Status onShardVersionMismatch(OperationContext* opCtx,
const NamespaceString& nss,
ChunkVersion shardVersionReceived,
bool forceRefreshFromThisThread) noexcept {
invariant(!opCtx->lockState()->isLocked());
invariant(!opCtx->getClient()->isInDirectClient());
auto const shardingState = ShardingState::get(opCtx);
invariant(shardingState->canAcceptShardedCommands());
LOG(2) << "Metadata refresh requested for " << nss.ns() << " at shard version "
<< shardVersionReceived;
ShardingStatistics::get(opCtx).countStaleConfigErrors.addAndFetch(1);
// Ensure any ongoing migrations have completed before trying to do the refresh. This wait is
// just an optimization so that MongoS does not exhaust its maximum number of StaleConfig retry
// attempts while the migration is being committed.
try {
auto& oss = OperationShardingState::get(opCtx);
oss.waitForMigrationCriticalSectionSignal(opCtx);
} catch (const DBException& ex) {
return ex.toStatus();
}
const auto currentShardVersion = [&] {
AutoGetCollection autoColl(opCtx, nss, MODE_IS);
const auto currentMetadata = CollectionShardingState::get(opCtx, nss)->getMetadata(opCtx);
if (currentMetadata) {
return currentMetadata->getShardVersion();
}
return ChunkVersion::UNSHARDED();
}();
if (currentShardVersion.epoch() == shardVersionReceived.epoch() &&
currentShardVersion.majorVersion() >= shardVersionReceived.majorVersion()) {
// Don't need to remotely reload if we're in the same epoch and the requested version is
// smaller than the one we know about. This means that the remote side is behind.
return Status::OK();
}
try {
forceShardFilteringMetadataRefresh(opCtx, nss, forceRefreshFromThisThread);
return Status::OK();
} catch (const DBException& ex) {
log() << "Failed to refresh metadata for collection" << nss << causedBy(redact(ex));
return ex.toStatus();
}
}
TEST_F(MergeChunkTests, CompoundMerge) {
const NamespaceString nss("foo.bar");
const BSONObj kp = BSON("x" << 1 << "y" << 1);
const OID epoch = OID::gen();
vector<KeyRange> ranges;
// Setup chunk metadata
ranges.push_back(
KeyRange(nss.ns(), BSON("x" << 0 << "y" << 1), BSON("x" << 1 << "y" << 0), kp));
ranges.push_back(
KeyRange(nss.ns(), BSON("x" << 1 << "y" << 0), BSON("x" << 2 << "y" << 1), kp));
storeCollectionRanges(nss, shardName(), ranges, ChunkVersion(1, 0, epoch));
// Get latest version
ChunkVersion latestVersion;
ShardingState::get(getGlobalServiceContext())
->refreshMetadataNow(&_txn, nss.ns(), &latestVersion);
ShardingState::get(getGlobalServiceContext())->resetMetadata(nss.ns());
// Do merge
string errMsg;
bool result = mergeChunks(
&_txn, nss, BSON("x" << 0 << "y" << 1), BSON("x" << 2 << "y" << 1), epoch, &errMsg);
ASSERT_EQUALS(errMsg, "");
ASSERT(result);
// Verify result
CollectionMetadataPtr metadata =
ShardingState::get(getGlobalServiceContext())->getCollectionMetadata(nss.ns());
ChunkType chunk;
ASSERT(metadata->getNextChunk(BSON("x" << 0 << "y" << 1), &chunk));
ASSERT(chunk.getMin().woCompare(BSON("x" << 0 << "y" << 1)) == 0);
ASSERT(chunk.getMax().woCompare(BSON("x" << 2 << "y" << 1)) == 0);
ASSERT_EQUALS(metadata->getNumChunks(), 1u);
ASSERT_EQUALS(metadata->getShardVersion().majorVersion(), latestVersion.majorVersion());
ASSERT_GREATER_THAN(metadata->getShardVersion().minorVersion(), latestVersion.minorVersion());
assertWrittenAsMerged(ranges);
}
MigrationSourceManager::MigrationSourceManager(OperationContext* txn, MoveChunkRequest request)
: _args(std::move(request)), _startTime() {
invariant(!txn->lockState()->isLocked());
const auto& oss = OperationShardingState::get(txn);
if (!oss.hasShardVersion()) {
uasserted(ErrorCodes::InvalidOptions, "collection version is missing");
}
// Even though the moveChunk command transmits a value in the operation's shardVersion field,
// this value does not actually contain the shard version, but the global collection version.
const ChunkVersion expectedCollectionVersion = oss.getShardVersion(_args.getNss());
log() << "Starting chunk migration for "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " with expected collection version " << expectedCollectionVersion;
// Now that the collection is locked, snapshot the metadata and fetch the latest versions
ShardingState* const shardingState = ShardingState::get(txn);
ChunkVersion shardVersion;
Status refreshStatus =
shardingState->refreshMetadataNow(txn, _args.getNss().ns(), &shardVersion);
if (!refreshStatus.isOK()) {
uasserted(refreshStatus.code(),
str::stream() << "cannot start migrate of chunk "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " due to "
<< refreshStatus.toString());
}
if (shardVersion.majorVersion() == 0) {
// If the major version is zero, this means we do not have any chunks locally to migrate in
// the first place
uasserted(ErrorCodes::IncompatibleShardingMetadata,
str::stream() << "cannot start migrate of chunk "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " with zero shard version");
}
// Snapshot the committed metadata from the time the migration starts
{
ScopedTransaction scopedXact(txn, MODE_IS);
AutoGetCollection autoColl(txn, _args.getNss(), MODE_IS);
auto css = CollectionShardingState::get(txn, _args.getNss());
_committedMetadata = css->getMetadata();
}
const ChunkVersion collectionVersion = _committedMetadata->getCollVersion();
if (expectedCollectionVersion.epoch() != collectionVersion.epoch()) {
throw SendStaleConfigException(
_args.getNss().ns(),
str::stream() << "cannot move chunk "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " because collection may have been dropped. "
<< "current epoch: "
<< collectionVersion.epoch()
<< ", cmd epoch: "
<< expectedCollectionVersion.epoch(),
expectedCollectionVersion,
collectionVersion);
}
// With nonzero shard version, we must have a coll version >= our shard version
invariant(collectionVersion >= shardVersion);
// With nonzero shard version, we must have a shard key
invariant(!_committedMetadata->getKeyPattern().isEmpty());
ChunkType origChunk;
if (!_committedMetadata->getNextChunk(_args.getMinKey(), &origChunk)) {
// If this assertion is hit, it means that whoever called the shard moveChunk command
// (mongos or the CSRS balancer) did not check whether the chunk actually belongs to this
// shard. It is a benign error and does not indicate data corruption.
uasserted(40145,
str::stream() << "Chunk with bounds "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " is not owned by this shard.");
}
uassert(40146,
str::stream() << "Unable to find chunk with the exact bounds "
<< ChunkRange(_args.getMinKey(), _args.getMaxKey()).toString()
<< " at collection version "
<< collectionVersion.toString()
<< ". This indicates corrupted metadata.",
origChunk.getMin().woCompare(_args.getMinKey()) == 0 &&
origChunk.getMax().woCompare(_args.getMaxKey()) == 0);
}
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , string& errmsg, ChunkVersion& received, ChunkVersion& wanted ) {
if ( ! shardingState.enabled() )
return true;
if ( ! isMasterNs( ns.c_str() ) ) {
// right now connections to secondaries aren't versioned at all
return true;
}
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO : all collections at some point, be sharded or not, will have a version
// (and a CollectionMetadata)
received = info->getVersion( ns );
wanted = shardingState.getVersion( ns );
if( received.isWriteCompatibleWith( wanted ) ) return true;
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
// Check epoch first, to send more meaningful message, since other parameters probably
// won't match either
if( ! wanted.hasCompatibleEpoch( received ) ){
errmsg = str::stream() << "version epoch mismatch detected for " << ns << ", "
<< "the collection may have been dropped and recreated";
return false;
}
if( ! wanted.isSet() && received.isSet() ){
errmsg = str::stream() << "this shard no longer contains chunks for " << ns << ", "
<< "the collection may have been dropped";
return false;
}
if( wanted.isSet() && ! received.isSet() ){
errmsg = str::stream() << "this shard contains versioned chunks for " << ns << ", "
<< "but no version set in request";
return false;
}
if( wanted.majorVersion() != received.majorVersion() ){
//
// Could be > or < - wanted is > if this is the source of a migration,
// wanted < if this is the target of a migration
//
errmsg = str::stream() << "version mismatch detected for " << ns << ", "
<< "stored major version " << wanted.majorVersion()
<< " does not match received " << received.majorVersion();
return false;
}
// Those are all the reasons the versions can mismatch
verify( false );
return false;
}