Status CachedPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) { // Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of // execution work that happens here, so this is needed for the time accounting to // make sense. ScopedTimer timer(&_commonStats.executionTimeMillis); // If we work this many times during the trial period, then we will replan the // query from scratch. size_t maxWorksBeforeReplan = static_cast<size_t>(internalQueryCacheEvictionRatio * _decisionWorks); // The trial period ends without replanning if the cached plan produces this many results. size_t numResults = MultiPlanStage::getTrialPeriodNumToReturn(*_canonicalQuery); for (size_t i = 0; i < maxWorksBeforeReplan; ++i) { // Might need to yield between calls to work due to the timer elapsing. Status yieldStatus = tryYield(yieldPolicy); if (!yieldStatus.isOK()) { return yieldStatus; } WorkingSetID id = WorkingSet::INVALID_ID; PlanStage::StageState state = child()->work(&id); if (PlanStage::ADVANCED == state) { // Save result for later. WorkingSetMember* member = _ws->get(id); // Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield. member->makeObjOwnedIfNeeded(); _results.push_back(id); if (_results.size() >= numResults) { // Once a plan returns enough results, stop working. Update cache with stats // from this run and return. updatePlanCache(); return Status::OK(); } } else if (PlanStage::IS_EOF == state) { // Cached plan hit EOF quickly enough. No need to replan. Update cache with stats // from this run and return. updatePlanCache(); return Status::OK(); } else if (PlanStage::NEED_YIELD == state) { if (id == WorkingSet::INVALID_ID) { if (!yieldPolicy->allowedToYield()) { throw WriteConflictException(); } } else { WorkingSetMember* member = _ws->get(id); invariant(member->hasFetcher()); // Transfer ownership of the fetcher and yield. _fetcher.reset(member->releaseFetcher()); } if (yieldPolicy->allowedToYield()) { yieldPolicy->forceYield(); } Status yieldStatus = tryYield(yieldPolicy); if (!yieldStatus.isOK()) { return yieldStatus; } } else if (PlanStage::FAILURE == state) { // On failure, fall back to replanning the whole query. We neither evict the // existing cache entry nor cache the result of replanning. BSONObj statusObj; WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj); LOG(1) << "Execution of cached plan failed, falling back to replan." << " query: " << _canonicalQuery->toStringShort() << " planSummary: " << Explain::getPlanSummary(child().get()) << " status: " << statusObj; const bool shouldCache = false; return replan(yieldPolicy, shouldCache); } else if (PlanStage::DEAD == state) { BSONObj statusObj; WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj); LOG(1) << "Execution of cached plan failed: PlanStage died" << ", query: " << _canonicalQuery->toStringShort() << " planSummary: " << Explain::getPlanSummary(child().get()) << " status: " << statusObj; return WorkingSetCommon::getMemberObjectStatus(statusObj); } else { invariant(PlanStage::NEED_TIME == state); } } // If we're here, the trial period took more than 'maxWorksBeforeReplan' work cycles. This // plan is taking too long, so we replan from scratch. LOG(1) << "Execution of cached plan required " << maxWorksBeforeReplan << " works, but was originally cached with only " << _decisionWorks << " works. Evicting cache entry and replanning query: " << _canonicalQuery->toStringShort() << " plan summary before replan: " << Explain::getPlanSummary(child().get()); const bool shouldCache = true; return replan(yieldPolicy, shouldCache); }
Status CachedPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) { // Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of // execution work that happens here, so this is needed for the time accounting to // make sense. ScopedTimer timer(getClock(), &_commonStats.executionTimeMillis); // During plan selection, the list of indices we are using to plan must remain stable, so the // query will die during yield recovery if any index has been dropped. However, once plan // selection completes successfully, we no longer need all indices to stick around. The selected // plan should safely die on yield recovery if it is using the dropped index. // // Dismiss the requirement that no indices can be dropped when this method returns. ON_BLOCK_EXIT([this] { releaseAllIndicesRequirement(); }); // If we work this many times during the trial period, then we will replan the // query from scratch. size_t maxWorksBeforeReplan = static_cast<size_t>(internalQueryCacheEvictionRatio * _decisionWorks); // The trial period ends without replanning if the cached plan produces this many results. size_t numResults = MultiPlanStage::getTrialPeriodNumToReturn(*_canonicalQuery); for (size_t i = 0; i < maxWorksBeforeReplan; ++i) { // Might need to yield between calls to work due to the timer elapsing. Status yieldStatus = tryYield(yieldPolicy); if (!yieldStatus.isOK()) { return yieldStatus; } WorkingSetID id = WorkingSet::INVALID_ID; PlanStage::StageState state = child()->work(&id); if (PlanStage::ADVANCED == state) { // Save result for later. WorkingSetMember* member = _ws->get(id); // Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield. member->makeObjOwnedIfNeeded(); _results.push(id); if (_results.size() >= numResults) { // Once a plan returns enough results, stop working. Update cache with stats // from this run and return. updatePlanCache(); return Status::OK(); } } else if (PlanStage::IS_EOF == state) { // Cached plan hit EOF quickly enough. No need to replan. Update cache with stats // from this run and return. updatePlanCache(); return Status::OK(); } else if (PlanStage::NEED_YIELD == state) { invariant(id == WorkingSet::INVALID_ID); if (!yieldPolicy->canAutoYield()) { throw WriteConflictException(); } if (yieldPolicy->canAutoYield()) { yieldPolicy->forceYield(); } Status yieldStatus = tryYield(yieldPolicy); if (!yieldStatus.isOK()) { return yieldStatus; } } else if (PlanStage::FAILURE == state) { // On failure, fall back to replanning the whole query. We neither evict the // existing cache entry nor cache the result of replanning. BSONObj statusObj; WorkingSetCommon::getStatusMemberObject(*_ws, id, &statusObj); LOG(1) << "Execution of cached plan failed, falling back to replan." << " query: " << redact(_canonicalQuery->toStringShort()) << " planSummary: " << Explain::getPlanSummary(child().get()) << " status: " << redact(statusObj); const bool shouldCache = false; return replan(yieldPolicy, shouldCache); } else { invariant(PlanStage::NEED_TIME == state); } } // If we're here, the trial period took more than 'maxWorksBeforeReplan' work cycles. This // plan is taking too long, so we replan from scratch. LOG(1) << "Execution of cached plan required " << maxWorksBeforeReplan << " works, but was originally cached with only " << _decisionWorks << " works. Evicting cache entry and replanning query: " << redact(_canonicalQuery->toStringShort()) << " plan summary before replan: " << Explain::getPlanSummary(child().get()); const bool shouldCache = true; return replan(yieldPolicy, shouldCache); }