) {
DocumentSourceNeedsMongod* needsMongod =
dynamic_cast<DocumentSourceNeedsMongod*>(source.get());
if (needsMongod) {
needsMongod->injectMongodInterface(std::make_shared<MongodImplementation>(pExpCtx));
}
}
if (!sources.empty()) {
if (sources.front()->isValidInitialSource()) {
if (dynamic_cast<DocumentSourceMergeCursors*>(sources.front().get())) {
// Enable the hooks for setting up authentication on the subsequent internal
// connections we are going to create. This would normally have been done
// when SetShardVersion was called, but since SetShardVersion is never called
// on secondaries, this is needed.
ShardedConnectionInfo::addHook();
}
return std::shared_ptr<PlanExecutor>(); // don't need a cursor
}
auto sampleStage = dynamic_cast<DocumentSourceSample*>(sources.front().get());
// Optimize an initial $sample stage if possible.
if (collection && sampleStage) {
const long long sampleSize = sampleStage->getSampleSize();
const long long numRecords = collection->getRecordStore()->numRecords(txn);
auto exec = createRandomCursorExecutor(collection, txn, sampleSize, numRecords);
if (exec) {
// Replace $sample stage with $sampleFromRandomCursor stage.
sources.pop_front();
std::string idString = collection->ns().isOplog() ? "ts" : "_id";
sources.emplace_front(DocumentSourceSampleFromRandomCursor::create(
pExpCtx, sampleSize, idString, numRecords));
const BSONObj initialQuery;
return addCursorSource(
pPipeline, pExpCtx, exec, pPipeline->getDependencies(initialQuery));
}
}
}
// Look for an initial match. This works whether we got an initial query or not. If not, it
// results in a "{}" query, which will be what we want in that case.
const BSONObj queryObj = pPipeline->getInitialQuery();
if (!queryObj.isEmpty()) {
if (dynamic_cast<DocumentSourceMatch*>(sources.front().get())) {
// If a $match query is pulled into the cursor, the $match is redundant, and can be
// removed from the pipeline.
sources.pop_front();
} else {
// A $geoNear stage, the only other stage that can produce an initial query, is also
// a valid initial stage and will be handled above.
MONGO_UNREACHABLE;
}
}
// Find the set of fields in the source documents depended on by this pipeline.
DepsTracker deps = pPipeline->getDependencies(queryObj);
BSONObj projForQuery = deps.toProjection();
/*
Look for an initial sort; we'll try to add this to the
Cursor we create. If we're successful in doing that (further down),
we'll remove the $sort from the pipeline, because the documents
will already come sorted in the specified order as a result of the
index scan.
*/
intrusive_ptr<DocumentSourceSort> sortStage;
BSONObj sortObj;
if (!sources.empty()) {
sortStage = dynamic_cast<DocumentSourceSort*>(sources.front().get());
if (sortStage) {
// build the sort key
sortObj = sortStage->serializeSortKey(/*explain*/ false).toBson();
}
}
// Create the PlanExecutor.
auto exec = prepareExecutor(txn,
collection,
nss,
pPipeline,
pExpCtx,
sortStage,
deps,
queryObj,
&sortObj,
&projForQuery);
return addCursorSource(pPipeline, pExpCtx, exec, deps, queryObj, sortObj, projForQuery);
}
