diff options
author | Paolo Polato <paolo.polato@mongodb.com> | 2021-05-20 10:25:06 +0000 |
---|---|---|
committer | Evergreen Agent <no-reply@evergreen.mongodb.com> | 2021-05-20 10:47:49 +0000 |
commit | 40e1693cb180fd17ec6fa0e1b32acc21769c0b85 (patch) | |
tree | 375a1c45870c16b2bbd92b6bd150d00268b0aeb9 | |
parent | 81bec8d5b1e03b2e31e38e202b65c3df62845a6d (diff) | |
download | mongo-40e1693cb180fd17ec6fa0e1b32acc21769c0b85.tar.gz |
SERVER-56307 Allow the donor to enter the critical section when the untransferred mods are within a convergence threshold.
4 files changed, 86 insertions, 7 deletions
diff --git a/src/mongo/db/s/migration_chunk_cloner_source_legacy.cpp b/src/mongo/db/s/migration_chunk_cloner_source_legacy.cpp index e02a0ee23e3..d65d7b6854d 100644 --- a/src/mongo/db/s/migration_chunk_cloner_source_legacy.cpp +++ b/src/mongo/db/s/migration_chunk_cloner_source_legacy.cpp @@ -41,11 +41,13 @@ #include "mongo/db/dbhelpers.h" #include "mongo/db/exec/plan_stage.h" #include "mongo/db/exec/working_set_common.h" +#include "mongo/db/index/index_access_method.h" #include "mongo/db/index/index_descriptor.h" #include "mongo/db/query/internal_plans.h" #include "mongo/db/repl/optime.h" #include "mongo/db/repl/replication_process.h" #include "mongo/db/s/start_chunk_clone_request.h" +#include "mongo/db/server_parameters.h" #include "mongo/db/service_context.h" #include "mongo/executor/remote_command_request.h" #include "mongo/executor/remote_command_response.h" @@ -63,6 +65,12 @@ namespace mongo { namespace { +/** + * The maximum percentage of untrasferred chunk mods at the end of a catch up iteration + * that may be deferred to the next phase of the migration protocol (where new writes get blocked). + */ +MONGO_EXPORT_SERVER_PARAMETER(maxCatchUpPercentageBeforeBlockingWrites, int, 10); + const char kRecvChunkStatus[] = "_recvChunkStatus"; const char kRecvChunkCommit[] = "_recvChunkCommit"; const char kRecvChunkAbort[] = "_recvChunkAbort"; @@ -296,6 +304,32 @@ Status MigrationChunkClonerSourceLegacy::awaitUntilCriticalSectionIsAppropriate( return Status::OK(); } + bool supportsCriticalSectionDuringCatchUp = false; + if (auto featureSupportedField = + res[StartChunkCloneRequest::kSupportsCriticalSectionDuringCatchUp]) { + if (!featureSupportedField.booleanSafe()) { + return {ErrorCodes::Error(563070), + str::stream() + << "Illegal value for " + << StartChunkCloneRequest::kSupportsCriticalSectionDuringCatchUp}; + } + supportsCriticalSectionDuringCatchUp = true; + } + + if (res["state"].String() == "catchup" && supportsCriticalSectionDuringCatchUp) { + int64_t estimatedUntransferredModsSize = _deleted.size() * _averageObjectIdSize + + _reload.size() * _averageObjectSizeForCloneLocs; + auto estimatedUntransferredChunkPercentage = + (std::min(_args.getMaxChunkSizeBytes(), estimatedUntransferredModsSize) * 100) / + _args.getMaxChunkSizeBytes(); + if (estimatedUntransferredChunkPercentage < + maxCatchUpPercentageBeforeBlockingWrites.load()) { + // The recipient is sufficiently caught-up with the writes on the donor. + // Block writes, so that it can drain everything. + return Status::OK(); + } + } + if (res["state"].String() == "fail") { return {ErrorCodes::OperationFailed, str::stream() << "Data transfer error: " << res["errmsg"].str()}; @@ -609,11 +643,11 @@ Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* opC // Allow multiKey based on the invariant that shard keys must be single-valued. Therefore, any // multi-key index prefixed by shard key cannot be multikey over the shard key fields. - IndexDescriptor* const idx = + IndexDescriptor* const shardKeyIdx = collection->getIndexCatalog()->findShardKeyPrefixedIndex(opCtx, _shardKeyPattern.toBSON(), false); // requireSingleKey - if (!idx) { + if (!shardKeyIdx) { return {ErrorCodes::IndexNotFound, str::stream() << "can't find index with prefix " << _shardKeyPattern.toBSON() << " in storeCurrentLocs for " @@ -634,7 +668,7 @@ Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* opC _deleteNotifyExec = std::move(statusWithDeleteNotificationPlanExecutor.getValue()); // Assume both min and max non-empty, append MinKey's to make them fit chosen index - const KeyPattern kp(idx->keyPattern()); + const KeyPattern kp(shardKeyIdx->keyPattern()); BSONObj min = Helpers::toKeyFormat(kp.extendRangeBound(_args.getMinKey(), false)); BSONObj max = Helpers::toKeyFormat(kp.extendRangeBound(_args.getMaxKey(), false)); @@ -643,7 +677,7 @@ Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* opC // being queued and will migrate in the 'transferMods' stage. auto exec = InternalPlanner::indexScan(opCtx, collection, - idx, + shardKeyIdx, min, max, BoundInclusion::kIncludeStartKeyOnly, @@ -698,6 +732,19 @@ Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* opC const uint64_t collectionAverageObjectSize = collection->averageObjectSize(opCtx); + uint64_t averageObjectIdSize = 0; + const uint64_t defaultObjectIdSize = OID::kOIDSize; + if (totalRecs > 0) { + const auto indexCatalog = collection->getIndexCatalog(); + const auto idIdx = indexCatalog->findIdIndex(opCtx); + if (!idIdx) { + return {ErrorCodes::IndexNotFound, + str::stream() << "can't find index '_id' in storeCurrentLocs for " + << _args.getNss().ns()}; + } + averageObjectIdSize = indexCatalog->getIndex(idIdx)->getSpaceUsedBytes(opCtx) / totalRecs; + } + if (isLargeChunk) { return { ErrorCodes::ChunkTooBig, @@ -719,8 +766,8 @@ Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* opC } stdx::lock_guard<stdx::mutex> lk(_mutex); - _averageObjectSizeForCloneLocs = collectionAverageObjectSize + 12; - + _averageObjectSizeForCloneLocs = collectionAverageObjectSize + defaultObjectIdSize; + _averageObjectIdSize = std::max(averageObjectIdSize, defaultObjectIdSize); return Status::OK(); } diff --git a/src/mongo/db/s/migration_chunk_cloner_source_legacy.h b/src/mongo/db/s/migration_chunk_cloner_source_legacy.h index 0e8c53feab1..aa9b5271119 100644 --- a/src/mongo/db/s/migration_chunk_cloner_source_legacy.h +++ b/src/mongo/db/s/migration_chunk_cloner_source_legacy.h @@ -252,6 +252,9 @@ private: // pre-allocation (initial clone). uint64_t _averageObjectSizeForCloneLocs{0}; + // The estimated average object _id size during the clone phase. + uint64_t _averageObjectIdSize{0}; + // List of _id of documents that were modified that must be re-cloned (xfer mods) std::list<BSONObj> _reload; diff --git a/src/mongo/db/s/migration_destination_manager.cpp b/src/mongo/db/s/migration_destination_manager.cpp index cf0a076e3f4..c2a19f9cf2b 100644 --- a/src/mongo/db/s/migration_destination_manager.cpp +++ b/src/mongo/db/s/migration_destination_manager.cpp @@ -69,6 +69,9 @@ #include "mongo/util/scopeguard.h" namespace mongo { + +constexpr StringData StartChunkCloneRequest::kSupportsCriticalSectionDuringCatchUp; + namespace { const auto getMigrationDestinationManager = @@ -295,6 +298,7 @@ void MigrationDestinationManager::report(BSONObjBuilder& b, b.append("min", _min); b.append("max", _max); b.append("shardKeyPattern", _shardKeyPattern); + b.append(StartChunkCloneRequest::kSupportsCriticalSectionDuringCatchUp, true); b.append("state", stateToString(_state)); @@ -462,6 +466,25 @@ Status MigrationDestinationManager::startCommit(const MigrationSessionId& sessio stdx::unique_lock<stdx::mutex> lock(_mutex); + const auto convergenceTimeout = + Shard::kDefaultConfigCommandTimeout + Shard::kDefaultConfigCommandTimeout / 4; + + // The donor may have started the commit while the recipient is still busy processing + // the last batch of mods sent in the catch up phase. Allow some time for synching up. + auto deadline = Date_t::now() + convergenceTimeout; + + while (_state == CATCHUP) { + if (stdx::cv_status::timeout == + _stateChangedCV.wait_until(lock, deadline.toSystemTimePoint())) { + return {ErrorCodes::CommandFailed, + str::stream() << "startCommit timed out waiting for the catch up completion. " + << "Sender's session is " + << sessionId.toString() + << ". Current session is " + << (_sessionId ? _sessionId->toString() : "none.")}; + } + } + if (_state != STEADY) { return {ErrorCodes::CommandFailed, str::stream() << "Migration startCommit attempted when not in STEADY state." @@ -489,7 +512,9 @@ Status MigrationDestinationManager::startCommit(const MigrationSessionId& sessio _state = COMMIT_START; _stateChangedCV.notify_all(); - auto const deadline = Date_t::now() + Seconds(30); + // Assigning a timeout slightly higher than the one used for network requests to the config + // server. Enough time to retry at least once in case of network failures (SERVER-51397). + deadline = Date_t::now() + convergenceTimeout; while (_sessionId) { if (stdx::cv_status::timeout == _isActiveCV.wait_until(lock, deadline.toSystemTimePoint())) { @@ -913,6 +938,7 @@ void MigrationDestinationManager::_migrateDriver(OperationContext* opCtx) { const auto& mods = res.response; if (mods["size"].number() == 0) { + // There are no more pending modifications to be applied. End the catchup phase break; } diff --git a/src/mongo/db/s/start_chunk_clone_request.h b/src/mongo/db/s/start_chunk_clone_request.h index c3d63fc807b..4a3e359372d 100644 --- a/src/mongo/db/s/start_chunk_clone_request.h +++ b/src/mongo/db/s/start_chunk_clone_request.h @@ -49,6 +49,9 @@ class StatusWith; */ class StartChunkCloneRequest { public: + static constexpr auto kSupportsCriticalSectionDuringCatchUp = + "supportsCriticalSectionDuringCatchUp"_sd; + /** * Parses the input command and produces a request corresponding to its arguments. */ |