void BatchWriteOp::buildBatchRequest(const TargetedWriteBatch& targetedBatch,
BatchedCommandRequest* request) const {
request->setNS(_clientRequest->getNS());
request->setShouldBypassValidation(_clientRequest->shouldBypassValidation());
const vector<TargetedWrite*>& targetedWrites = targetedBatch.getWrites();
for (vector<TargetedWrite*>::const_iterator it = targetedWrites.begin();
it != targetedWrites.end();
++it) {
const WriteOpRef& writeOpRef = (*it)->writeOpRef;
BatchedCommandRequest::BatchType batchType = _clientRequest->getBatchType();
// NOTE: We copy the batch items themselves here from the client request
// TODO: This could be inefficient, maybe we want to just reference in the future
if (batchType == BatchedCommandRequest::BatchType_Insert) {
BatchedInsertRequest* clientInsertRequest = _clientRequest->getInsertRequest();
BSONObj insertDoc = clientInsertRequest->getDocumentsAt(writeOpRef.first);
request->getInsertRequest()->addToDocuments(insertDoc);
} else if (batchType == BatchedCommandRequest::BatchType_Update) {
BatchedUpdateRequest* clientUpdateRequest = _clientRequest->getUpdateRequest();
BatchedUpdateDocument* updateDoc = new BatchedUpdateDocument;
clientUpdateRequest->getUpdatesAt(writeOpRef.first)->cloneTo(updateDoc);
request->getUpdateRequest()->addToUpdates(updateDoc);
} else {
dassert(batchType == BatchedCommandRequest::BatchType_Delete);
BatchedDeleteRequest* clientDeleteRequest = _clientRequest->getDeleteRequest();
BatchedDeleteDocument* deleteDoc = new BatchedDeleteDocument;
clientDeleteRequest->getDeletesAt(writeOpRef.first)->cloneTo(deleteDoc);
request->getDeleteRequest()->addToDeletes(deleteDoc);
}
// TODO: We can add logic here to allow aborting individual ops
// if ( NULL == response ) {
// ->responses.erase( it++ );
// continue;
//}
}
if (_clientRequest->isWriteConcernSet()) {
if (_clientRequest->isVerboseWC()) {
request->setWriteConcern(_clientRequest->getWriteConcern());
} else {
// Mongos needs to send to the shard with w > 0 so it will be able to
// see the writeErrors.
request->setWriteConcern(upgradeWriteConcern(_clientRequest->getWriteConcern()));
}
}
if (!request->isOrderedSet()) {
request->setOrdered(_clientRequest->getOrdered());
}
unique_ptr<BatchedRequestMetadata> requestMetadata(new BatchedRequestMetadata());
requestMetadata->setShardVersion(
ChunkVersionAndOpTime(targetedBatch.getEndpoint().shardVersion));
requestMetadata->setSession(0);
request->setMetadata(requestMetadata.release());
}
// Helper function to cancel all the write ops of targeted batches in a map
static void cancelBatches(const WriteErrorDetail& why,
WriteOp* writeOps,
TargetedBatchMap* batchMap) {
set<WriteOp*> targetedWriteOps;
// Collect all the writeOps that are currently targeted
for (TargetedBatchMap::iterator it = batchMap->begin(); it != batchMap->end();) {
TargetedWriteBatch* batch = it->second;
const vector<TargetedWrite*>& writes = batch->getWrites();
for (vector<TargetedWrite*>::const_iterator writeIt = writes.begin();
writeIt != writes.end();
++writeIt) {
TargetedWrite* write = *writeIt;
// NOTE: We may repeatedly cancel a write op here, but that's fast and we want to
// cancel before erasing the TargetedWrite* (which owns the cancelled targeting
// info) for reporting reasons.
writeOps[write->writeOpRef.first].cancelWrites(&why);
}
// Note that we need to *erase* first, *then* delete, since the map keys are ptrs from
// the values
batchMap->erase(it++);
delete batch;
}
batchMap->clear();
}
// Helper function to cancel all the write ops of targeted batches in a map
static void cancelBatches( const BatchedErrorDetail& why,
WriteOp* writeOps,
TargetedBatchMap* batchMap ) {
set<WriteOp*> targetedWriteOps;
// Collect all the writeOps that are currently targeted
for ( TargetedBatchMap::iterator it = batchMap->begin(); it != batchMap->end(); ) {
TargetedWriteBatch* batch = it->second;
const vector<TargetedWrite*>& writes = batch->getWrites();
for ( vector<TargetedWrite*>::const_iterator writeIt = writes.begin();
writeIt != writes.end(); ++writeIt ) {
TargetedWrite* write = *writeIt;
targetedWriteOps.insert( &writeOps[write->writeOpRef.first] );
}
// Note that we need to *erase* first, *then* delete, since the map keys are ptrs from
// the values
batchMap->erase( it++ );
delete batch;
}
batchMap->clear();
// Cancel all the write ops we found above
for ( set<WriteOp*>::iterator it = targetedWriteOps.begin(); it != targetedWriteOps.end();
++it ) {
WriteOp* writeOp = *it;
writeOp->cancelWrites( &why );
}
}
void BatchWriteOp::buildBatchRequest( const TargetedWriteBatch& targetedBatch,
BatchedCommandRequest* request ) const {
request->setNS( _clientRequest->getNS() );
request->setShardVersion( targetedBatch.getEndpoint().shardVersion );
const vector<TargetedWrite*>& targetedWrites = targetedBatch.getWrites();
for ( vector<TargetedWrite*>::const_iterator it = targetedWrites.begin();
it != targetedWrites.end(); ++it ) {
const WriteOpRef& writeOpRef = ( *it )->writeOpRef;
BatchedCommandRequest::BatchType batchType = _clientRequest->getBatchType();
// NOTE: We copy the batch items themselves here from the client request
// TODO: This could be inefficient, maybe we want to just reference in the future
if ( batchType == BatchedCommandRequest::BatchType_Insert ) {
BatchedInsertRequest* clientInsertRequest = _clientRequest->getInsertRequest();
BSONObj insertDoc = clientInsertRequest->getDocumentsAt( writeOpRef.first );
request->getInsertRequest()->addToDocuments( insertDoc );
}
else if ( batchType == BatchedCommandRequest::BatchType_Update ) {
BatchedUpdateRequest* clientUpdateRequest = _clientRequest->getUpdateRequest();
BatchedUpdateDocument* updateDoc = new BatchedUpdateDocument;
clientUpdateRequest->getUpdatesAt( writeOpRef.first )->cloneTo( updateDoc );
request->getUpdateRequest()->addToUpdates( updateDoc );
}
else {
dassert( batchType == BatchedCommandRequest::BatchType_Delete );
BatchedDeleteRequest* clientDeleteRequest = _clientRequest->getDeleteRequest();
BatchedDeleteDocument* deleteDoc = new BatchedDeleteDocument;
clientDeleteRequest->getDeletesAt( writeOpRef.first )->cloneTo( deleteDoc );
request->getDeleteRequest()->addToDeletes( deleteDoc );
}
// TODO: We can add logic here to allow aborting individual ops
//if ( NULL == response ) {
// ->responses.erase( it++ );
// continue;
//}
}
if ( _clientRequest->isWriteConcernSet() ) {
request->setWriteConcern( _clientRequest->getWriteConcern() );
}
if ( _clientRequest->isContinueOnErrorSet() ) {
request->setContinueOnError( _clientRequest->getContinueOnError() );
}
request->setSession( 0 );
}
void BatchWriteOp::noteBatchError(const TargetedWriteBatch& targetedBatch,
const WriteErrorDetail& error) {
// Treat errors to get a batch response as failures of the contained writes
BatchedCommandResponse emulatedResponse;
toWriteErrorResponse(
error, _clientRequest->getOrdered(), targetedBatch.getWrites().size(), &emulatedResponse);
noteBatchResponse(targetedBatch, emulatedResponse, NULL);
}
// Helper function to cancel all the write ops of targeted batch.
static void cancelBatch( const TargetedWriteBatch& targetedBatch,
WriteOp* writeOps,
const WriteErrorDetail& why ) {
const vector<TargetedWrite*>& writes = targetedBatch.getWrites();
for ( vector<TargetedWrite*>::const_iterator writeIt = writes.begin();
writeIt != writes.end(); ++writeIt ) {
TargetedWrite* write = *writeIt;
// NOTE: We may repeatedly cancel a write op here, but that's fast.
writeOps[write->writeOpRef.first].cancelWrites( &why );
}
}
void BatchWriteOp::noteBatchResponse(const TargetedWriteBatch& targetedBatch,
const BatchedCommandResponse& response,
TrackedErrors* trackedErrors) {
if (!response.getOk()) {
WriteErrorDetail error;
cloneCommandErrorTo(response, &error);
// Treat command errors exactly like other failures of the batch
// Note that no errors will be tracked from these failures - as-designed
noteBatchError(targetedBatch, error);
return;
}
dassert(response.getOk());
// Stop tracking targeted batch
_targeted.erase(&targetedBatch);
// Increment stats for this batch
incBatchStats(_clientRequest->getBatchType(), response, _stats.get());
//
// Assign errors to particular items.
// Write Concern errors are stored and handled later.
//
// Special handling for write concern errors, save for later
if (response.isWriteConcernErrorSet()) {
unique_ptr<ShardWCError> wcError(
new ShardWCError(targetedBatch.getEndpoint(), *response.getWriteConcernError()));
_wcErrors.mutableVector().push_back(wcError.release());
}
vector<WriteErrorDetail*> itemErrors;
// Handle batch and per-item errors
if (response.isErrDetailsSet()) {
// Per-item errors were set
itemErrors.insert(
itemErrors.begin(), response.getErrDetails().begin(), response.getErrDetails().end());
// Sort per-item errors by index
std::sort(itemErrors.begin(), itemErrors.end(), WriteErrorDetailComp());
}
//
// Go through all pending responses of the op and sorted remote reponses, populate errors
// This will either set all errors to the batch error or apply per-item errors as-needed
//
// If the batch is ordered, cancel all writes after the first error for retargeting.
//
bool ordered = _clientRequest->getOrdered();
vector<WriteErrorDetail*>::iterator itemErrorIt = itemErrors.begin();
int index = 0;
WriteErrorDetail* lastError = NULL;
for (vector<TargetedWrite*>::const_iterator it = targetedBatch.getWrites().begin();
it != targetedBatch.getWrites().end();
++it, ++index) {
const TargetedWrite* write = *it;
WriteOp& writeOp = _writeOps[write->writeOpRef.first];
dassert(writeOp.getWriteState() == WriteOpState_Pending);
// See if we have an error for the write
WriteErrorDetail* writeError = NULL;
if (itemErrorIt != itemErrors.end() && (*itemErrorIt)->getIndex() == index) {
// We have an per-item error for this write op's index
writeError = *itemErrorIt;
++itemErrorIt;
}
// Finish the response (with error, if needed)
if (NULL == writeError) {
if (!ordered || !lastError) {
writeOp.noteWriteComplete(*write);
} else {
// We didn't actually apply this write - cancel so we can retarget
dassert(writeOp.getNumTargeted() == 1u);
writeOp.cancelWrites(lastError);
}
} else {
writeOp.noteWriteError(*write, *writeError);
lastError = writeError;
}
}
// Track errors we care about, whether batch or individual errors
if (NULL != trackedErrors) {
trackErrors(targetedBatch.getEndpoint(), itemErrors, trackedErrors);
}
// Track upserted ids if we need to
if (response.isUpsertDetailsSet()) {
const vector<BatchedUpsertDetail*>& upsertedIds = response.getUpsertDetails();
for (vector<BatchedUpsertDetail*>::const_iterator it = upsertedIds.begin();
it != upsertedIds.end();
++it) {
// The child upserted details don't have the correct index for the full batch
const BatchedUpsertDetail* childUpsertedId = *it;
// Work backward from the child batch item index to the batch item index
int childBatchIndex = childUpsertedId->getIndex();
int batchIndex = targetedBatch.getWrites()[childBatchIndex]->writeOpRef.first;
// Push the upserted id with the correct index into the batch upserted ids
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex(batchIndex);
upsertedId->setUpsertedID(childUpsertedId->getUpsertedID());
_upsertedIds.mutableVector().push_back(upsertedId);
}
}
}
Status BatchWriteOp::targetBatch(OperationContext* txn,
const NSTargeter& targeter,
bool recordTargetErrors,
vector<TargetedWriteBatch*>* targetedBatches) {
//
// Targeting of unordered batches is fairly simple - each remaining write op is targeted,
// and each of those targeted writes are grouped into a batch for a particular shard
// endpoint.
//
// Targeting of ordered batches is a bit more complex - to respect the ordering of the
// batch, we can only send:
// A) a single targeted batch to one shard endpoint
// B) multiple targeted batches, but only containing targeted writes for a single write op
//
// This means that any multi-shard write operation must be targeted and sent one-by-one.
// Subsequent single-shard write operations can be batched together if they go to the same
// place.
//
// Ex: ShardA : { skey : a->k }, ShardB : { skey : k->z }
//
// Ordered insert batch of: [{ skey : a }, { skey : b }, { skey : x }]
// broken into:
// [{ skey : a }, { skey : b }],
// [{ skey : x }]
//
// Ordered update Batch of :
// [{ skey : a }{ $push },
// { skey : b }{ $push },
// { skey : [c, x] }{ $push },
// { skey : y }{ $push },
// { skey : z }{ $push }]
// broken into:
// [{ skey : a }, { skey : b }],
// [{ skey : [c,x] }],
// [{ skey : y }, { skey : z }]
//
const bool ordered = _clientRequest->getOrdered();
TargetedBatchMap batchMap;
TargetedBatchSizeMap batchSizes;
int numTargetErrors = 0;
size_t numWriteOps = _clientRequest->sizeWriteOps();
for (size_t i = 0; i < numWriteOps; ++i) {
WriteOp& writeOp = _writeOps[i];
// Only target _Ready ops
if (writeOp.getWriteState() != WriteOpState_Ready)
continue;
//
// Get TargetedWrites from the targeter for the write operation
//
// TargetedWrites need to be owned once returned
OwnedPointerVector<TargetedWrite> writesOwned;
vector<TargetedWrite*>& writes = writesOwned.mutableVector();
Status targetStatus = writeOp.targetWrites(txn, targeter, &writes);
if (!targetStatus.isOK()) {
WriteErrorDetail targetError;
buildTargetError(targetStatus, &targetError);
if (!recordTargetErrors) {
// Cancel current batch state with an error
cancelBatches(targetError, _writeOps, &batchMap);
dassert(batchMap.empty());
return targetStatus;
} else if (!ordered || batchMap.empty()) {
// Record an error for this batch
writeOp.setOpError(targetError);
++numTargetErrors;
if (ordered)
return Status::OK();
continue;
} else {
dassert(ordered && !batchMap.empty());
// Send out what we have, but don't record an error yet, since there may be an
// error in the writes before this point.
writeOp.cancelWrites(&targetError);
break;
}
}
//
// If ordered and we have a previous endpoint, make sure we don't need to send these
// targeted writes to any other endpoints.
//
if (ordered && !batchMap.empty()) {
dassert(batchMap.size() == 1u);
if (isNewBatchRequired(writes, batchMap)) {
writeOp.cancelWrites(NULL);
break;
}
}
//
// If this write will push us over some sort of size limit, stop targeting
//
int writeSizeBytes = getWriteSizeBytes(writeOp);
if (wouldMakeBatchesTooBig(writes, writeSizeBytes, batchSizes)) {
invariant(!batchMap.empty());
writeOp.cancelWrites(NULL);
break;
}
//
// Targeting went ok, add to appropriate TargetedBatch
//
for (vector<TargetedWrite*>::iterator it = writes.begin(); it != writes.end(); ++it) {
TargetedWrite* write = *it;
TargetedBatchMap::iterator batchIt = batchMap.find(&write->endpoint);
TargetedBatchSizeMap::iterator batchSizeIt = batchSizes.find(&write->endpoint);
if (batchIt == batchMap.end()) {
TargetedWriteBatch* newBatch = new TargetedWriteBatch(write->endpoint);
batchIt = batchMap.insert(make_pair(&newBatch->getEndpoint(), newBatch)).first;
batchSizeIt =
batchSizes.insert(make_pair(&newBatch->getEndpoint(), BatchSize())).first;
}
TargetedWriteBatch* batch = batchIt->second;
BatchSize& batchSize = batchSizeIt->second;
++batchSize.numOps;
batchSize.sizeBytes += writeSizeBytes;
batch->addWrite(write);
}
// Relinquish ownership of TargetedWrites, now the TargetedBatches own them
writesOwned.mutableVector().clear();
//
// Break if we're ordered and we have more than one endpoint - later writes cannot be
// enforced as ordered across multiple shard endpoints.
//
if (ordered && batchMap.size() > 1u)
break;
}
//
// Send back our targeted batches
//
for (TargetedBatchMap::iterator it = batchMap.begin(); it != batchMap.end(); ++it) {
TargetedWriteBatch* batch = it->second;
if (batch->getWrites().empty())
continue;
// Remember targeted batch for reporting
_targeted.insert(batch);
// Send the handle back to caller
targetedBatches->push_back(batch);
}
return Status::OK();
}
void BatchWriteExec::executeBatch( const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse ) {
BatchWriteOp batchOp;
batchOp.initClientRequest( &clientRequest );
// Current batch status
bool refreshedTargeter = false;
int rounds = 0;
int numCompletedOps = 0;
int numRoundsWithoutProgress = 0;
while ( !batchOp.isFinished() ) {
//
// Get child batches to send using the targeter
//
// Targeting errors can be caused by remote metadata changing (the collection could have
// been dropped and recreated, for example with a new shard key). If a remote metadata
// change occurs *before* a client sends us a batch, we need to make sure that we don't
// error out just because we're staler than the client - otherwise mongos will be have
// unpredictable behavior.
//
// (If a metadata change happens *during* or *after* a client sends us a batch, however,
// we make no guarantees about delivery.)
//
// For this reason, we don't record targeting errors until we've refreshed our targeting
// metadata at least once *after* receiving the client batch - at that point, we know:
//
// 1) our new metadata is the same as the metadata when the client sent a batch, and so
// targeting errors are real.
// OR
// 2) our new metadata is a newer version than when the client sent a batch, and so
// the metadata must have changed after the client batch was sent. We don't need to
// deliver in this case, since for all the client knows we may have gotten the batch
// exactly when the metadata changed.
//
vector<TargetedWriteBatch*> childBatches;
// If we've already had a targeting error, we've refreshed the metadata once and can
// record target errors definitively.
bool recordTargetErrors = refreshedTargeter;
Status targetStatus = batchOp.targetBatch( *_targeter,
recordTargetErrors,
&childBatches );
if ( !targetStatus.isOK() ) {
// Don't do anything until a targeter refresh
_targeter->noteCouldNotTarget();
refreshedTargeter = true;
++_stats->numTargetErrors;
dassert( childBatches.size() == 0u );
}
//
// Send all child batches
//
size_t numSent = 0;
size_t numToSend = childBatches.size();
bool remoteMetadataChanging = false;
while ( numSent != numToSend ) {
// Collect batches out on the network, mapped by endpoint
HostBatchMap pendingBatches;
//
// Send side
//
// Get as many batches as we can at once
for ( vector<TargetedWriteBatch*>::iterator it = childBatches.begin();
it != childBatches.end(); ++it ) {
//
// Collect the info needed to dispatch our targeted batch
//
TargetedWriteBatch* nextBatch = *it;
// If the batch is NULL, we sent it previously, so skip
if ( nextBatch == NULL ) continue;
// Figure out what host we need to dispatch our targeted batch
ConnectionString shardHost;
Status resolveStatus = _resolver->chooseWriteHost( nextBatch->getEndpoint()
.shardName,
&shardHost );
if ( !resolveStatus.isOK() ) {
++_stats->numResolveErrors;
// Record a resolve failure
// TODO: It may be necessary to refresh the cache if stale, or maybe just
// cancel and retarget the batch
WriteErrorDetail error;
buildErrorFrom( resolveStatus, &error );
batchOp.noteBatchError( *nextBatch, error );
// We're done with this batch
*it = NULL;
--numToSend;
continue;
}
// If we already have a batch for this host, wait until the next time
HostBatchMap::iterator pendingIt = pendingBatches.find( shardHost );
if ( pendingIt != pendingBatches.end() ) continue;
//
// We now have all the info needed to dispatch the batch
//
BatchedCommandRequest request( clientRequest.getBatchType() );
batchOp.buildBatchRequest( *nextBatch, &request );
// Internally we use full namespaces for request/response, but we send the
// command to a database with the collection name in the request.
NamespaceString nss( request.getNS() );
request.setNS( nss.coll() );
_dispatcher->addCommand( shardHost, nss.db(), request );
// Indicate we're done by setting the batch to NULL
// We'll only get duplicate hostEndpoints if we have broadcast and non-broadcast
// endpoints for the same host, so this should be pretty efficient without
// moving stuff around.
*it = NULL;
// Recv-side is responsible for cleaning up the nextBatch when used
pendingBatches.insert( make_pair( shardHost, nextBatch ) );
}
// Send them all out
_dispatcher->sendAll();
numSent += pendingBatches.size();
//
// Recv side
//
while ( _dispatcher->numPending() > 0 ) {
// Get the response
ConnectionString shardHost;
BatchedCommandResponse response;
Status dispatchStatus = _dispatcher->recvAny( &shardHost, &response );
// Get the TargetedWriteBatch to find where to put the response
dassert( pendingBatches.find( shardHost ) != pendingBatches.end() );
TargetedWriteBatch* batchRaw = pendingBatches.find( shardHost )->second;
scoped_ptr<TargetedWriteBatch> batch( batchRaw );
if ( dispatchStatus.isOK() ) {
TrackedErrors trackedErrors;
trackedErrors.startTracking( ErrorCodes::StaleShardVersion );
// Dispatch was ok, note response
batchOp.noteBatchResponse( *batch, response, &trackedErrors );
// Note if anything was stale
const vector<ShardError*>& staleErrors =
trackedErrors.getErrors( ErrorCodes::StaleShardVersion );
if ( staleErrors.size() > 0 ) {
noteStaleResponses( staleErrors, _targeter );
++_stats->numStaleBatches;
}
// Remember if the shard is actively changing metadata right now
if ( isShardMetadataChanging( staleErrors ) ) {
remoteMetadataChanging = true;
}
// Remember that we successfully wrote to this shard
// NOTE: This will record lastOps for shards where we actually didn't update
// or delete any documents, which preserves old behavior but is conservative
_stats->noteWriteAt( shardHost,
response.isLastOpSet() ?
response.getLastOp() : OpTime(),
response.isElectionIdSet() ?
response.getElectionId() : OID());
}
else {
// Error occurred dispatching, note it
WriteErrorDetail error;
buildErrorFrom( dispatchStatus, &error );
batchOp.noteBatchError( *batch, error );
}
}
}
++rounds;
++_stats->numRounds;
// If we're done, get out
if ( batchOp.isFinished() )
break;
// MORE WORK TO DO
//
// Refresh the targeter if we need to (no-op if nothing stale)
//
bool targeterChanged = false;
Status refreshStatus = _targeter->refreshIfNeeded( &targeterChanged );
if ( !refreshStatus.isOK() ) {
// It's okay if we can't refresh, we'll just record errors for the ops if
// needed.
warning() << "could not refresh targeter" << causedBy( refreshStatus.reason() )
<< endl;
}
//
// Ensure progress is being made toward completing the batch op
//
int currCompletedOps = batchOp.numWriteOpsIn( WriteOpState_Completed );
if ( currCompletedOps == numCompletedOps && !targeterChanged
&& !remoteMetadataChanging ) {
++numRoundsWithoutProgress;
}
else {
numRoundsWithoutProgress = 0;
}
numCompletedOps = currCompletedOps;
if ( numRoundsWithoutProgress > kMaxRoundsWithoutProgress ) {
stringstream msg;
msg << "no progress was made executing batch write op in " << clientRequest.getNS()
<< " after " << kMaxRoundsWithoutProgress << " rounds (" << numCompletedOps
<< " ops completed in " << rounds << " rounds total)";
WriteErrorDetail error;
buildErrorFrom( Status( ErrorCodes::NoProgressMade, msg.str() ), &error );
batchOp.setBatchError( error );
break;
}
}
batchOp.buildClientResponse( clientResponse );
}
void BatchWriteExec::executeBatch( const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse ) {
BatchWriteOp batchOp;
batchOp.initClientRequest( &clientRequest );
int numTargetErrors = 0;
int numStaleBatches = 0;
for ( int rounds = 0; !batchOp.isFinished(); rounds++ ) {
//
// Refresh the targeter if we need to (no-op if nothing stale)
//
Status refreshStatus = _targeter->refreshIfNeeded();
if ( !refreshStatus.isOK() ) {
// It's okay if we can't refresh, we'll just record errors for the ops if
// needed.
warning() << "could not refresh targeter" << causedBy( refreshStatus.reason() )
<< endl;
}
//
// Get child batches to send
//
vector<TargetedWriteBatch*> childBatches;
//
// Targeting errors can be caused by remote metadata changing (the collection could have
// been dropped and recreated, for example with a new shard key). If a remote metadata
// change occurs *before* a client sends us a batch, we need to make sure that we don't
// error out just because we're staler than the client - otherwise mongos will be have
// unpredictable behavior.
//
// (If a metadata change happens *during* or *after* a client sends us a batch, however,
// we make no guarantees about delivery.)
//
// For this reason, we don't record targeting errors until we've refreshed our targeting
// metadata at least once *after* receiving the client batch - at that point, we know:
//
// 1) our new metadata is the same as the metadata when the client sent a batch, and so
// targeting errors are real.
// OR
// 2) our new metadata is a newer version than when the client sent a batch, and so
// the metadata must have changed after the client batch was sent. We don't need to
// deliver in this case, since for all the client knows we may have gotten the batch
// exactly when the metadata changed.
//
// If we've had a targeting error or stale error, we've refreshed the metadata once and
// can record target errors.
bool recordTargetErrors = numTargetErrors > 0 || numStaleBatches > 0;
Status targetStatus = batchOp.targetBatch( *_targeter,
recordTargetErrors,
&childBatches );
if ( !targetStatus.isOK() ) {
_targeter->noteCouldNotTarget();
++numTargetErrors;
continue;
}
//
// Send all child batches
//
size_t numSent = 0;
while ( numSent != childBatches.size() ) {
// Collect batches out on the network, mapped by endpoint
EndpointBatchMap pendingBatches;
//
// Send side
//
// Get as many batches as we can at once
for ( vector<TargetedWriteBatch*>::iterator it = childBatches.begin();
it != childBatches.end(); ++it ) {
TargetedWriteBatch* nextBatch = *it;
// If the batch is NULL, we sent it previously, so skip
if ( nextBatch == NULL ) continue;
const ConnectionString& hostEndpoint = nextBatch->getEndpoint().shardHost;
EndpointBatchMap::iterator pendingIt = pendingBatches.find( &hostEndpoint );
// If we already have a batch for this endpoint, continue
if ( pendingIt != pendingBatches.end() ) continue;
// Otherwise send it out to the endpoint via a command to a database
BatchedCommandRequest request( clientRequest.getBatchType() );
batchOp.buildBatchRequest( *nextBatch, &request );
// Internally we use full namespaces for request/response, but we send the
// command to a database with the collection name in the request.
NamespaceString nss( request.getNS() );
request.setNS( nss.coll() );
_dispatcher->addCommand( hostEndpoint, nss.db(), request );
// Indicate we're done by setting the batch to NULL
// We'll only get duplicate hostEndpoints if we have broadcast and non-broadcast
// endpoints for the same host, so this should be pretty efficient without
// moving stuff around.
*it = NULL;
// Recv-side is responsible for cleaning up the nextBatch when used
pendingBatches.insert( make_pair( &hostEndpoint, nextBatch ) );
}
// Send them all out
_dispatcher->sendAll();
numSent += pendingBatches.size();
//
// Recv side
//
while ( _dispatcher->numPending() > 0 ) {
// Get the response
ConnectionString endpoint;
BatchedCommandResponse response;
Status dispatchStatus = _dispatcher->recvAny( &endpoint, &response );
// Get the TargetedWriteBatch to find where to put the response
TargetedWriteBatch* batchRaw = pendingBatches.find( &endpoint )->second;
scoped_ptr<TargetedWriteBatch> batch( batchRaw );
if ( dispatchStatus.isOK() ) {
TrackedErrors trackedErrors;
trackedErrors.startTracking( ErrorCodes::StaleShardVersion );
// Dispatch was ok, note response
batchOp.noteBatchResponse( *batch, response, &trackedErrors );
// Note if anything was stale
const vector<ShardError*>& staleErrors =
trackedErrors.getErrors( ErrorCodes::StaleShardVersion );
if ( staleErrors.size() > 0 ) {
noteStaleResponses( staleErrors, _targeter );
++numStaleBatches;
}
}
else {
// Error occurred dispatching, note it
BatchedErrorDetail error;
buildErrorFrom( dispatchStatus, &error );
batchOp.noteBatchError( *batch, error );
}
}
}
}
batchOp.buildClientResponse( clientResponse );
}
void BatchWriteOp::noteBatchResponse( const TargetedWriteBatch& targetedBatch,
const BatchedCommandResponse& response,
TrackedErrors* trackedErrors ) {
//
// Organize errors based on error code.
// We may have *either* a batch error or errors per-item.
// (Write Concern errors are stored and handled later.)
//
vector<BatchedErrorDetail*> itemErrors;
scoped_ptr<BatchedErrorDetail> batchError;
if ( !response.getOk() ) {
int errCode = response.getErrCode();
bool isWCError = isWCErrCode( errCode );
// Special handling for write concern errors, save for later
if ( isWCError ) {
BatchedErrorDetail error;
cloneBatchErrorTo( response, &error );
ShardError* wcError = new ShardError( targetedBatch.getEndpoint(), error );
_wcErrors.mutableVector().push_back( wcError );
}
// Handle batch and per-item errors
if ( response.isErrDetailsSet() ) {
// Per-item errors were set
itemErrors.insert( itemErrors.begin(),
response.getErrDetails().begin(),
response.getErrDetails().end() );
// Sort per-item errors by index
std::sort( itemErrors.begin(), itemErrors.end(), BatchedErrorDetailComp() );
}
else if ( !isWCError ) {
// Per-item errors were not set and this error is not a WC error
// => this is a full-batch error
batchError.reset( new BatchedErrorDetail );
cloneBatchErrorTo( response, batchError.get() );
}
}
// We can't have both a batch error and per-item errors
dassert( !( batchError && !itemErrors.empty() ) );
//
// Go through all pending responses of the op and sorted remote reponses, populate errors
// This will either set all errors to the batch error or apply per-item errors as-needed
//
vector<BatchedErrorDetail*>::iterator itemErrorIt = itemErrors.begin();
int index = 0;
for ( vector<TargetedWrite*>::const_iterator it = targetedBatch.getWrites().begin();
it != targetedBatch.getWrites().end(); ++it, ++index ) {
const TargetedWrite* write = *it;
WriteOp& writeOp = _writeOps[write->writeOpRef.first];
dassert( writeOp.getWriteState() == WriteOpState_Pending );
// See if we have an error for the write
BatchedErrorDetail* writeError = NULL;
if ( batchError ) {
// Default to batch error, if it exists
writeError = batchError.get();
}
else if ( itemErrorIt != itemErrors.end() && ( *itemErrorIt )->getIndex() == index ) {
// We have an per-item error for this write op's index
writeError = *itemErrorIt;
++itemErrorIt;
}
// Finish the response (with error, if needed)
if ( NULL == writeError ) {
writeOp.noteWriteComplete( *write );
}
else {
writeOp.noteWriteError( *write, *writeError );
}
}
// Track errors we care about, whether batch or individual errors
if ( NULL != trackedErrors ) {
trackErrors( targetedBatch.getEndpoint(), batchError.get(), itemErrors, trackedErrors );
}
// Stop tracking targeted batch
_targeted.erase( &targetedBatch );
}
Status BatchWriteOp::targetBatch( const NSTargeter& targeter,
bool recordTargetErrors,
vector<TargetedWriteBatch*>* targetedBatches ) {
TargetedBatchMap batchMap;
size_t numWriteOps = _clientRequest->sizeWriteOps();
for ( size_t i = 0; i < numWriteOps; ++i ) {
// Only do one-at-a-time ops if COE is false
if ( !_clientRequest->getContinueOnError() && !batchMap.empty() ) break;
WriteOp& writeOp = _writeOps[i];
// Only target _Ready ops
if ( writeOp.getWriteState() != WriteOpState_Ready ) continue;
//
// Get TargetedWrites from the targeter for the write operation
//
// TargetedWrites need to be owned once returned
OwnedPointerVector<TargetedWrite> writesOwned;
vector<TargetedWrite*>& writes = writesOwned.mutableVector();
Status targetStatus = writeOp.targetWrites( targeter, &writes );
if ( !targetStatus.isOK() ) {
//
// We're not sure how to target here, so either record the error or cancel the
// current batches.
//
BatchedErrorDetail targetError;
buildTargetError( targetStatus, &targetError );
if ( recordTargetErrors ) {
writeOp.setOpError( targetError );
continue;
}
else {
// Cancel current batch state with an error
cancelBatches( targetError, _writeOps, &batchMap );
dassert( batchMap.empty() );
return targetStatus;
}
}
//
// Targeting went ok, add to appropriate TargetedBatch
//
for ( vector<TargetedWrite*>::iterator it = writes.begin(); it != writes.end(); ++it ) {
TargetedWrite* write = *it;
TargetedBatchMap::iterator seenIt = batchMap.find( &write->endpoint );
if ( seenIt == batchMap.end() ) {
TargetedWriteBatch* newBatch = new TargetedWriteBatch( write->endpoint );
seenIt = batchMap.insert( make_pair( &newBatch->getEndpoint(), //
newBatch ) ).first;
}
TargetedWriteBatch* batch = seenIt->second;
batch->addWrite( write );
}
// Relinquish ownership of TargetedWrites, now the TargetedBatches own them
writesOwned.mutableVector().clear();
}
//
// Send back our targeted batches
//
for ( TargetedBatchMap::iterator it = batchMap.begin(); it != batchMap.end(); ++it ) {
TargetedWriteBatch* batch = it->second;
// Remember targeted batch for reporting
_targeted.insert( batch );
// Send the handle back to caller
targetedBatches->push_back( batch );
}
return Status::OK();
}
