PlanStage::StageState NearStage::advanceNext(WorkingSetID* toReturn) {
if (_resultBuffer.empty()) {
getNearStats()->intervalStats.push_back(*_nextIntervalStats);
_nextIntervalStats.reset();
_searchState = SearchState_Buffering;
return PlanStage::NEED_TIME;
}
*toReturn = _resultBuffer.top().resultID;
_resultBuffer.pop();
return PlanStage::ADVANCED;
}
GeoNear2DSphereStage::GeoNear2DSphereStage(const GeoNearParams& nearParams,
OperationContext* txn,
WorkingSet* workingSet,
Collection* collection,
IndexDescriptor* s2Index)
: NearStage(txn,
workingSet,
collection,
new PlanStageStats(CommonStats(kS2IndexNearStage.c_str()),
STAGE_GEO_NEAR_2DSPHERE)),
_nearParams(nearParams),
_s2Index(s2Index),
_fullBounds(geoNearDistanceBounds(nearParams.nearQuery)),
_currBounds(_fullBounds.center(), -1, _fullBounds.getInner()),
_boundsIncrement(twoDSphereBoundsIncrement(s2Index)) {
getNearStats()->keyPattern = s2Index->keyPattern();
}
GeoNear2DStage::GeoNear2DStage(const GeoNearParams& nearParams,
OperationContext* txn,
WorkingSet* workingSet,
Collection* collection,
IndexDescriptor* twoDIndex)
: NearStage(txn,
workingSet,
collection,
new PlanStageStats(CommonStats(kTwoDIndexNearStage.c_str()),
STAGE_GEO_NEAR_2D)),
_nearParams(nearParams),
_twoDIndex(twoDIndex),
_fullBounds(twoDDistanceBounds(nearParams, twoDIndex)),
_currBounds(_fullBounds.center(), -1, _fullBounds.getInner()),
_boundsIncrement(twoDBoundsIncrement(nearParams)) {
getNearStats()->keyPattern = twoDIndex->keyPattern();
}
StatusWith<NearStage::CoveredInterval*> //
GeoNear2DSphereStage::nextInterval(OperationContext* txn,
WorkingSet* workingSet,
Collection* collection) {
// The search is finished if we searched at least once and all the way to the edge
if (_currBounds.getInner() >= 0 && _currBounds.getOuter() == _fullBounds.getOuter()) {
return StatusWith<CoveredInterval*>(NULL);
}
//
// Setup the next interval
//
const NearStats* stats = getNearStats();
if (!stats->intervalStats.empty()) {
const IntervalStats& lastIntervalStats = stats->intervalStats.back();
// TODO: Generally we want small numbers of results fast, then larger numbers later
if (lastIntervalStats.numResultsBuffered < 300)
_boundsIncrement *= 2;
else if (lastIntervalStats.numResultsBuffered > 600)
_boundsIncrement /= 2;
}
R2Annulus nextBounds(_currBounds.center(),
_currBounds.getOuter(),
min(_currBounds.getOuter() + _boundsIncrement,
_fullBounds.getOuter()));
bool isLastInterval = (nextBounds.getOuter() == _fullBounds.getOuter());
_currBounds = nextBounds;
//
// Setup the covering region and stages for this interval
//
IndexScanParams scanParams;
scanParams.descriptor = _s2Index;
scanParams.direction = 1;
// We use a filter on the key. The filter rejects keys that don't intersect with the
// annulus. An object that is in the annulus might have a key that's not in it and a key
// that's in it. As such we can't just look at one key per object.
//
// This does force us to do our own deduping of results, though.
scanParams.doNotDedup = true;
scanParams.bounds = _nearParams.baseBounds;
// Because the planner doesn't yet set up 2D index bounds, do it ourselves here
const string s2Field = _nearParams.nearQuery.field;
const int s2FieldPosition = getFieldPosition(_s2Index, s2Field);
scanParams.bounds.fields[s2FieldPosition].intervals.clear();
OrderedIntervalList* coveredIntervals = &scanParams.bounds.fields[s2FieldPosition];
TwoDSphereKeyInRegionExpression* keyMatcher =
new TwoDSphereKeyInRegionExpression(_currBounds, s2Field);
ExpressionMapping::cover2dsphere(keyMatcher->getRegion(),
_s2Index->infoObj(),
coveredIntervals);
// IndexScan owns the hash matcher
IndexScan* scan = new IndexScanWithMatch(txn, scanParams, workingSet, keyMatcher);
// FetchStage owns index scan
FetchStage* fetcher(new FetchStage(workingSet, scan, _nearParams.filter, collection));
return StatusWith<CoveredInterval*>(new CoveredInterval(fetcher,
true,
nextBounds.getInner(),
nextBounds.getOuter(),
isLastInterval));
}
StatusWith<NearStage::CoveredInterval*> //
GeoNear2DStage::nextInterval(OperationContext* txn,
WorkingSet* workingSet,
Collection* collection) {
// The search is finished if we searched at least once and all the way to the edge
if (_currBounds.getInner() >= 0 && _currBounds.getOuter() == _fullBounds.getOuter()) {
return StatusWith<CoveredInterval*>(NULL);
}
//
// Setup the next interval
//
const NearStats* stats = getNearStats();
if (!stats->intervalStats.empty()) {
const IntervalStats& lastIntervalStats = stats->intervalStats.back();
// TODO: Generally we want small numbers of results fast, then larger numbers later
if (lastIntervalStats.numResultsBuffered < 300)
_boundsIncrement *= 2;
else if (lastIntervalStats.numResultsBuffered > 600)
_boundsIncrement /= 2;
}
_boundsIncrement = max(_boundsIncrement,
min2DBoundsIncrement(_nearParams.nearQuery, _twoDIndex));
R2Annulus nextBounds(_currBounds.center(),
_currBounds.getOuter(),
min(_currBounds.getOuter() + _boundsIncrement,
_fullBounds.getOuter()));
const bool isLastInterval = (nextBounds.getOuter() == _fullBounds.getOuter());
_currBounds = nextBounds;
//
// Get a covering region for this interval
//
const CRS queryCRS = _nearParams.nearQuery.centroid.crs;
auto_ptr<R2Region> coverRegion;
if (FLAT == queryCRS) {
// NOTE: Due to floating point math issues, FLAT searches of a 2D index need to treat
// containment and distance separately.
// Ex: (distance) 54.001 - 54 > 0.001, but (containment) 54 + 0.001 <= 54.001
// The idea is that a $near search with bounds is really a $within search, sorted by
// distance. We attach a custom $within : annulus matcher to do the $within search,
// and adjust max/min bounds slightly since, as above, containment does not mean the
// distance calculation won't slightly overflow the boundary.
//
// The code below adjusts:
// 1) Overall min/max bounds of the generated distance intervals to be more inclusive
// 2) Bounds of the interval covering to be more inclusive
// ... and later on we add the custom $within : annulus matcher.
//
// IMPORTANT: The *internal* interval distance bounds are *exact thresholds* - these
// should not be adjusted.
// TODO: Maybe integrate annuluses as a standard shape, and literally transform $near
// internally into a $within query with $near just as sort.
// Compute the maximum axis-aligned distance error
const double epsilon = std::numeric_limits<double>::epsilon()
* (max(abs(_fullBounds.center().x), abs(_fullBounds.center().y))
+ _fullBounds.getOuter());
if (nextBounds.getInner() > 0 && nextBounds.getInner() == _fullBounds.getInner()) {
nextBounds = R2Annulus(nextBounds.center(),
max(0.0, nextBounds.getInner() - epsilon),
nextBounds.getOuter());
}
if (nextBounds.getOuter() > 0 && nextBounds.getOuter() == _fullBounds.getOuter()) {
// We're at the max bound of the search, adjust interval maximum
nextBounds = R2Annulus(nextBounds.center(),
nextBounds.getInner(),
nextBounds.getOuter() + epsilon);
}
// *Always* adjust the covering bounds to be more inclusive
coverRegion.reset(new R2Annulus(nextBounds.center(),
max(0.0, nextBounds.getInner() - epsilon),
nextBounds.getOuter() + epsilon));
}
else {
invariant(SPHERE == queryCRS);
// TODO: As above, make this consistent with $within : $centerSphere
// Our intervals aren't in the same CRS as our index, so we need to adjust them
coverRegion.reset(new R2Annulus(projectBoundsToTwoDDegrees(nextBounds)));
}
//
// Setup the stages for this interval
//
IndexScanParams scanParams;
scanParams.descriptor = _twoDIndex;
scanParams.direction = 1;
// We use a filter on the key. The filter rejects keys that don't intersect with the
// annulus. An object that is in the annulus might have a key that's not in it and a key
// that's in it. As such we can't just look at one key per object.
//
// This does force us to do our own deduping of results, though.
scanParams.doNotDedup = true;
// Scan bounds on 2D indexes are only over the 2D field - other bounds aren't applicable.
// This is handled in query planning.
scanParams.bounds = _nearParams.baseBounds;
// The "2d" field is always the first in the index
const string twoDFieldName = _nearParams.nearQuery.field;
const int twoDFieldPosition = 0;
OrderedIntervalList coveredIntervals;
coveredIntervals.name = scanParams.bounds.fields[twoDFieldPosition].name;
ExpressionMapping::cover2d(*coverRegion,
_twoDIndex->infoObj(),
internalGeoNearQuery2DMaxCoveringCells,
&coveredIntervals);
// Intersect the $near bounds we just generated into the bounds we have for anything else
// in the scan (i.e. $within)
IndexBoundsBuilder::intersectize(coveredIntervals,
&scanParams.bounds.fields[twoDFieldPosition]);
// These parameters are stored by the index, and so must be ok
GeoHashConverter::Parameters hashParams;
GeoHashConverter::parseParameters(_twoDIndex->infoObj(), &hashParams);
MatchExpression* keyMatcher =
new TwoDKeyInRegionExpression(coverRegion.release(),
hashParams,
twoDFieldName);
// 2D indexes support covered search over additional fields they contain
// TODO: Don't need to clone, can just attach to custom matcher above
if (_nearParams.filter) {
AndMatchExpression* andMatcher = new AndMatchExpression();
andMatcher->add(keyMatcher);
andMatcher->add(_nearParams.filter->shallowClone());
keyMatcher = andMatcher;
}
// IndexScanWithMatch owns the matcher
IndexScan* scan = new IndexScanWithMatch(txn, scanParams, workingSet, keyMatcher);
MatchExpression* docMatcher = NULL;
// FLAT searches need to add an additional annulus $within matcher, see above
if (FLAT == queryCRS) {
docMatcher = new TwoDPtInAnnulusExpression(_fullBounds, twoDFieldName);
}
// FetchStage owns index scan
FetchStage* fetcher(new FetchStageWithMatch(workingSet,
scan,
docMatcher,
collection));
return StatusWith<CoveredInterval*>(new CoveredInterval(fetcher,
true,
nextBounds.getInner(),
nextBounds.getOuter(),
isLastInterval));
}
