Status getRunnerDistinct(Collection* collection,
const BSONObj& query,
const string& field,
Runner** out) {
// This should'a been checked by the distinct command.
verify(collection);
// TODO: check for idhack here?
// When can we do a fast distinct hack?
// 1. There is a plan with just one leaf and that leaf is an ixscan.
// 2. The ixscan indexes the field we're interested in.
// 2a: We are correct if the index contains the field but for now we look for prefix.
// 3. The query is covered/no fetch.
//
// We go through normal planning (with limited parameters) to see if we can produce
// a soln with the above properties.
QueryPlannerParams plannerParams;
plannerParams.options = QueryPlannerParams::NO_TABLE_SCAN;
IndexCatalog::IndexIterator ii = collection->getIndexCatalog()->getIndexIterator(false);
while (ii.more()) {
const IndexDescriptor* desc = ii.next();
// The distinct hack can work if any field is in the index but it's not always clear
// if it's a win unless it's the first field.
if (desc->keyPattern().firstElement().fieldName() == field) {
plannerParams.indices.push_back(IndexEntry(desc->keyPattern(),
desc->isMultikey(),
desc->isSparse(),
desc->indexName(),
desc->infoObj()));
}
}
// If there are no suitable indices for the distinct hack and the _id index is the only
// index in the catalog, a collection scan is the only possible solution. Bail out now
// into regular planning with no projection. Projection is unnecessary because there
// is no possibility of a covered index scan.
if (plannerParams.indices.empty() &&
collection->getIndexCatalog()->numIndexesTotal() == 1 &&
collection->getIndexCatalog()->haveIdIndex()) {
CanonicalQuery* cq;
Status status = CanonicalQuery::canonicalize(collection->ns().ns(), query, BSONObj(),
BSONObj(), &cq);
if (!status.isOK()) {
return status;
}
// Takes ownership of cq.
return getRunner(cq, out);
}
// We only care about the field that we're projecting over. Have to drop the _id field
// explicitly because those are .find() semantics.
//
// Applying a projection allows the planner to try to give us covered plans.
BSONObj projection;
if ("_id" == field) {
projection = BSON("_id" << 1);
}
else {
projection = BSON("_id" << 0 << field << 1);
}
// Apply a projection of the key. Empty BSONObj() is for the sort.
CanonicalQuery* cq;
Status status = CanonicalQuery::canonicalize(collection->ns().ns(), query, BSONObj(), projection, &cq);
if (!status.isOK()) {
return status;
}
// No index has the field we're looking for. Punt to normal planning.
if (plannerParams.indices.empty()) {
// Takes ownership of cq.
return getRunner(cq, out);
}
// If we're here, we have an index prefixed by the field we're distinct-ing over.
// If there's no query, we can just distinct-scan one of the indices.
// Not every index in plannerParams.indices may be suitable. Refer to
// getDistinctNodeIndex().
size_t distinctNodeIndex = 0;
if (query.isEmpty() &&
getDistinctNodeIndex(plannerParams.indices, &distinctNodeIndex)) {
DistinctNode* dn = new DistinctNode();
dn->indexKeyPattern = plannerParams.indices[distinctNodeIndex].keyPattern;
dn->direction = 1;
IndexBoundsBuilder::allValuesBounds(dn->indexKeyPattern, &dn->bounds);
dn->fieldNo = 0;
QueryPlannerParams params;
// Takes ownership of 'dn'.
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(*cq, params, dn);
verify(soln);
WorkingSet* ws;
PlanStage* root;
verify(StageBuilder::build(*soln, &root, &ws));
*out = new SingleSolutionRunner(collection, cq, soln, root, ws);
return Status::OK();
}
// See if we can answer the query in a fast-distinct compatible fashion.
vector<QuerySolution*> solutions;
status = QueryPlanner::plan(*cq, plannerParams, &solutions);
if (!status.isOK()) {
return getRunner(cq, out);
}
// XXX: why do we need to do this? planner should prob do this internally
cq->root()->resetTag();
// We look for a solution that has an ixscan we can turn into a distinctixscan
for (size_t i = 0; i < solutions.size(); ++i) {
if (turnIxscanIntoDistinctIxscan(solutions[i], field)) {
// Great, we can use solutions[i]. Clean up the other QuerySolution(s).
for (size_t j = 0; j < solutions.size(); ++j) {
if (j != i) {
delete solutions[j];
}
}
// Build and return the SSR over solutions[i].
WorkingSet* ws;
PlanStage* root;
verify(StageBuilder::build(*solutions[i], &root, &ws));
*out = new SingleSolutionRunner(collection, cq, solutions[i], root, ws);
return Status::OK();
}
}
// If we're here, the planner made a soln with the restricted index set but we couldn't
// translate any of them into a distinct-compatible soln. So, delete the solutions and just
// go through normal planning.
for (size_t i = 0; i < solutions.size(); ++i) {
delete solutions[i];
}
return getRunner(cq, out);
}
Status getExecutorDistinct(OperationContext* txn,
Collection* collection,
const BSONObj& query,
const std::string& field,
PlanExecutor** out) {
// This should'a been checked by the distinct command.
invariant(collection);
// TODO: check for idhack here?
// When can we do a fast distinct hack?
// 1. There is a plan with just one leaf and that leaf is an ixscan.
// 2. The ixscan indexes the field we're interested in.
// 2a: We are correct if the index contains the field but for now we look for prefix.
// 3. The query is covered/no fetch.
//
// We go through normal planning (with limited parameters) to see if we can produce
// a soln with the above properties.
QueryPlannerParams plannerParams;
plannerParams.options = QueryPlannerParams::NO_TABLE_SCAN;
IndexCatalog::IndexIterator ii = collection->getIndexCatalog()->getIndexIterator(false);
while (ii.more()) {
const IndexDescriptor* desc = ii.next();
// The distinct hack can work if any field is in the index but it's not always clear
// if it's a win unless it's the first field.
if (desc->keyPattern().firstElement().fieldName() == field) {
plannerParams.indices.push_back(IndexEntry(desc->keyPattern(),
desc->getAccessMethodName(),
desc->isMultikey(),
desc->isSparse(),
desc->indexName(),
desc->infoObj()));
}
}
const WhereCallbackReal whereCallback(collection->ns().db());
// If there are no suitable indices for the distinct hack bail out now into regular planning
// with no projection.
if (plannerParams.indices.empty()) {
CanonicalQuery* cq;
Status status = CanonicalQuery::canonicalize(
collection->ns().ns(), query, &cq, whereCallback);
if (!status.isOK()) {
return status;
}
// Takes ownership of 'cq'.
return getExecutor(txn, collection, cq, out);
}
//
// If we're here, we have an index prefixed by the field we're distinct-ing over.
//
// Applying a projection allows the planner to try to give us covered plans that we can turn
// into the projection hack. getDistinctProjection deals with .find() projection semantics
// (ie _id:1 being implied by default).
BSONObj projection = getDistinctProjection(field);
// Apply a projection of the key. Empty BSONObj() is for the sort.
CanonicalQuery* cq;
Status status = CanonicalQuery::canonicalize(collection->ns().ns(),
query,
BSONObj(),
projection,
&cq,
whereCallback);
if (!status.isOK()) {
return status;
}
auto_ptr<CanonicalQuery> autoCq(cq);
// If there's no query, we can just distinct-scan one of the indices.
// Not every index in plannerParams.indices may be suitable. Refer to
// getDistinctNodeIndex().
size_t distinctNodeIndex = 0;
if (query.isEmpty() &&
getDistinctNodeIndex(plannerParams.indices, field, &distinctNodeIndex)) {
DistinctNode* dn = new DistinctNode();
dn->indexKeyPattern = plannerParams.indices[distinctNodeIndex].keyPattern;
dn->direction = 1;
IndexBoundsBuilder::allValuesBounds(dn->indexKeyPattern, &dn->bounds);
dn->fieldNo = 0;
QueryPlannerParams params;
// Takes ownership of 'dn'.
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(*cq, params, dn);
invariant(soln);
LOG(2) << "Using fast distinct: " << cq->toStringShort()
<< ", planSummary: " << getPlanSummary(*soln);
WorkingSet* ws = new WorkingSet();
PlanStage* root;
verify(StageBuilder::build(txn, collection, *soln, ws, &root));
// Takes ownership of its arguments (except for 'collection').
*out = new PlanExecutor(ws, root, soln, autoCq.release(), collection);
return Status::OK();
}
// See if we can answer the query in a fast-distinct compatible fashion.
vector<QuerySolution*> solutions;
status = QueryPlanner::plan(*cq, plannerParams, &solutions);
if (!status.isOK()) {
return getExecutor(txn, collection, autoCq.release(), out);
}
// We look for a solution that has an ixscan we can turn into a distinctixscan
for (size_t i = 0; i < solutions.size(); ++i) {
if (turnIxscanIntoDistinctIxscan(solutions[i], field)) {
// Great, we can use solutions[i]. Clean up the other QuerySolution(s).
for (size_t j = 0; j < solutions.size(); ++j) {
if (j != i) {
delete solutions[j];
}
}
LOG(2) << "Using fast distinct: " << cq->toStringShort()
<< ", planSummary: " << getPlanSummary(*solutions[i]);
// Build and return the SSR over solutions[i].
WorkingSet* ws = new WorkingSet();
PlanStage* root;
verify(StageBuilder::build(txn, collection, *solutions[i], ws, &root));
// Takes ownership of its arguments (except for 'collection').
*out = new PlanExecutor(ws, root, solutions[i], autoCq.release(), collection);
return Status::OK();
}
}
// If we're here, the planner made a soln with the restricted index set but we couldn't
// translate any of them into a distinct-compatible soln. So, delete the solutions and just
// go through normal planning.
for (size_t i = 0; i < solutions.size(); ++i) {
delete solutions[i];
}
// We drop the projection from the 'cq'. Unfortunately this is not trivial.
status = CanonicalQuery::canonicalize(collection->ns().ns(), query, &cq, whereCallback);
if (!status.isOK()) {
return status;
}
autoCq.reset(cq);
// Takes ownership of 'autoCq'.
return getExecutor(txn, collection, autoCq.release(), out);
}
