// TestSendRPCRetry verifies that sendRPC failed on first address but succeed on
// second address, the second reply should be successfully returned back.
func TestSendRPCRetry(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
if err := g.SetNodeDescriptor(&proto.NodeDescriptor{NodeID: 1}); err != nil {
t.Fatal(err)
}
// Fill RangeDescriptor with 2 replicas
var descriptor = proto.RangeDescriptor{
RaftID: 1,
StartKey: proto.Key("a"),
EndKey: proto.Key("z"),
}
for i := 1; i <= 2; i++ {
addr := util.MakeUnresolvedAddr("tcp", fmt.Sprintf("node%d", i))
nd := &proto.NodeDescriptor{
NodeID: proto.NodeID(i),
Address: proto.Addr{
Network: addr.Network(),
Address: addr.String(),
},
}
if err := g.AddInfo(gossip.MakeNodeIDKey(proto.NodeID(i)), nd, time.Hour); err != nil {
t.Fatal(err)
}
descriptor.Replicas = append(descriptor.Replicas, proto.Replica{
NodeID: proto.NodeID(i),
StoreID: proto.StoreID(i),
})
}
// Define our rpcSend stub which returns success on the second address.
var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) interface{}, getReply func() interface{}, _ *rpc.Context) ([]interface{}, error) {
if method == "Node.Scan" {
// reply from first address failed
_ = getReply()
// reply from second address succeed
reply := getReply()
reply.(*proto.ScanResponse).Rows = append([]proto.KeyValue{}, proto.KeyValue{Key: proto.Key("b"), Value: proto.Value{}})
return []interface{}{reply}, nil
}
return nil, util.Errorf("Not expected method %v", method)
}
ctx := &DistSenderContext{
rpcSend: testFn,
rangeDescriptorDB: mockRangeDescriptorDB(func(_ proto.Key, _ lookupOptions) ([]proto.RangeDescriptor, error) {
return []proto.RangeDescriptor{descriptor}, nil
}),
}
ds := NewDistSender(ctx, g)
call := proto.ScanCall(proto.Key("a"), proto.Key("d"), 1)
sr := call.Reply.(*proto.ScanResponse)
ds.Send(context.Background(), call)
if err := sr.GoError(); err != nil {
t.Fatal(err)
}
if l := len(sr.Rows); l != 1 {
t.Fatalf("expected 1 row; got %d", l)
}
}
// TestMultiRangeScanReverseScanInconsistent verifies that a Scan/ReverseScan
// across ranges that doesn't require read consistency will set a timestamp
// using the clock local to the distributed sender.
func TestMultiRangeScanReverseScanInconsistent(t *testing.T) {
defer leaktest.AfterTest(t)
s, db := setupMultipleRanges(t, "b")
defer s.Stop()
// Write keys "a" and "b", the latter of which is the first key in the
// second range.
keys := []string{"a", "b"}
ts := []time.Time{}
b := &client.Batch{}
for _, key := range keys {
b.Put(key, "value")
}
if err := db.Run(b); err != nil {
t.Fatal(err)
}
for i := range keys {
ts = append(ts, b.Results[i].Rows[0].Timestamp())
log.Infof("%d: %s", i, b.Results[i].Rows[0].Timestamp())
}
// Do an inconsistent Scan/ReverseScan from a new DistSender and verify
// it does the read at its local clock and doesn't receive an
// OpRequiresTxnError. We set the local clock to the timestamp of
// the first key to verify it's used to read only key "a".
manual := hlc.NewManualClock(ts[1].UnixNano() - 1)
clock := hlc.NewClock(manual.UnixNano)
ds := kv.NewDistSender(&kv.DistSenderContext{Clock: clock}, s.Gossip())
// Scan.
call := proto.ScanCall(proto.Key("a"), proto.Key("c"), 0)
sr := call.Reply.(*proto.ScanResponse)
sa := call.Args.(*proto.ScanRequest)
sa.ReadConsistency = proto.INCONSISTENT
if err := client.SendCall(ds, call); err != nil {
t.Fatal(err)
}
if l := len(sr.Rows); l != 1 {
t.Fatalf("expected 1 row; got %d", l)
}
if key := string(sr.Rows[0].Key); keys[0] != key {
t.Errorf("expected key %q; got %q", keys[0], key)
}
// ReverseScan.
call = proto.ReverseScanCall(proto.Key("a"), proto.Key("c"), 0)
rsr := call.Reply.(*proto.ReverseScanResponse)
rsa := call.Args.(*proto.ReverseScanRequest)
rsa.ReadConsistency = proto.INCONSISTENT
if err := client.SendCall(ds, call); err != nil {
t.Fatal(err)
}
if l := len(rsr.Rows); l != 1 {
t.Fatalf("expected 1 row; got %d", l)
}
if key := string(rsr.Rows[0].Key); keys[0] != key {
t.Errorf("expected key %q; got %q", keys[0], key)
}
}
// TestMultiRangeScanWithMaxResults tests that commands which access multiple
// ranges with MaxResults parameter are carried out properly.
func TestMultiRangeScanWithMaxResults(t *testing.T) {
defer leaktest.AfterTest(t)
testCases := []struct {
splitKeys []proto.Key
keys []proto.Key
}{
{[]proto.Key{proto.Key("m")},
[]proto.Key{proto.Key("a"), proto.Key("z")}},
{[]proto.Key{proto.Key("h"), proto.Key("q")},
[]proto.Key{proto.Key("b"), proto.Key("f"), proto.Key("k"),
proto.Key("r"), proto.Key("w"), proto.Key("y")}},
}
for i, tc := range testCases {
s := StartTestServer(t)
ds := kv.NewDistSender(&kv.DistSenderContext{Clock: s.Clock()}, s.Gossip())
tds := kv.NewTxnCoordSender(ds, s.Clock(), testContext.Linearizable, s.stopper)
for _, sk := range tc.splitKeys {
if err := s.node.ctx.DB.AdminSplit(sk); err != nil {
t.Fatal(err)
}
}
var call proto.Call
for _, k := range tc.keys {
call = proto.PutCall(k, proto.Value{Bytes: k})
call.Args.Header().User = storage.UserRoot
tds.Send(context.Background(), call)
if err := call.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
}
// Try every possible ScanRequest startKey.
for start := 0; start < len(tc.keys); start++ {
// Try every possible maxResults, from 1 to beyond the size of key array.
for maxResults := 1; maxResults <= len(tc.keys)-start+1; maxResults++ {
scan := proto.ScanCall(tc.keys[start], tc.keys[len(tc.keys)-1].Next(),
int64(maxResults))
scan.Args.Header().Timestamp = call.Reply.Header().Timestamp
scan.Args.Header().User = storage.UserRoot
tds.Send(context.Background(), scan)
if err := scan.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
rows := scan.Reply.(*proto.ScanResponse).Rows
if start+maxResults <= len(tc.keys) && len(rows) != maxResults {
t.Fatalf("%d: start=%s: expected %d rows, but got %d", i, tc.keys[start], maxResults, len(rows))
} else if start+maxResults == len(tc.keys)+1 && len(rows) != maxResults-1 {
t.Fatalf("%d: expected %d rows, but got %d", i, maxResults-1, len(rows))
}
}
}
defer s.Stop()
}
}
// TestRetryOnWrongReplicaError sets up a DistSender on a minimal gossip
// network and a mock of rpc.Send, and verifies that the DistSender correctly
// retries upon encountering a stale entry in its range descriptor cache.
func TestRetryOnWrongReplicaError(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
// Updated below, after it has first been returned.
badStartKey := proto.Key("m")
newRangeDescriptor := testRangeDescriptor
goodStartKey := newRangeDescriptor.StartKey
newRangeDescriptor.StartKey = badStartKey
descStale := true
var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) gogoproto.Message, getReply func() gogoproto.Message, _ *rpc.Context) ([]gogoproto.Message, error) {
ba := getArgs(testAddress).(*proto.BatchRequest)
if _, ok := ba.GetArg(proto.RangeLookup); ok {
if !descStale && bytes.HasPrefix(ba.Key, keys.Meta2Prefix) {
t.Errorf("unexpected extra lookup for non-stale replica descriptor at %s",
ba.Key)
}
br := getReply().(*proto.BatchResponse)
r := &proto.RangeLookupResponse{}
r.Ranges = append(r.Ranges, newRangeDescriptor)
br.Add(r)
// If we just returned the stale descriptor, set up returning the
// good one next time.
if bytes.HasPrefix(ba.Key, keys.Meta2Prefix) {
if newRangeDescriptor.StartKey.Equal(badStartKey) {
newRangeDescriptor.StartKey = goodStartKey
} else {
descStale = false
}
}
return []gogoproto.Message{br}, nil
}
// When the Scan first turns up, update the descriptor for future
// range descriptor lookups.
if !newRangeDescriptor.StartKey.Equal(goodStartKey) {
return nil, &proto.RangeKeyMismatchError{RequestStartKey: ba.Key,
RequestEndKey: ba.EndKey}
}
return []gogoproto.Message{getReply()}, nil
}
ctx := &DistSenderContext{
RPCSend: testFn,
}
ds := NewDistSender(ctx, g)
call := proto.ScanCall(proto.Key("a"), proto.Key("d"), 0)
sr := call.Reply.(*proto.ScanResponse)
client.SendCallConverted(ds, context.Background(), call)
if err := sr.GoError(); err != nil {
t.Errorf("scan encountered error: %s", err)
}
}
// TestRetryOnWrongReplicaError sets up a DistSender on a minimal gossip
// network and a mock of rpc.Send, and verifies that the DistSender correctly
// retries upon encountering a stale entry in its range descriptor cache.
func TestRetryOnWrongReplicaError(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
// Updated below, after it has first been returned.
newRangeDescriptor := testRangeDescriptor
newEndKey := proto.Key("m")
descStale := true
var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) interface{}, getReply func() interface{}, _ *rpc.Context) ([]interface{}, error) {
header := getArgs(testAddress).(proto.Request).Header()
if method == "Node.InternalRangeLookup" {
// If the non-broken descriptor has already been returned, that's
// an error.
if !descStale && bytes.HasPrefix(header.Key, keys.Meta2Prefix) {
t.Errorf("unexpected extra lookup for non-stale replica descriptor at %s",
header.Key)
}
r := getReply().(*proto.InternalRangeLookupResponse)
// The fresh descriptor is about to be returned.
if bytes.HasPrefix(header.Key, keys.Meta2Prefix) &&
newRangeDescriptor.StartKey.Equal(newEndKey) {
descStale = false
}
r.Ranges = append(r.Ranges, newRangeDescriptor)
return []interface{}{r}, nil
}
// When the Scan first turns up, update the descriptor for future
// range descriptor lookups.
if !newRangeDescriptor.StartKey.Equal(newEndKey) {
newRangeDescriptor = *gogoproto.Clone(&testRangeDescriptor).(*proto.RangeDescriptor)
newRangeDescriptor.StartKey = newEndKey
return nil, &proto.RangeKeyMismatchError{RequestStartKey: header.Key,
RequestEndKey: header.EndKey}
}
return []interface{}{getReply()}, nil
}
ctx := &DistSenderContext{
rpcSend: testFn,
}
ds := NewDistSender(ctx, g)
call := proto.ScanCall(proto.Key("a"), proto.Key("d"), 0)
sr := call.Reply.(*proto.ScanResponse)
ds.Send(context.Background(), call)
if err := sr.GoError(); err != nil {
t.Errorf("scan encountered error: %s", err)
}
}
func (b *Batch) scan(s, e interface{}, maxRows int64, isReverse bool) {
begin, err := marshalKey(s)
if err != nil {
b.initResult(0, 0, err)
return
}
end, err := marshalKey(e)
if err != nil {
b.initResult(0, 0, err)
return
}
if !isReverse {
b.calls = append(b.calls, proto.ScanCall(proto.Key(begin), proto.Key(end), maxRows))
} else {
b.calls = append(b.calls, proto.ReverseScanCall(proto.Key(begin), proto.Key(end), maxRows))
}
b.initResult(1, 0, nil)
}
// TestMultiRangeScanDeleteRange tests that commands which access multiple
// ranges are carried out properly.
func TestMultiRangeScanDeleteRange(t *testing.T) {
defer leaktest.AfterTest(t)
s := StartTestServer(t)
defer s.Stop()
ds := kv.NewDistSender(&kv.DistSenderContext{Clock: s.Clock()}, s.Gossip())
tds := kv.NewTxnCoordSender(ds, s.Clock(), testContext.Linearizable, s.stopper)
if err := s.node.ctx.DB.AdminSplit("m"); err != nil {
t.Fatal(err)
}
writes := []proto.Key{proto.Key("a"), proto.Key("z")}
get := proto.Call{
Args: &proto.GetRequest{
RequestHeader: proto.RequestHeader{
Key: writes[0],
},
},
Reply: &proto.GetResponse{},
}
get.Args.Header().User = storage.UserRoot
get.Args.Header().EndKey = writes[len(writes)-1]
tds.Send(context.Background(), get)
if err := get.Reply.Header().GoError(); err == nil {
t.Errorf("able to call Get with a key range: %v", get)
}
var call proto.Call
for i, k := range writes {
call = proto.PutCall(k, proto.Value{Bytes: k})
call.Args.Header().User = storage.UserRoot
tds.Send(context.Background(), call)
if err := call.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
scan := proto.ScanCall(writes[0], writes[len(writes)-1].Next(), 0)
// The Put ts may have been pushed by tsCache,
// so make sure we see their values in our Scan.
scan.Args.Header().Timestamp = call.Reply.Header().Timestamp
scan.Args.Header().User = storage.UserRoot
tds.Send(context.Background(), scan)
if err := scan.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
if scan.Reply.Header().Txn == nil {
t.Errorf("expected Scan to be wrapped in a Transaction")
}
if rows := scan.Reply.(*proto.ScanResponse).Rows; len(rows) != i+1 {
t.Fatalf("expected %d rows, but got %d", i+1, len(rows))
}
}
del := proto.Call{
Args: &proto.DeleteRangeRequest{
RequestHeader: proto.RequestHeader{
User: storage.UserRoot,
Key: writes[0],
EndKey: proto.Key(writes[len(writes)-1]).Next(),
Timestamp: call.Reply.Header().Timestamp,
},
},
Reply: &proto.DeleteRangeResponse{},
}
tds.Send(context.Background(), del)
if err := del.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
if del.Reply.Header().Txn == nil {
t.Errorf("expected DeleteRange to be wrapped in a Transaction")
}
if n := del.Reply.(*proto.DeleteRangeResponse).NumDeleted; n != int64(len(writes)) {
t.Errorf("expected %d keys to be deleted, but got %d instead",
len(writes), n)
}
scan := proto.ScanCall(writes[0], writes[len(writes)-1].Next(), 0)
scan.Args.Header().Timestamp = del.Reply.Header().Timestamp
scan.Args.Header().User = storage.UserRoot
scan.Args.Header().Txn = &proto.Transaction{Name: "MyTxn"}
tds.Send(context.Background(), scan)
if err := scan.Reply.Header().GoError(); err != nil {
t.Fatal(err)
}
if txn := scan.Reply.Header().Txn; txn == nil || txn.Name != "MyTxn" {
t.Errorf("wanted Txn to persist, but it changed to %v", txn)
}
if rows := scan.Reply.(*proto.ScanResponse).Rows; len(rows) > 0 {
t.Fatalf("scan after delete returned rows: %v", rows)
}
}
// TestMultiRangeMergeStaleDescriptor simulates the situation in which the
// DistSender executes a multi-range scan which encounters the stale descriptor
// of a range which has since incorporated its right neighbor by means of a
// merge. It is verified that the DistSender scans the correct keyrange exactly
// once.
func TestMultiRangeMergeStaleDescriptor(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
// Assume we have two ranges, [a-b) and [b-KeyMax).
merged := false
// The stale first range descriptor which is unaware of the merge.
var firstRange = proto.RangeDescriptor{
RaftID: 1,
StartKey: proto.Key("a"),
EndKey: proto.Key("b"),
Replicas: []proto.Replica{
{
NodeID: 1,
StoreID: 1,
},
},
}
// The merged descriptor, which will be looked up after having processed
// the stale range [a,b).
var mergedRange = proto.RangeDescriptor{
RaftID: 1,
StartKey: proto.Key("a"),
EndKey: proto.KeyMax,
Replicas: []proto.Replica{
{
NodeID: 1,
StoreID: 1,
},
},
}
// Assume we have two key-value pairs, a=1 and c=2.
existingKVs := []proto.KeyValue{
{Key: proto.Key("a"), Value: proto.Value{Bytes: []byte("1")}},
{Key: proto.Key("c"), Value: proto.Value{Bytes: []byte("2")}},
}
var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) interface{}, getReply func() interface{}, _ *rpc.Context) ([]interface{}, error) {
if method != "Node.Scan" {
t.Fatalf("unexpected method:%s", method)
}
header := getArgs(testAddress).(proto.Request).Header()
reply := getReply().(*proto.ScanResponse)
results := []proto.KeyValue{}
for _, curKV := range existingKVs {
if header.Key.Less(curKV.Key.Next()) && curKV.Key.Less(header.EndKey) {
results = append(results, curKV)
}
}
reply.Rows = results
return []interface{}{reply}, nil
}
ctx := &DistSenderContext{
rpcSend: testFn,
rangeDescriptorDB: mockRangeDescriptorDB(func(key proto.Key, _ lookupOptions) ([]proto.RangeDescriptor, error) {
if !merged {
// Asume a range merge operation happened
merged = true
return []proto.RangeDescriptor{firstRange}, nil
}
return []proto.RangeDescriptor{mergedRange}, nil
}),
}
ds := NewDistSender(ctx, g)
call := proto.ScanCall(proto.Key("a"), proto.Key("d"), 10)
// Set the Txn info to avoid an OpRequiresTxnError.
call.Args.Header().Txn = &proto.Transaction{}
reply := call.Reply.(*proto.ScanResponse)
ds.Send(context.Background(), call)
if err := reply.GoError(); err != nil {
t.Fatalf("scan encountered error: %s", err)
}
if !reflect.DeepEqual(existingKVs, reply.Rows) {
t.Fatalf("expect get %v, actual get %v", existingKVs, reply.Rows)
}
}
// ScanStruct scans the specified columns from the structured table identified
// by the destination slice. The slice element type, start and end key types
// must be identical. The primary key columns within start and end are used to
// identify which rows to scan. The type must have previously been bound to a
// table using BindModel. If columns is empty all of the columns in the table
// are scanned.
func (b *Batch) ScanStruct(dest, start, end interface{}, maxRows int64, columns ...string) {
sliceV := reflect.ValueOf(dest)
if sliceV.Kind() != reflect.Ptr {
b.initResult(0, 0, fmt.Errorf("dest must be a pointer to a slice: %T", dest))
return
}
sliceV = sliceV.Elem()
if sliceV.Kind() != reflect.Slice {
b.initResult(0, 0, fmt.Errorf("dest must be a pointer to a slice: %T", dest))
return
}
modelT := sliceV.Type().Elem()
// Are we returning a slice of structs or pointers to structs?
ptrResults := modelT.Kind() == reflect.Ptr
if ptrResults {
modelT = modelT.Elem()
}
m, err := b.DB.getModel(modelT, false)
if err != nil {
b.initResult(0, 0, err)
return
}
var scanColIDs map[uint32]bool
if len(columns) > 0 {
lowerStrings(columns)
scanColIDs = make(map[uint32]bool, len(columns))
for _, colName := range columns {
col, ok := m.columnsByName[colName]
if !ok {
b.initResult(0, 0, fmt.Errorf("%s: unable to find column %s", m.name, colName))
return
}
scanColIDs[col.ID] = true
}
}
startV := reflect.Indirect(reflect.ValueOf(start))
if modelT != startV.Type() {
b.initResult(0, 0, fmt.Errorf("incompatible start key type: %s != %s", modelT, startV.Type()))
return
}
endV := reflect.Indirect(reflect.ValueOf(end))
if modelT != endV.Type() {
b.initResult(0, 0, fmt.Errorf("incompatible end key type: %s != %s", modelT, endV.Type()))
return
}
startKey, err := m.encodePrimaryKey(startV)
if err != nil {
b.initResult(0, 0, err)
return
}
endKey, err := m.encodePrimaryKey(endV)
if err != nil {
b.initResult(0, 0, err)
return
}
if log.V(2) {
log.Infof("Scan %q %q", startKey, endKey)
}
c := proto.ScanCall(proto.Key(startKey), proto.Key(endKey), maxRows)
c.Post = func() error {
reply := c.Reply.(*proto.ScanResponse)
if len(reply.Rows) == 0 {
return nil
}
var primaryKey []byte
resultPtr := reflect.New(modelT)
result := resultPtr.Elem()
zero := reflect.Zero(result.Type())
for _, row := range reply.Rows {
if primaryKey != nil && !bytes.HasPrefix(row.Key, primaryKey) {
if ptrResults {
sliceV = reflect.Append(sliceV, resultPtr)
resultPtr = reflect.New(modelT)
result = resultPtr.Elem()
} else {
sliceV = reflect.Append(sliceV, result)
result.Set(zero)
}
_, err := m.decodePrimaryKey(primaryKey, result)
if err != nil {
return err
}
}
remaining, err := m.decodePrimaryKey([]byte(row.Key), result)
if err != nil {
return err
}
primaryKey = []byte(row.Key[:len(row.Key)-len(remaining)])
_, colID := roachencoding.DecodeUvarint(remaining)
if err != nil {
return err
}
if scanColIDs != nil && !scanColIDs[uint32(colID)] {
continue
}
col, ok := m.columnsByID[uint32(colID)]
if !ok {
return fmt.Errorf("%s: unable to find column %d", m.name, colID)
}
if err := unmarshalValue(&row.Value, result.FieldByIndex(col.field.Index)); err != nil {
return err
}
}
if ptrResults {
sliceV = reflect.Append(sliceV, resultPtr)
} else {
sliceV = reflect.Append(sliceV, result)
}
reflect.ValueOf(dest).Elem().Set(sliceV)
return nil
}
b.calls = append(b.calls, c)
b.initResult(1, 0, nil)
}