bool canStartRollback(OplogReader& r, GTID idToRollbackTo) {
shared_ptr<DBClientConnection> conn(r.conn_shared());
// before we start rollback, let's make sure that the minUnapplied on the remote
// server is past the id that we are rolling back to. Otherwise, the snapshot
// we create will not be up to date, and the rollback algorithm will not work
BSONObjBuilder b;
b.append("_id", "minUnapplied");
// Note that another way to get this information is to
// request a heartbeat. That one will technically return
// a more up to date value for minUnapplied
BSONObj res = findOneFromConn(conn.get(), rsReplInfo, Query(b.done()));
GTID minUnapplied = getGTIDFromBSON("GTID", res);
if (GTID::cmp(minUnapplied, idToRollbackTo) < 0) {
log() << "Remote server has minUnapplied " << minUnapplied.toString() << \
" we want to rollback to " << idToRollbackTo.toString() << \
". Therefore, exiting and retrying." << rsLog;
return false;
}
return true;
}
/**
* Do the initial sync for this member.
* This code can use a little refactoring. bit ugly
*/
bool ReplSetImpl::_syncDoInitialSync() {
sethbmsg("initial sync pending",0);
bool needsFullSync = gtidManager->getLiveState().isInitial();
bool needGapsFilled = needsFullSync || replSettings.fastsync;
// if this is the first node, it may have already become primary
if ( box.getState().primary() ) {
sethbmsg("I'm already primary, no need for initial sync",0);
return true;
}
const Member *source = NULL;
OplogReader r;
string sourceHostname;
// only bother making a connection if we need to connect for some reason
if (needGapsFilled) {
source = getMemberToSyncTo();
if (!source) {
sethbmsg("initial sync need a member to be primary or secondary to do our initial sync", 0);
sleepsecs(15);
return false;
}
sourceHostname = source->h().toString();
if( !r.connect(sourceHostname, 0) ) {
sethbmsg( str::stream() << "initial sync couldn't connect to " << source->h().toString() , 0);
sleepsecs(15);
return false;
}
}
if( needsFullSync ) {
BSONObj lastOp = r.getLastOp();
if( lastOp.isEmpty() ) {
sethbmsg("initial sync couldn't read remote oplog", 0);
sleepsecs(15);
return false;
}
{
LOCK_REASON(lockReason, "repl: initial sync drop all databases");
Lock::GlobalWrite lk(lockReason);
Client::Transaction dropTransaction(DB_SERIALIZABLE);
sethbmsg("initial sync drop all databases", 0);
dropAllDatabasesExceptLocal();
dropTransaction.commit();
}
// now deal with creation of oplog
// first delete any existing data in the oplog
{
LOCK_REASON(lockReason, "repl: create oplog");
Lock::DBWrite lk("local", lockReason);
Client::Transaction fileOpsTransaction(DB_SERIALIZABLE);
deleteOplogFiles();
fileOpsTransaction.commit(0);
}
try {
sethbmsg("initial sync clone all databases", 0);
shared_ptr<DBClientConnection> conn(r.conn_shared());
RemoteTransaction rtxn(*conn, "mvcc");
list<string> dbs = conn->getDatabaseNamesForRepl();
//
// Not sure if it is necessary to have a separate fileOps
// transaction and clone transaction. The cloneTransaction
// has a higher chance of failing, and I don't know at the moment
// if it is ok to do fileops successfully, and then an operation (cloning) that
// later causes an abort. So, to be cautious, they are separate
{
LOCK_REASON(lockReason, "repl: initial sync");
Lock::GlobalWrite lk(lockReason);
Client::Transaction cloneTransaction(DB_SERIALIZABLE);
bool ret = _syncDoInitialSync_clone(sourceHostname.c_str(), dbs, conn);
if (!ret) {
veto(source->fullName(), 600);
cloneTransaction.abort();
sleepsecs(300);
return false;
}
// at this point, we have copied all of the data from the
// remote machine. Now we need to copy the replication information
// on the remote machine's local database, we need to copy
// the entire (small) replInfo dictionary, and the necessary portion
// of the oplog
// first copy the replInfo, as we will use its information
// to determine how much of the opLog to copy
{
Client::Context ctx( "local" );
BSONObj q;
cloneCollection(conn,
"local",
rsReplInfo,
q,
true, //copyIndexes
false //logForRepl
);
// copy entire oplog (probably overkill)
cloneCollection(conn,
"local",
rsoplog,
q,
true, //copyIndexes
false //logForRepl
);
// copy entire oplog.refs (probably overkill)
cloneCollection(conn,
"local",
rsOplogRefs,
q,
true, //copyIndexes
false //logForRepl
);
}
cloneTransaction.commit(0);
}
bool ok = rtxn.commit();
verify(ok); // absolutely no reason this should fail, it was read only
// data should now be consistent
}
catch (DBException &e) {
sethbmsg("exception trying to copy data", 0);
LOG(0) << e.getCode() << ": " << e.what() << endl;
sleepsecs(1);
return false;
}
}
if (needGapsFilled) {
_fillGaps(&r);
}
GTID dummy;
applyMissingOpsInOplog(dummy);
sethbmsg("initial sync done",0);
return true;
}