// StoreData writes the supplied time series data to the cockroach server.
// Stored data will be sampled at the supplied resolution.
func (db *DB) StoreData(ctx context.Context, r Resolution, data []tspb.TimeSeriesData) error {
var kvs []roachpb.KeyValue
// Process data collection: data is converted to internal format, and a key
// is generated for each internal message.
for _, d := range data {
idatas, err := d.ToInternal(r.SlabDuration(), r.SampleDuration())
if err != nil {
return err
}
for _, idata := range idatas {
var value roachpb.Value
if err := value.SetProto(&idata); err != nil {
return err
}
kvs = append(kvs, roachpb.KeyValue{
Key: MakeDataKey(d.Name, d.Source, r, idata.StartTimestampNanos),
Value: value,
})
}
}
// Send the individual internal merge requests.
b := &client.Batch{}
for _, kv := range kvs {
b.AddRawRequest(&roachpb.MergeRequest{
Span: roachpb.Span{
Key: kv.Key,
},
Value: kv.Value,
})
}
return db.db.Run(ctx, b)
}
func setAppliedIndex(
ctx context.Context,
eng engine.ReadWriter,
ms *enginepb.MVCCStats,
rangeID roachpb.RangeID,
appliedIndex,
leaseAppliedIndex uint64,
) error {
var value roachpb.Value
value.SetInt(int64(appliedIndex))
if err := engine.MVCCPut(ctx, eng, ms,
keys.RaftAppliedIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */); err != nil {
return err
}
value.SetInt(int64(leaseAppliedIndex))
return engine.MVCCPut(ctx, eng, ms,
keys.LeaseAppliedIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */)
}
// insertCPutFn is used by insertRow when conflicts should be respected.
// logValue is used for pretty printing.
func insertCPutFn(ctx context.Context, b puter, key *roachpb.Key, value *roachpb.Value) {
// TODO(dan): We want do this V(2) log everywhere in sql. Consider making a
// client.Batch wrapper instead of inlining it everywhere.
if log.V(2) {
log.InfofDepth(ctx, 1, "CPut %s -> %s", *key, value.PrettyPrint())
}
b.CPut(key, value, nil)
}
func maybeUnmarshalInline(v []byte, dest proto.Message) error {
var meta enginepb.MVCCMetadata
if err := meta.Unmarshal(v); err != nil {
return err
}
value := roachpb.Value{
RawBytes: meta.RawBytes,
}
return value.GetProto(dest)
}
func setLastIndex(
ctx context.Context, eng engine.ReadWriter, rangeID roachpb.RangeID, lastIndex uint64,
) error {
var value roachpb.Value
value.SetInt(int64(lastIndex))
return engine.MVCCPut(ctx, eng, nil, keys.RaftLastIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */)
}
func newInfo(val float64) Info {
now := timeutil.Now()
v := roachpb.Value{Timestamp: hlc.Timestamp{WallTime: now.UnixNano()}}
v.SetFloat(val)
return Info{
Value: v,
OrigStamp: now.UnixNano(),
TTLStamp: now.Add(time.Millisecond).UnixNano(),
}
}
// Create the key/value pairs for the default zone config entry.
func createDefaultZoneConfig() []roachpb.KeyValue {
var ret []roachpb.KeyValue
value := roachpb.Value{}
desc := config.DefaultZoneConfig()
if err := value.SetProto(&desc); err != nil {
log.Fatalf(context.TODO(), "could not marshal %v", desc)
}
ret = append(ret, roachpb.KeyValue{
Key: MakeZoneKey(keys.RootNamespaceID),
Value: value,
})
return ret
}
func tryMeta(kv engine.MVCCKeyValue) (string, error) {
if !bytes.HasPrefix(kv.Key.Key, keys.Meta1Prefix) && !bytes.HasPrefix(kv.Key.Key, keys.Meta2Prefix) {
return "", errors.New("not a meta key")
}
value := roachpb.Value{
Timestamp: kv.Key.Timestamp,
RawBytes: kv.Value,
}
var desc roachpb.RangeDescriptor
if err := value.GetProto(&desc); err != nil {
return "", err
}
return descStr(desc), nil
}
// BenchmarkMVCCMergeTimeSeries computes performance of merging time series data.
// Uses an in-memory engine.
func BenchmarkMVCCMergeTimeSeries_RocksDB(b *testing.B) {
ts := &roachpb.InternalTimeSeriesData{
StartTimestampNanos: 0,
SampleDurationNanos: 1000,
Samples: []roachpb.InternalTimeSeriesSample{
{Offset: 0, Count: 1, Sum: 5.0},
},
}
var value roachpb.Value
if err := value.SetProto(ts); err != nil {
b.Fatal(err)
}
runMVCCMerge(setupMVCCInMemRocksDB, &value, 1024, b)
}
func setFrozenStatus(
ctx context.Context,
eng engine.ReadWriter,
ms *enginepb.MVCCStats,
rangeID roachpb.RangeID,
frozen storagebase.ReplicaState_FrozenEnum,
) error {
if frozen == storagebase.ReplicaState_FROZEN_UNSPECIFIED {
return errors.New("cannot persist unspecified FrozenStatus")
}
var val roachpb.Value
val.SetBool(frozen == storagebase.ReplicaState_FROZEN)
return engine.MVCCPut(ctx, eng, ms,
keys.RangeFrozenStatusKey(rangeID), hlc.ZeroTimestamp, val, nil)
}
// storeGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) storeGossipUpdate(_ string, content roachpb.Value) {
var storeDesc roachpb.StoreDescriptor
if err := content.GetProto(&storeDesc); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
// Does this storeDetail exist yet?
detail := sp.getStoreDetailLocked(storeDesc.StoreID)
detail.markAlive(sp.clock.Now(), &storeDesc)
sp.mu.queue.enqueue(detail)
}
// storeGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) storeGossipUpdate(_ string, content roachpb.Value) {
var storeDesc roachpb.StoreDescriptor
if err := content.GetProto(&storeDesc); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
detail := sp.getStoreDetailLocked(storeDesc.StoreID)
detail.desc = &storeDesc
detail.lastUpdatedTime = sp.clock.PhysicalTime()
sp.mu.nodeLocalities[storeDesc.Node.NodeID] = storeDesc.Node.Locality
}
// append the given entries to the raft log. Takes the previous values of
// r.mu.lastIndex and r.mu.raftLogSize, and returns new values. We do this
// rather than modifying them directly because these modifications need to be
// atomic with the commit of the batch.
func (r *Replica) append(
ctx context.Context,
batch engine.ReadWriter,
prevLastIndex uint64,
prevRaftLogSize int64,
entries []raftpb.Entry,
) (uint64, int64, error) {
if len(entries) == 0 {
return prevLastIndex, prevRaftLogSize, nil
}
var diff enginepb.MVCCStats
var value roachpb.Value
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := value.SetProto(ent); err != nil {
return 0, 0, err
}
value.InitChecksum(key)
var err error
if ent.Index > prevLastIndex {
err = engine.MVCCBlindPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
} else {
err = engine.MVCCPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
}
if err != nil {
return 0, 0, err
}
}
// Delete any previously appended log entries which never committed.
lastIndex := entries[len(entries)-1].Index
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(ctx, batch, &diff, keys.RaftLogKey(r.RangeID, i),
hlc.ZeroTimestamp, nil /* txn */)
if err != nil {
return 0, 0, err
}
}
if err := setLastIndex(ctx, batch, r.RangeID, lastIndex); err != nil {
return 0, 0, err
}
raftLogSize := prevRaftLogSize + diff.SysBytes
return lastIndex, raftLogSize, nil
}
// deadReplicasGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) deadReplicasGossipUpdate(_ string, content roachpb.Value) {
var replicas roachpb.StoreDeadReplicas
if err := content.GetProto(&replicas); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
detail := sp.getStoreDetailLocked(replicas.StoreID)
deadReplicas := make(map[roachpb.RangeID][]roachpb.ReplicaDescriptor)
for _, r := range replicas.Replicas {
deadReplicas[r.RangeID] = append(deadReplicas[r.RangeID], r.Replica)
}
detail.deadReplicas = deadReplicas
}
// Benchmark batch time series merge operations. This benchmark does not
// perform any reads and is only used to measure the cost of the periodic time
// series updates.
func runMVCCBatchTimeSeries(emk engineMaker, batchSize int, b *testing.B) {
// Precompute keys so we don't waste time formatting them at each iteration.
numKeys := batchSize
keys := make([]roachpb.Key, numKeys)
for i := 0; i < numKeys; i++ {
keys[i] = roachpb.Key(fmt.Sprintf("key-%d", i))
}
// We always write the same time series data (containing a single unchanging
// sample). This isn't realistic but is fine because we're never reading the
// data.
var value roachpb.Value
if err := value.SetProto(&roachpb.InternalTimeSeriesData{
StartTimestampNanos: 0,
SampleDurationNanos: 1000,
Samples: []roachpb.InternalTimeSeriesSample{
{Offset: 0, Count: 1, Sum: 5.0},
},
}); err != nil {
b.Fatal(err)
}
eng := emk(b, fmt.Sprintf("batch_merge_%d", batchSize))
defer eng.Close()
b.ResetTimer()
ts := hlc.Timestamp{}
for i := 0; i < b.N; i++ {
batch := eng.NewBatch()
for j := 0; j < batchSize; j++ {
ts.Logical++
if err := MVCCMerge(context.Background(), batch, nil, keys[j], ts, value); err != nil {
b.Fatalf("failed put: %s", err)
}
}
if err := batch.Commit(); err != nil {
b.Fatal(err)
}
batch.Close()
}
b.StopTimer()
}
// livenessGossipUpdate is the gossip callback used to keep the
// in-memory liveness info up to date.
func (nl *NodeLiveness) livenessGossipUpdate(key string, content roachpb.Value) {
var liveness Liveness
if err := content.GetProto(&liveness); err != nil {
log.Error(context.TODO(), err)
return
}
// If there's an existing liveness record, only update the received
// timestamp if this is our first receipt of this node's liveness
// or if the expiration or epoch was advanced.
nl.mu.Lock()
defer nl.mu.Unlock()
exLiveness, ok := nl.mu.nodes[liveness.NodeID]
if !ok || exLiveness.Expiration.Less(liveness.Expiration) || exLiveness.Epoch < liveness.Epoch {
nl.mu.nodes[liveness.NodeID] = liveness
}
}
// MergeInternalTimeSeriesData exports the engine's C++ merge logic for
// InternalTimeSeriesData to higher level packages. This is intended primarily
// for consumption by high level testing of time series functionality.
func MergeInternalTimeSeriesData(
sources ...roachpb.InternalTimeSeriesData,
) (roachpb.InternalTimeSeriesData, error) {
// Wrap each proto in an inlined MVCC value, and marshal each wrapped value
// to bytes. This is the format required by the engine.
srcBytes := make([][]byte, 0, len(sources))
for _, src := range sources {
var val roachpb.Value
if err := val.SetProto(&src); err != nil {
return roachpb.InternalTimeSeriesData{}, err
}
bytes, err := protoutil.Marshal(&enginepb.MVCCMetadata{
RawBytes: val.RawBytes,
})
if err != nil {
return roachpb.InternalTimeSeriesData{}, err
}
srcBytes = append(srcBytes, bytes)
}
// Merge every element into a nil byte slice, one at a time.
var (
mergedBytes []byte
err error
)
for _, bytes := range srcBytes {
mergedBytes, err = goMerge(mergedBytes, bytes)
if err != nil {
return roachpb.InternalTimeSeriesData{}, err
}
}
// Unmarshal merged bytes and extract the time series value within.
var meta enginepb.MVCCMetadata
if err := proto.Unmarshal(mergedBytes, &meta); err != nil {
return roachpb.InternalTimeSeriesData{}, err
}
mergedTS, err := MakeValue(meta).GetTimeseries()
if err != nil {
return roachpb.InternalTimeSeriesData{}, err
}
return mergedTS, nil
}
// updateNodeAddress is a gossip callback which fires with each
// update to the node address. This allows us to compute the
// total size of the gossip network (for determining max peers
// each gossip node is allowed to have), as well as to create
// new resolvers for each encountered host and to write the
// set of gossip node addresses to persistent storage when it
// changes.
func (g *Gossip) updateNodeAddress(_ string, content roachpb.Value) {
ctx := g.AnnotateCtx(context.TODO())
var desc roachpb.NodeDescriptor
if err := content.GetProto(&desc); err != nil {
log.Error(ctx, err)
return
}
g.mu.Lock()
defer g.mu.Unlock()
// Skip if the node has already been seen.
if _, ok := g.nodeDescs[desc.NodeID]; ok {
return
}
g.nodeDescs[desc.NodeID] = &desc
// Recompute max peers based on size of network and set the max
// sizes for incoming and outgoing node sets.
maxPeers := g.maxPeers(len(g.nodeDescs))
g.mu.incoming.setMaxSize(maxPeers)
g.outgoing.setMaxSize(maxPeers)
// Skip if it's our own address.
if desc.Address == g.mu.is.NodeAddr {
return
}
// Add this new node address (if it's not already there) to our list
// of resolvers so we can keep connecting to gossip if the original
// resolvers go offline.
g.maybeAddResolver(desc.Address)
// Add new address (if it's not already there) to bootstrap info and
// persist if possible.
if g.storage != nil && g.maybeAddBootstrapAddress(desc.Address) {
if err := g.storage.WriteBootstrapInfo(&g.bootstrapInfo); err != nil {
log.Error(ctx, err)
}
}
}
// MigrateZoneConfig migrates the legacy ZoneConfig format into the new one.
func MigrateZoneConfig(value *roachpb.Value) (ZoneConfig, error) {
var zone ZoneConfig
if err := value.GetProto(&zone); err != nil {
return ZoneConfig{}, err
}
if len(zone.ReplicaAttrs) > 0 {
if zone.NumReplicas > 0 || len(zone.Constraints.Constraints) > 0 {
return ZoneConfig{}, errors.New("migration to new ZoneConfig failed due to previous partial upgrade")
}
zone.NumReplicas = int32(len(zone.ReplicaAttrs))
if zone.NumReplicas > 0 {
attrs := zone.ReplicaAttrs[0].Attrs
zone.Constraints.Constraints = make([]Constraint, len(attrs))
for i, attr := range attrs {
zone.Constraints.Constraints[i].Value = attr
}
}
zone.ReplicaAttrs = nil
}
return zone, nil
}
// updateSystemConfig is the raw gossip info callback.
// Unmarshal the system config, and if successfully, update out
// copy and run the callbacks.
func (g *Gossip) updateSystemConfig(key string, content roachpb.Value) {
ctx := g.AnnotateCtx(context.TODO())
if key != KeySystemConfig {
log.Fatalf(ctx, "wrong key received on SystemConfig callback: %s", key)
}
cfg := config.SystemConfig{}
if err := content.GetProto(&cfg); err != nil {
log.Errorf(ctx, "could not unmarshal system config on callback: %s", err)
return
}
g.systemConfigMu.Lock()
defer g.systemConfigMu.Unlock()
g.systemConfig = cfg
g.systemConfigSet = true
for _, c := range g.systemConfigChannels {
select {
case c <- struct{}{}:
default:
}
}
}
func (tm *testModel) storeInModel(r Resolution, data tspb.TimeSeriesData) {
// Note the source, used to construct keys for model queries.
tm.seenSources[data.Source] = struct{}{}
// Process and store data in the model.
internalData, err := data.ToInternal(r.SlabDuration(), r.SampleDuration())
if err != nil {
tm.t.Fatalf("test could not convert time series to internal format: %s", err.Error())
}
for _, idata := range internalData {
key := MakeDataKey(data.Name, data.Source, r, idata.StartTimestampNanos)
keyStr := string(key)
existing, ok := tm.modelData[keyStr]
var newTs roachpb.InternalTimeSeriesData
if ok {
existingTs, err := existing.GetTimeseries()
if err != nil {
tm.t.Fatalf("test could not extract time series from existing model value: %s", err.Error())
}
newTs, err = engine.MergeInternalTimeSeriesData(existingTs, idata)
if err != nil {
tm.t.Fatalf("test could not merge time series into model value: %s", err.Error())
}
} else {
newTs, err = engine.MergeInternalTimeSeriesData(idata)
if err != nil {
tm.t.Fatalf("test could not merge time series into model value: %s", err.Error())
}
}
var val roachpb.Value
if err := val.SetProto(&newTs); err != nil {
tm.t.Fatal(err)
}
tm.modelData[keyStr] = val
}
}
func timeSeriesAsValue(start int64, duration int64, samples ...tsSample) roachpb.Value {
ts := &roachpb.InternalTimeSeriesData{
StartTimestampNanos: start,
SampleDurationNanos: duration,
}
for _, sample := range samples {
newSample := roachpb.InternalTimeSeriesSample{
Offset: sample.offset,
Count: sample.count,
Sum: sample.sum,
}
if sample.count > 1 {
newSample.Max = proto.Float64(sample.max)
newSample.Min = proto.Float64(sample.min)
}
ts.Samples = append(ts.Samples, newSample)
}
var v roachpb.Value
if err := v.SetProto(ts); err != nil {
panic(err)
}
return v
}
// GetInitialValues returns the set of initial K/V values which should be added to
// a bootstrapping CockroachDB cluster in order to create the tables contained
// in the schema.
func (ms MetadataSchema) GetInitialValues() []roachpb.KeyValue {
var ret []roachpb.KeyValue
// Save the ID generator value, which will generate descriptor IDs for user
// objects.
value := roachpb.Value{}
value.SetInt(int64(keys.MaxReservedDescID + 1))
ret = append(ret, roachpb.KeyValue{
Key: keys.DescIDGenerator,
Value: value,
})
// addDescriptor generates the needed KeyValue objects to install a
// descriptor on a new cluster.
addDescriptor := func(parentID ID, desc DescriptorProto) {
// Create name metadata key.
value := roachpb.Value{}
value.SetInt(int64(desc.GetID()))
ret = append(ret, roachpb.KeyValue{
Key: MakeNameMetadataKey(parentID, desc.GetName()),
Value: value,
})
// Create descriptor metadata key.
value = roachpb.Value{}
wrappedDesc := WrapDescriptor(desc)
if err := value.SetProto(wrappedDesc); err != nil {
log.Fatalf(context.TODO(), "could not marshal %v", desc)
}
ret = append(ret, roachpb.KeyValue{
Key: MakeDescMetadataKey(desc.GetID()),
Value: value,
})
}
// Generate initial values for system databases and tables, which have
// static descriptors that were generated elsewhere.
for _, sysObj := range ms.descs {
addDescriptor(sysObj.parentID, sysObj.desc)
}
// Other key/value generation that doesn't fit into databases and
// tables. This can be used to add initial entries to a table.
ret = append(ret, ms.otherKV...)
// Sort returned key values; this is valuable because it matches the way the
// objects would be sorted if read from the engine.
sort.Sort(roachpb.KeyValueByKey(ret))
return ret
}
func TestMigrateZoneConfig(t *testing.T) {
defer leaktest.AfterTest(t)()
testCases := []struct {
input, want proto.Message
}{
{
&ZoneConfig{
ReplicaAttrs: []roachpb.Attributes{
{Attrs: []string{"foo"}},
{},
{},
},
},
&ZoneConfig{
NumReplicas: 3,
Constraints: Constraints{
Constraints: []Constraint{
{
Type: Constraint_POSITIVE,
Value: "foo",
},
},
},
},
},
{
&ZoneConfig{
NumReplicas: 3,
Constraints: Constraints{
Constraints: []Constraint{
{
Type: Constraint_POSITIVE,
Value: "foo",
},
},
},
},
&ZoneConfig{
NumReplicas: 3,
Constraints: Constraints{
Constraints: []Constraint{
{
Type: Constraint_POSITIVE,
Value: "foo",
},
},
},
},
},
}
for i, tc := range testCases {
var val roachpb.Value
if err := val.SetProto(tc.input); err != nil {
t.Fatal(err)
}
out, err := MigrateZoneConfig(&val)
if err != nil {
t.Fatal(err)
}
if !proto.Equal(tc.want, &out) {
t.Errorf("%d: MigrateZoneConfig(%+v) = %+v; not %+v", i, tc.input, out, tc.want)
}
}
}
func getProtoValue(data []byte, msg proto.Message) error {
value := roachpb.Value{
RawBytes: data,
}
return value.GetProto(msg)
}
func tryRangeIDKey(kv engine.MVCCKeyValue) (string, error) {
if kv.Key.Timestamp != hlc.ZeroTimestamp {
return "", fmt.Errorf("range ID keys shouldn't have timestamps: %s", kv.Key)
}
_, _, suffix, _, err := keys.DecodeRangeIDKey(kv.Key.Key)
if err != nil {
return "", err
}
// All range ID keys are stored inline on the metadata.
var meta enginepb.MVCCMetadata
if err := meta.Unmarshal(kv.Value); err != nil {
return "", err
}
value := roachpb.Value{RawBytes: meta.RawBytes}
// Values encoded as protobufs set msg and continue outside the
// switch. Other types are handled inside the switch and return.
var msg proto.Message
switch {
case bytes.Equal(suffix, keys.LocalLeaseAppliedIndexSuffix):
fallthrough
case bytes.Equal(suffix, keys.LocalRaftAppliedIndexSuffix):
i, err := value.GetInt()
if err != nil {
return "", err
}
return strconv.FormatInt(i, 10), nil
case bytes.Equal(suffix, keys.LocalRangeFrozenStatusSuffix):
b, err := value.GetBool()
if err != nil {
return "", err
}
return strconv.FormatBool(b), nil
case bytes.Equal(suffix, keys.LocalAbortCacheSuffix):
msg = &roachpb.AbortCacheEntry{}
case bytes.Equal(suffix, keys.LocalRangeLastGCSuffix):
msg = &hlc.Timestamp{}
case bytes.Equal(suffix, keys.LocalRaftTombstoneSuffix):
msg = &roachpb.RaftTombstone{}
case bytes.Equal(suffix, keys.LocalRaftTruncatedStateSuffix):
msg = &roachpb.RaftTruncatedState{}
case bytes.Equal(suffix, keys.LocalRangeLeaseSuffix):
msg = &roachpb.Lease{}
case bytes.Equal(suffix, keys.LocalRangeStatsSuffix):
msg = &enginepb.MVCCStats{}
case bytes.Equal(suffix, keys.LocalRaftHardStateSuffix):
msg = &raftpb.HardState{}
case bytes.Equal(suffix, keys.LocalRaftLastIndexSuffix):
i, err := value.GetInt()
if err != nil {
return "", err
}
return strconv.FormatInt(i, 10), nil
case bytes.Equal(suffix, keys.LocalRangeLastVerificationTimestampSuffixDeprecated):
msg = &hlc.Timestamp{}
case bytes.Equal(suffix, keys.LocalRangeLastReplicaGCTimestampSuffix):
msg = &hlc.Timestamp{}
default:
return "", fmt.Errorf("unknown raft id key %s", suffix)
}
if err := value.GetProto(msg); err != nil {
return "", err
}
return msg.String(), nil
}
func TestSystemConfigGossip(t *testing.T) {
defer leaktest.AfterTest(t)()
t.Skip("#12351")
s, _, kvDB := serverutils.StartServer(t, base.TestServerArgs{})
defer s.Stopper().Stop()
ts := s.(*TestServer)
ctx := context.TODO()
key := sqlbase.MakeDescMetadataKey(keys.MaxReservedDescID)
valAt := func(i int) *sqlbase.DatabaseDescriptor {
return &sqlbase.DatabaseDescriptor{Name: "foo", ID: sqlbase.ID(i)}
}
// Register a callback for gossip updates.
resultChan := ts.Gossip().RegisterSystemConfigChannel()
// The span gets gossiped when it first shows up.
select {
case <-resultChan:
case <-time.After(500 * time.Millisecond):
t.Fatal("did not receive gossip message")
}
// Try a plain KV write first.
if err := kvDB.Put(ctx, key, valAt(0)); err != nil {
t.Fatal(err)
}
// Now do it as part of a transaction, but without the trigger set.
if err := kvDB.Txn(ctx, func(txn *client.Txn) error {
return txn.Put(key, valAt(1))
}); err != nil {
t.Fatal(err)
}
// Gossip channel should be dormant.
// TODO(tschottdorf): This test is likely flaky. Why can't some other
// process trigger gossip? It seems that a new range lease being
// acquired will gossip a new system config since the hash changed and fail
// the test (seen in practice during some buggy WIP).
var systemConfig config.SystemConfig
select {
case <-resultChan:
systemConfig, _ = ts.gossip.GetSystemConfig()
t.Fatalf("unexpected message received on gossip channel: %v", systemConfig)
case <-time.After(50 * time.Millisecond):
}
// This time mark the transaction as having a Gossip trigger.
if err := kvDB.Txn(ctx, func(txn *client.Txn) error {
txn.SetSystemConfigTrigger()
return txn.Put(key, valAt(2))
}); err != nil {
t.Fatal(err)
}
// New system config received.
select {
case <-resultChan:
systemConfig, _ = ts.gossip.GetSystemConfig()
case <-time.After(500 * time.Millisecond):
t.Fatal("did not receive gossip message")
}
// Now check the new config.
var val *roachpb.Value
for _, kv := range systemConfig.Values {
if bytes.Equal(key, kv.Key) {
val = &kv.Value
break
}
}
if val == nil {
t.Fatal("key not found in gossiped info")
}
// Make sure the returned value is valAt(2).
got := new(sqlbase.DatabaseDescriptor)
if err := val.GetProto(got); err != nil {
t.Fatal(err)
}
if expected := valAt(2); !reflect.DeepEqual(got, expected) {
t.Fatalf("mismatch: expected %+v, got %+v", *expected, *got)
}
}
// Ingest loads some data in an sstable into an empty range. Only the keys
// between startKey and endKey are loaded. If newTableID is non-zero, every
// row's key is rewritten to be for that table.
func Ingest(
ctx context.Context,
txn *client.Txn,
path string,
checksum uint32,
startKey, endKey roachpb.Key,
newTableID sqlbase.ID,
) error {
// TODO(mjibson): An appropriate value for this should be determined. The
// current value was guessed at but appears to work well.
const batchSize = 10000
// TODO(dan): Check if the range being ingested into is empty. If newTableID
// is non-zero, it'll have to be derived from startKey and endKey.
f, err := os.Open(path)
if err != nil {
return err
}
defer f.Close()
crc := crc32.New(crc32.MakeTable(crc32.Castagnoli))
if _, err := io.Copy(crc, f); err != nil {
return nil
}
if c := crc.Sum32(); c != checksum {
return errors.Errorf("%s: checksum mismatch got %d expected %d", path, c, checksum)
}
sst, err := engine.MakeRocksDBSstFileReader()
if err != nil {
return err
}
defer sst.Close()
if err := sst.AddFile(path); err != nil {
return err
}
b := txn.NewBatch()
var v roachpb.Value
count := 0
ingestFunc := func(kv engine.MVCCKeyValue) (bool, error) {
v = roachpb.Value{RawBytes: kv.Value}
v.ClearChecksum()
if log.V(3) {
log.Infof(ctx, "Put %s %s\n", kv.Key.Key, v.PrettyPrint())
}
b.Put(kv.Key.Key, &v)
count++
if count > batchSize {
if err := txn.Run(b); err != nil {
return true, err
}
b = txn.NewBatch()
count = 0
}
return false, nil
}
if newTableID != 0 {
// MakeRekeyMVCCKeyValFunc modifies the keys, but this is safe because
// the one we get back from rocksDBIterator.Key is a copy (not a
// reference to the mmaped file.)
ingestFunc = MakeRekeyMVCCKeyValFunc(newTableID, ingestFunc)
}
startKeyMVCC, endKeyMVCC := engine.MVCCKey{Key: startKey}, engine.MVCCKey{Key: endKey}
if err := sst.Iterate(startKeyMVCC, endKeyMVCC, ingestFunc); err != nil {
return err
}
return txn.Run(b)
}
// MarshalColumnValue returns a Go primitive value equivalent of val, of the
// type expected by col. If val's type is incompatible with col, or if
// col's type is not yet implemented, an error is returned.
func MarshalColumnValue(col ColumnDescriptor, val parser.Datum) (roachpb.Value, error) {
var r roachpb.Value
if val == parser.DNull {
return r, nil
}
switch col.Type.Kind {
case ColumnType_BOOL:
if v, ok := val.(*parser.DBool); ok {
r.SetBool(bool(*v))
return r, nil
}
case ColumnType_INT:
if v, ok := val.(*parser.DInt); ok {
r.SetInt(int64(*v))
return r, nil
}
case ColumnType_FLOAT:
if v, ok := val.(*parser.DFloat); ok {
r.SetFloat(float64(*v))
return r, nil
}
case ColumnType_DECIMAL:
if v, ok := val.(*parser.DDecimal); ok {
err := r.SetDecimal(&v.Dec)
return r, err
}
case ColumnType_STRING:
if v, ok := val.(*parser.DString); ok {
r.SetString(string(*v))
return r, nil
}
case ColumnType_BYTES:
if v, ok := val.(*parser.DBytes); ok {
r.SetString(string(*v))
return r, nil
}
if v, ok := val.(*parser.DString); ok {
r.SetString(string(*v))
return r, nil
}
case ColumnType_DATE:
if v, ok := val.(*parser.DDate); ok {
r.SetInt(int64(*v))
return r, nil
}
case ColumnType_TIMESTAMP:
if v, ok := val.(*parser.DTimestamp); ok {
r.SetTime(v.Time)
return r, nil
}
case ColumnType_TIMESTAMPTZ:
if v, ok := val.(*parser.DTimestampTZ); ok {
r.SetTime(v.Time)
return r, nil
}
case ColumnType_INTERVAL:
if v, ok := val.(*parser.DInterval); ok {
err := r.SetDuration(v.Duration)
return r, err
}
default:
return r, errors.Errorf("unsupported column type: %s", col.Type.Kind)
}
return r, fmt.Errorf("value type %s doesn't match type %s of column %q",
val.ResolvedType(), col.Type.Kind, col.Name)
}
// UnmarshalColumnValue decodes the value from a key-value pair using the type
// expected by the column. An error is returned if the value's type does not
// match the column's type.
func UnmarshalColumnValue(
a *DatumAlloc, kind ColumnType_Kind, value *roachpb.Value,
) (parser.Datum, error) {
if value == nil {
return parser.DNull, nil
}
switch kind {
case ColumnType_BOOL:
v, err := value.GetBool()
if err != nil {
return nil, err
}
return parser.MakeDBool(parser.DBool(v)), nil
case ColumnType_INT:
v, err := value.GetInt()
if err != nil {
return nil, err
}
return a.NewDInt(parser.DInt(v)), nil
case ColumnType_FLOAT:
v, err := value.GetFloat()
if err != nil {
return nil, err
}
return a.NewDFloat(parser.DFloat(v)), nil
case ColumnType_DECIMAL:
v, err := value.GetDecimal()
if err != nil {
return nil, err
}
dd := a.NewDDecimal(parser.DDecimal{})
dd.Set(v)
return dd, nil
case ColumnType_STRING:
v, err := value.GetBytes()
if err != nil {
return nil, err
}
return a.NewDString(parser.DString(v)), nil
case ColumnType_BYTES:
v, err := value.GetBytes()
if err != nil {
return nil, err
}
return a.NewDBytes(parser.DBytes(v)), nil
case ColumnType_DATE:
v, err := value.GetInt()
if err != nil {
return nil, err
}
return a.NewDDate(parser.DDate(v)), nil
case ColumnType_TIMESTAMP:
v, err := value.GetTime()
if err != nil {
return nil, err
}
return a.NewDTimestamp(parser.DTimestamp{Time: v}), nil
case ColumnType_TIMESTAMPTZ:
v, err := value.GetTime()
if err != nil {
return nil, err
}
return a.NewDTimestampTZ(parser.DTimestampTZ{Time: v}), nil
case ColumnType_INTERVAL:
d, err := value.GetDuration()
if err != nil {
return nil, err
}
return a.NewDInterval(parser.DInterval{Duration: d}), nil
default:
return nil, errors.Errorf("unsupported column type: %s", kind)
}
}