Allow cluster to start when only n/2 servers are up (#4066)

Fixes #3234.

Relaxes the quorum requirement to start the object layer, and skips
quick-healing at start-up (as no write quorum is present).

cmd/prepare-storage.go

@@ -132,13 +132,14 @@ func prepForInitXL(firstDisk bool, sErrs []error, diskCount int) InitActions {
 	}
 
 	quorum := diskCount/2 + 1
+	readQuorum := diskCount / 2
 	disksOffline := errMap[errDiskNotFound]
 	disksFormatted := errMap[nil]
 	disksUnformatted := errMap[errUnformattedDisk]
 	disksCorrupted := errMap[errCorruptedFormat]
 
 	// No Quorum lots of offline disks, wait for quorum.
-	if disksOffline >= quorum {
+	if disksOffline > readQuorum {
 		return WaitForQuorum
 	}
 
@@ -168,7 +169,7 @@ func prepForInitXL(firstDisk bool, sErrs []error, diskCount int) InitActions {
 	}
 
 	// Already formatted and in quorum, proceed to initialization of object layer.
-	if disksFormatted >= quorum {
+	if disksFormatted >= readQuorum {
 		if disksFormatted+disksOffline == diskCount {
 			return InitObjectLayer
 		}
@@ -293,7 +294,7 @@ func initStorageDisks(endpoints []*url.URL) ([]StorageAPI, error) {
 	return storageDisks, nil
 }
 
-// Format disks before initialization object layer.
+// Format disks before initialization of object layer.
 func waitForFormatXLDisks(firstDisk bool, endpoints []*url.URL, storageDisks []StorageAPI) (formattedDisks []StorageAPI, err error) {
 	if len(endpoints) == 0 {
 		return nil, errInvalidArgument

cmd/xl-v1.go

@@ -155,6 +155,16 @@ func newXLObjects(storageDisks []StorageAPI) (ObjectLayer, error) {
 	xl.readQuorum = readQuorum
 	xl.writeQuorum = writeQuorum
 
+	// If the number of offline servers is equal to the readQuorum
+	// (i.e. the number of online servers also equals the
+	// readQuorum), we cannot perform quick-heal (no
+	// write-quorum). However reads may still be possible, so we
+	// skip quick-heal in this case, and continue.
+	offlineCount := len(newStorageDisks) - diskCount(newStorageDisks)
+	if offlineCount == readQuorum {
+		return xl, nil
+	}
+
 	// Do a quick heal on the buckets themselves for any discrepancies.
 	return xl, quickHeal(xl.storageDisks, xl.writeQuorum, xl.readQuorum)
 }