// TestTxnMultipleCoord checks that a coordinator uses the Writing flag to
// enforce that only one coordinator can be used for transactional writes.
func TestTxnMultipleCoord(t *testing.T) {
defer leaktest.AfterTest(t)
s := createTestDB(t)
defer s.Stop()
for i, tc := range []struct {
args proto.Request
writing bool
ok bool
}{
{proto.NewGet(proto.Key("a")), true, true},
{proto.NewGet(proto.Key("a")), false, true},
{proto.NewPut(proto.Key("a"), proto.Value{}), false, true},
{proto.NewPut(proto.Key("a"), proto.Value{}), true, false},
} {
{
txn := newTxn(s.Clock, proto.Key("a"))
txn.Writing = tc.writing
tc.args.Header().Txn = txn
}
reply, err := batchutil.SendWrapped(s.Sender, tc.args)
if err == nil != tc.ok {
t.Errorf("%d: %T (writing=%t): success_expected=%t, but got: %v",
i, tc.args, tc.writing, tc.ok, err)
}
if err != nil {
continue
}
txn := reply.Header().Txn
// The transaction should come back rw if it started rw or if we just
// wrote.
isWrite := proto.IsTransactionWrite(tc.args)
if (tc.writing || isWrite) != txn.Writing {
t.Errorf("%d: unexpected writing state: %s", i, txn)
}
if !isWrite {
continue
}
// Abort for clean shutdown.
if _, err := batchutil.SendWrapped(s.Sender, &proto.EndTransactionRequest{
RequestHeader: proto.RequestHeader{
Key: txn.Key,
Timestamp: txn.Timestamp,
Txn: txn,
},
Commit: false,
}); err != nil {
t.Fatal(err)
}
}
}
// 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, nil, s.stopper)
for _, sk := range tc.splitKeys {
if err := s.node.ctx.DB.AdminSplit(sk); err != nil {
t.Fatal(err)
}
}
var reply proto.Response
for _, k := range tc.keys {
put := proto.NewPut(k, proto.Value{Bytes: k})
var err error
reply, err = batchutil.SendWrapped(tds, put)
if 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.NewScan(tc.keys[start], tc.keys[len(tc.keys)-1].Next(),
int64(maxResults))
scan.Header().Timestamp = reply.Header().Timestamp
reply, err := batchutil.SendWrapped(tds, scan)
if err != nil {
t.Fatal(err)
}
rows := reply.(*proto.ScanResponse).Rows
if start+maxResults <= len(tc.keys) && len(rows) != maxResults {
t.Errorf("%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.Errorf("%d: expected %d rows, but got %d", i, maxResults-1, len(rows))
}
}
}
defer s.Stop()
}
}
// Put sets the value for a key.
//
// A new result will be appended to the batch which will contain a single row
// and Result.Err will indicate success or failure.
//
// key can be either a byte slice or a string. value can be any key type, a
// proto.Message or any Go primitive type (bool, int, etc).
func (b *Batch) Put(key, value interface{}) {
k, err := marshalKey(key)
if err != nil {
b.initResult(0, 1, err)
return
}
v, err := marshalValue(value)
if err != nil {
b.initResult(0, 1, err)
return
}
b.reqs = append(b.reqs, proto.NewPut(k, v))
b.initResult(1, 1, nil)
}
// TestRetryOnDescriptorLookupError verifies that the DistSender retries a descriptor
// lookup on retryable errors.
func TestRetryOnDescriptorLookupError(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
var testFn rpcSendFn = func(_ rpc.Options, _ string, _ []net.Addr, _ func(addr net.Addr) gogoproto.Message, getReply func() gogoproto.Message, _ *rpc.Context) ([]gogoproto.Message, error) {
return []gogoproto.Message{getReply()}, nil
}
errors := []error{
errors.New("fatal boom"),
&proto.RangeKeyMismatchError{}, // retryable
nil,
}
ctx := &DistSenderContext{
RPCSend: testFn,
RangeDescriptorDB: mockRangeDescriptorDB(func(k proto.Key, _ lookupOptions) ([]proto.RangeDescriptor, error) {
// Return next error and truncate the prefix of the errors array.
var err error
if k != nil {
err = errors[0]
errors = errors[1:]
}
return []proto.RangeDescriptor{testRangeDescriptor}, err
}),
}
ds := NewDistSender(ctx, g)
put := proto.NewPut(proto.Key("a"), proto.Value{Bytes: []byte("value")})
// Fatal error on descriptor lookup, propagated to reply.
if _, err := batchutil.SendWrapped(ds, put); err.Error() != "fatal boom" {
t.Errorf("unexpected error: %s", err)
}
// Retryable error on descriptor lookup, second attempt successful.
if _, err := batchutil.SendWrapped(ds, put); err != nil {
t.Errorf("unexpected error: %s", err)
}
if len(errors) != 0 {
t.Fatalf("expected more descriptor lookups, leftover errors: %+v", errors)
}
}
// TestRetryOnNotLeaderError verifies that the DistSender correctly updates the
// leader cache and retries when receiving a NotLeaderError.
func TestRetryOnNotLeaderError(t *testing.T) {
defer leaktest.AfterTest(t)
g, s := makeTestGossip(t)
defer s()
leader := proto.Replica{
NodeID: 99,
StoreID: 999,
}
first := 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) {
if first {
reply := getReply()
reply.(proto.Response).Header().SetGoError(
&proto.NotLeaderError{Leader: &leader, Replica: &proto.Replica{}})
first = false
return []gogoproto.Message{reply}, nil
}
return []gogoproto.Message{getReply()}, nil
}
ctx := &DistSenderContext{
RPCSend: testFn,
RangeDescriptorDB: mockRangeDescriptorDB(func(_ proto.Key, _ lookupOptions) ([]proto.RangeDescriptor, error) {
return []proto.RangeDescriptor{testRangeDescriptor}, nil
}),
}
ds := NewDistSender(ctx, g)
put := proto.NewPut(proto.Key("a"), proto.Value{Bytes: []byte("value")})
if _, err := batchutil.SendWrapped(ds, put); err != nil {
t.Errorf("put encountered error: %s", err)
}
if first {
t.Errorf("The command did not retry")
}
if cur := ds.leaderCache.Lookup(1); cur.StoreID != leader.StoreID {
t.Errorf("leader cache was not updated: expected %v, got %v",
&leader, cur)
}
}
func TestNodeEventFeed(t *testing.T) {
defer leaktest.AfterTest(t)
nodeDesc := proto.NodeDescriptor{
NodeID: proto.NodeID(99),
}
// A testCase corresponds to a single Store event type. Each case contains a
// method which publishes a single event to the given storeEventPublisher,
// and an expected result interface which should match the produced
// event.
testCases := []struct {
publishTo func(status.NodeEventFeed)
expected interface{}
}{
{
publishTo: func(nef status.NodeEventFeed) {
nef.StartNode(nodeDesc, 100)
},
expected: &status.StartNodeEvent{
Desc: nodeDesc,
StartedAt: 100,
},
},
{
publishTo: func(nef status.NodeEventFeed) {
nef.CallComplete(wrap(proto.NewGet(proto.Key("abc"))), nil)
},
expected: &status.CallSuccessEvent{
NodeID: proto.NodeID(1),
Method: proto.Get,
},
},
{
publishTo: func(nef status.NodeEventFeed) {
nef.CallComplete(wrap(proto.NewPut(proto.Key("abc"), proto.Value{Bytes: []byte("def")})), nil)
},
expected: &status.CallSuccessEvent{
NodeID: proto.NodeID(1),
Method: proto.Put,
},
},
{
publishTo: func(nef status.NodeEventFeed) {
nef.CallComplete(wrap(proto.NewGet(proto.Key("abc"))), proto.NewError(util.Errorf("error")))
},
expected: &status.CallErrorEvent{
NodeID: proto.NodeID(1),
Method: proto.Batch,
},
},
{
publishTo: func(nef status.NodeEventFeed) {
nef.CallComplete(wrap(proto.NewGet(proto.Key("abc"))), &proto.Error{
Index: &proto.ErrPosition{Index: 0},
Message: "boo",
})
},
expected: &status.CallErrorEvent{
NodeID: proto.NodeID(1),
Method: proto.Get,
},
},
}
// Compile expected events into a single slice.
expectedEvents := make([]interface{}, len(testCases))
for i := range testCases {
expectedEvents[i] = testCases[i].expected
}
events := make([]interface{}, 0, len(expectedEvents))
// Run test cases directly through a feed.
stopper := stop.NewStopper()
defer stopper.Stop()
feed := util.NewFeed(stopper)
feed.Subscribe(func(event interface{}) {
events = append(events, event)
})
nodefeed := status.NewNodeEventFeed(proto.NodeID(1), feed)
for _, tc := range testCases {
tc.publishTo(nodefeed)
}
feed.Flush()
if a, e := events, expectedEvents; !reflect.DeepEqual(a, e) {
t.Errorf("received incorrect events.\nexpected: %v\nactual: %v", e, a)
}
}
func TestEvictCacheOnError(t *testing.T) {
defer leaktest.AfterTest(t)
// if rpcError is true, the first attempt gets an RPC error, otherwise
// the RPC call succeeds but there is an error in the RequestHeader.
// Currently leader and cached range descriptor are treated equally.
testCases := []struct{ rpcError, retryable, shouldClearLeader, shouldClearReplica bool }{
{false, false, false, false}, // non-retryable replica error
{false, true, false, false}, // retryable replica error
{true, false, true, true}, // RPC error aka all nodes dead
{true, true, false, false}, // retryable RPC error
}
for i, tc := range testCases {
g, s := makeTestGossip(t)
defer s()
leader := proto.Replica{
NodeID: 99,
StoreID: 999,
}
first := true
var testFn rpcSendFn = func(_ rpc.Options, _ string, _ []net.Addr, _ func(addr net.Addr) gogoproto.Message, getReply func() gogoproto.Message, _ *rpc.Context) ([]gogoproto.Message, error) {
if !first {
return []gogoproto.Message{getReply()}, nil
}
first = false
if tc.rpcError {
return nil, rpc.NewSendError("boom", tc.retryable)
}
var err error
if tc.retryable {
err = &proto.RangeKeyMismatchError{}
} else {
err = errors.New("boom")
}
reply := getReply()
reply.(proto.Response).Header().SetGoError(err)
return []gogoproto.Message{reply}, nil
}
ctx := &DistSenderContext{
RPCSend: testFn,
RangeDescriptorDB: mockRangeDescriptorDB(func(_ proto.Key, _ lookupOptions) ([]proto.RangeDescriptor, error) {
return []proto.RangeDescriptor{testRangeDescriptor}, nil
}),
}
ds := NewDistSender(ctx, g)
ds.updateLeaderCache(1, leader)
put := proto.NewPut(proto.Key("a"), proto.Value{Bytes: []byte("value")}).(*proto.PutRequest)
if _, err := batchutil.SendWrapped(ds, put); err != nil && !testutils.IsError(err, "boom") {
t.Errorf("put encountered unexpected error: %s", err)
}
if cur := ds.leaderCache.Lookup(1); reflect.DeepEqual(cur, &proto.Replica{}) && !tc.shouldClearLeader {
t.Errorf("%d: leader cache eviction: shouldClearLeader=%t, but value is %v", i, tc.shouldClearLeader, cur)
}
_, cachedDesc := ds.rangeCache.getCachedRangeDescriptor(put.Key, false /* !inclusive */)
if cachedDesc == nil != tc.shouldClearReplica {
t.Errorf("%d: unexpected second replica lookup behaviour: wanted=%t", i, tc.shouldClearReplica)
}
}
}
// 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, nil, 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.GetRequest{
RequestHeader: proto.RequestHeader{Key: writes[0]},
}
get.EndKey = writes[len(writes)-1]
if _, err := batchutil.SendWrapped(tds, get); err == nil {
t.Errorf("able to call Get with a key range: %v", get)
}
var delTS proto.Timestamp
for i, k := range writes {
put := proto.NewPut(k, proto.Value{Bytes: k})
reply, err := batchutil.SendWrapped(tds, put)
if err != nil {
t.Fatal(err)
}
scan := proto.NewScan(writes[0], writes[len(writes)-1].Next(), 0).(*proto.ScanRequest)
// The Put ts may have been pushed by tsCache,
// so make sure we see their values in our Scan.
delTS = reply.(*proto.PutResponse).Timestamp
scan.Timestamp = delTS
reply, err = batchutil.SendWrapped(tds, scan)
if err != nil {
t.Fatal(err)
}
sr := reply.(*proto.ScanResponse)
if sr.Txn != nil {
// This was the other way around at some point in the past.
// Same below for Delete, etc.
t.Errorf("expected no transaction in response header")
}
if rows := sr.Rows; len(rows) != i+1 {
t.Fatalf("expected %d rows, but got %d", i+1, len(rows))
}
}
del := &proto.DeleteRangeRequest{
RequestHeader: proto.RequestHeader{
Key: writes[0],
EndKey: proto.Key(writes[len(writes)-1]).Next(),
Timestamp: delTS,
},
}
reply, err := batchutil.SendWrapped(tds, del)
if err != nil {
t.Fatal(err)
}
dr := reply.(*proto.DeleteRangeResponse)
if dr.Txn != nil {
t.Errorf("expected no transaction in response header")
}
if n := dr.NumDeleted; n != int64(len(writes)) {
t.Errorf("expected %d keys to be deleted, but got %d instead",
len(writes), n)
}
scan := proto.NewScan(writes[0], writes[len(writes)-1].Next(), 0).(*proto.ScanRequest)
scan.Timestamp = dr.Timestamp
scan.Txn = &proto.Transaction{Name: "MyTxn"}
reply, err = batchutil.SendWrapped(tds, scan)
if err != nil {
t.Fatal(err)
}
sr := reply.(*proto.ScanResponse)
if txn := sr.Txn; txn == nil || txn.Name != "MyTxn" {
t.Errorf("wanted Txn to persist, but it changed to %v", txn)
}
if rows := sr.Rows; len(rows) > 0 {
t.Fatalf("scan after delete returned rows: %v", rows)
}
}