// ProcessTaskOverNetwork is used to process the query and get the result from
// the instance which stores posting list corresponding to the predicate in the
// query.
func ProcessTaskOverNetwork(ctx context.Context, q *task.Query) (*task.Result, error) {
attr := q.Attr
gid := group.BelongsTo(attr)
x.Trace(ctx, "attr: %v groupId: %v", attr, gid)
if groups().ServesGroup(gid) {
// No need for a network call, as this should be run from within this instance.
return processTask(q)
}
// Send this over the network.
// TODO: Send the request to multiple servers as described in Jeff Dean's talk.
addr := groups().AnyServer(gid)
pl := pools().get(addr)
conn, err := pl.Get()
if err != nil {
return &emptyResult, x.Wrapf(err, "ProcessTaskOverNetwork: while retrieving connection.")
}
defer pl.Put(conn)
x.Trace(ctx, "Sending request to %v", addr)
c := NewWorkerClient(conn)
reply, err := c.ServeTask(ctx, q)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while calling Worker.ServeTask"))
return &emptyResult, err
}
x.Trace(ctx, "Reply from server. length: %v Addr: %v Attr: %v",
len(reply.UidMatrix), addr, attr)
return reply, nil
}
func BackupOverNetwork(ctx context.Context) error {
// If we haven't even had a single membership update, don't run backup.
if len(*peer) > 0 && groups().LastUpdate() == 0 {
x.Trace(ctx, "This server hasn't yet been fully initiated. Please retry later.")
return x.Errorf("Uninitiated server. Please retry later")
}
// Let's first collect all groups.
gids := groups().KnownGroups()
ch := make(chan *BackupPayload, len(gids))
for _, gid := range gids {
go func(group uint32) {
reqId := uint64(rand.Int63())
ch <- handleBackupForGroup(ctx, reqId, group)
}(gid)
}
for i := 0; i < len(gids); i++ {
bp := <-ch
if bp.Status != BackupPayload_SUCCESS {
x.Trace(ctx, "Backup status: %v for group id: %d", bp.Status, bp.GroupId)
return fmt.Errorf("Backup status: %v for group id: %d", bp.Status, bp.GroupId)
} else {
x.Trace(ctx, "Backup successful for group: %v", bp.GroupId)
}
}
x.Trace(ctx, "DONE backup")
return nil
}
// writeBatch performs a batch write of key value pairs to RocksDB.
func writeBatch(ctx context.Context, kv chan *task.KV, che chan error) {
wb := pstore.NewWriteBatch()
defer wb.Destroy()
batchSize := 0
batchWriteNum := 1
for i := range kv {
wb.Put(i.Key, i.Val)
batchSize += len(i.Key) + len(i.Val)
// We write in batches of size 32MB.
if batchSize >= 32*MB {
x.Trace(ctx, "Doing batch write %d.", batchWriteNum)
if err := pstore.WriteBatch(wb); err != nil {
che <- err
return
}
batchWriteNum++
// Resetting batch size after a batch write.
batchSize = 0
// Since we are writing data in batches, we need to clear up items enqueued
// for batch write after every successful write.
wb.Clear()
}
}
// After channel is closed the above loop would exit, we write the data in
// write batch here.
if batchSize > 0 {
x.Trace(ctx, "Doing batch write %d.", batchWriteNum)
che <- pstore.WriteBatch(wb)
return
}
che <- nil
}
// PopulateShard gets data for predicate pred from server with id serverId and
// writes it to RocksDB.
func populateShard(ctx context.Context, pl *pool, group uint32) (int, error) {
gkeys, err := generateGroup(group)
if err != nil {
return 0, x.Wrapf(err, "While generating keys group")
}
conn, err := pl.Get()
if err != nil {
return 0, err
}
defer pl.Put(conn)
c := NewWorkerClient(conn)
stream, err := c.PredicateData(context.Background(), gkeys)
if err != nil {
return 0, err
}
x.Trace(ctx, "Streaming data for group: %v", group)
kvs := make(chan *task.KV, 1000)
che := make(chan error)
go writeBatch(ctx, kvs, che)
// We can use count to check the number of posting lists returned in tests.
count := 0
for {
kv, err := stream.Recv()
if err == io.EOF {
break
}
if err != nil {
close(kvs)
return count, err
}
count++
// We check for errors, if there are no errors we send value to channel.
select {
case kvs <- kv:
// OK
case <-ctx.Done():
x.TraceError(ctx, x.Errorf("Context timed out while streaming group: %v", group))
close(kvs)
return count, ctx.Err()
case err := <-che:
x.TraceError(ctx, x.Errorf("Error while doing a batch write for group: %v", group))
close(kvs)
return count, err
}
}
close(kvs)
if err := <-che; err != nil {
x.TraceError(ctx, x.Errorf("Error while doing a batch write for group: %v", group))
return count, err
}
x.Trace(ctx, "Streaming complete for group: %v", group)
return count, nil
}
func (l *List) CommitIfDirty(ctx context.Context) (committed bool, err error) {
if atomic.LoadInt64(&l.dirtyTs) == 0 {
x.Trace(ctx, "Not committing")
return false, nil
}
x.Trace(ctx, "Committing")
return l.commit()
}
func convertToNQuad(ctx context.Context, mutation string) ([]rdf.NQuad, error) {
var nquads []rdf.NQuad
r := strings.NewReader(mutation)
reader := bufio.NewReader(r)
x.Trace(ctx, "Converting to NQuad")
var strBuf bytes.Buffer
var err error
for {
err = x.ReadLine(reader, &strBuf)
if err != nil {
break
}
ln := strings.Trim(strBuf.String(), " \t")
if len(ln) == 0 {
continue
}
nq, err := rdf.Parse(ln)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while parsing RDF"))
return nquads, err
}
nquads = append(nquads, nq)
}
if err != io.EOF {
return nquads, err
}
return nquads, nil
}
// ServeTask is used to respond to a query.
func (w *grpcWorker) ServeTask(ctx context.Context, q *task.Query) (*task.Result, error) {
if ctx.Err() != nil {
return &emptyResult, ctx.Err()
}
gid := group.BelongsTo(q.Attr)
x.Trace(ctx, "Attribute: %q NumUids: %v groupId: %v ServeTask", q.Attr, len(q.Uids), gid)
var reply *task.Result
x.AssertTruef(groups().ServesGroup(gid),
"attr: %q groupId: %v Request sent to wrong server.", q.Attr, gid)
c := make(chan error, 1)
go func() {
var err error
reply, err = processTask(q)
c <- err
}()
select {
case <-ctx.Done():
return reply, ctx.Err()
case err := <-c:
return reply, err
}
}
func (n *node) ProposeAndWait(ctx context.Context, proposal *task.Proposal) error {
if n.raft == nil {
return x.Errorf("RAFT isn't initialized yet")
}
proposal.Id = rand.Uint32()
slice := slicePool.Get().([]byte)
if len(slice) < proposal.Size() {
slice = make([]byte, proposal.Size())
}
defer slicePool.Put(slice)
upto, err := proposal.MarshalTo(slice)
if err != nil {
return err
}
proposalData := slice[:upto]
che := make(chan error, 1)
n.props.Store(proposal.Id, che)
if err = n.raft.Propose(ctx, proposalData); err != nil {
return x.Wrapf(err, "While proposing")
}
// Wait for the proposal to be committed.
if proposal.Mutations != nil {
x.Trace(ctx, "Waiting for the proposal: mutations.")
} else {
x.Trace(ctx, "Waiting for the proposal: membership update.")
}
select {
case err = <-che:
x.TraceError(ctx, err)
return err
case <-ctx.Done():
return ctx.Err()
}
}
// SortOverNetwork sends sort query over the network.
func SortOverNetwork(ctx context.Context, q *task.Sort) (*task.SortResult, error) {
gid := group.BelongsTo(q.Attr)
x.Trace(ctx, "worker.Sort attr: %v groupId: %v", q.Attr, gid)
if groups().ServesGroup(gid) {
// No need for a network call, as this should be run from within this instance.
return processSort(q)
}
// Send this over the network.
// TODO: Send the request to multiple servers as described in Jeff Dean's talk.
addr := groups().AnyServer(gid)
pl := pools().get(addr)
conn, err := pl.Get()
if err != nil {
return &emptySortResult, x.Wrapf(err, "SortOverNetwork: while retrieving connection.")
}
defer pl.Put(conn)
x.Trace(ctx, "Sending request to %v", addr)
c := NewWorkerClient(conn)
var reply *task.SortResult
cerr := make(chan error, 1)
go func() {
var err error
reply, err = c.Sort(ctx, q)
cerr <- err
}()
select {
case <-ctx.Done():
return &emptySortResult, ctx.Err()
case err := <-cerr:
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while calling Worker.Sort"))
}
return reply, err
}
}
// AddMutation adds mutation to mutation layers. Note that it does not write
// anything to disk. Some other background routine will be responsible for merging
// changes in mutation layers to RocksDB. Returns whether any mutation happens.
func (l *List) AddMutation(ctx context.Context, t *task.DirectedEdge, op uint32) (bool, error) {
if atomic.LoadInt32(&l.deleteMe) == 1 {
x.TraceError(ctx, x.Errorf("DELETEME set to true. Temporary error."))
return false, ErrRetry
}
x.Trace(ctx, "AddMutation called.")
defer x.Trace(ctx, "AddMutation done.")
// All edges with a value set, have the same uid. In other words,
// an (entity, attribute) can only have one value.
if !bytes.Equal(t.Value, nil) {
t.ValueId = math.MaxUint64
}
if t.ValueId == 0 {
err := x.Errorf("ValueId cannot be zero")
x.TraceError(ctx, err)
return false, err
}
mpost := newPosting(t, op)
// Mutation arrives:
// - Check if we had any(SET/DEL) before this, stored in the mutation list.
// - If yes, then replace that mutation. Jump to a)
// a) check if the entity exists in main posting list.
// - If yes, store the mutation.
// - If no, disregard this mutation.
l.Lock()
defer l.Unlock()
hasMutated := l.updateMutationLayer(mpost)
if len(l.mlayer) > 0 {
atomic.StoreInt64(&l.dirtyTs, time.Now().UnixNano())
if dirtyChan != nil {
dirtyChan <- l.ghash
}
}
return hasMutated, nil
}
func (n *node) Run() {
firstRun := true
ticker := time.NewTicker(time.Second)
for {
select {
case <-ticker.C:
n.raft.Tick()
case rd := <-n.raft.Ready():
x.Check(n.wal.Store(n.gid, rd.Snapshot, rd.HardState, rd.Entries))
n.saveToStorage(rd.Snapshot, rd.HardState, rd.Entries)
rcBytes, err := n.raftContext.Marshal()
for _, msg := range rd.Messages {
// NOTE: We can do some optimizations here to drop messages.
x.Check(err)
msg.Context = rcBytes
n.send(msg)
}
if !raft.IsEmptySnap(rd.Snapshot) {
n.processSnapshot(rd.Snapshot)
}
if len(rd.CommittedEntries) > 0 {
x.Trace(n.ctx, "Found %d committed entries", len(rd.CommittedEntries))
}
for _, entry := range rd.CommittedEntries {
// Just queue up to be processed. Don't wait on them.
n.commitCh <- entry
}
n.raft.Advance()
if firstRun && n.canCampaign {
go n.raft.Campaign(n.ctx)
firstRun = false
}
case <-n.done:
return
}
}
}
// newGraph returns the SubGraph and its task query.
func newGraph(ctx context.Context, gq *gql.GraphQuery) (*SubGraph, error) {
euid, exid := gq.UID, gq.XID
// This would set the Result field in SubGraph,
// and populate the children for attributes.
if len(exid) > 0 {
x.AssertTruef(!strings.HasPrefix(exid, "_new_:"), "Query shouldn't contain _new_")
euid = farm.Fingerprint64([]byte(exid))
x.Trace(ctx, "Xid: %v Uid: %v", exid, euid)
}
if euid == 0 && gq.Func == nil {
err := x.Errorf("Invalid query, query internal id is zero and generator is nil")
x.TraceError(ctx, err)
return nil, err
}
// For the root, the name to be used in result is stored in Alias, not Attr.
// The attr at root (if present) would stand for the source functions attr.
args := params{
AttrType: schema.TypeOf(gq.Alias),
isDebug: gq.Alias == "debug",
Alias: gq.Alias,
}
sg := &SubGraph{
Params: args,
}
if gq.Func != nil {
sg.Attr = gq.Func.Attr
sg.SrcFunc = append(sg.SrcFunc, gq.Func.Name)
sg.SrcFunc = append(sg.SrcFunc, gq.Func.Args...)
}
if euid > 0 {
// euid is the root UID.
sg.SrcUIDs = &task.List{Uids: []uint64{euid}}
sg.uidMatrix = []*task.List{&task.List{Uids: []uint64{euid}}}
}
sg.values = createNilValuesList(1)
return sg, nil
}
// This method is used to execute the query and return the response to the
// client as a protocol buffer message.
func (s *grpcServer) Query(ctx context.Context,
req *graph.Request) (*graph.Response, error) {
var allocIds map[string]uint64
if rand.Float64() < *tracing {
tr := trace.New("Dgraph", "GrpcQuery")
defer tr.Finish()
ctx = trace.NewContext(ctx, tr)
}
resp := new(graph.Response)
if len(req.Query) == 0 && req.Mutation == nil {
x.TraceError(ctx, x.Errorf("Empty query and mutation."))
return resp, fmt.Errorf("Empty query and mutation.")
}
var l query.Latency
l.Start = time.Now()
x.Trace(ctx, "Query received: %v", req.Query)
gq, mu, err := gql.Parse(req.Query)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while parsing query"))
return resp, err
}
// If mutations are part of the query, we run them through the mutation handler
// same as the http client.
if mu != nil && (len(mu.Set) > 0 || len(mu.Del) > 0) {
if allocIds, err = mutationHandler(ctx, mu); err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while handling mutations"))
return resp, err
}
}
// If mutations are sent as part of the mutation object in the request we run
// them here.
if req.Mutation != nil && (len(req.Mutation.Set) > 0 || len(req.Mutation.Del) > 0) {
if allocIds, err = runMutations(ctx, req.Mutation); err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while handling mutations"))
return resp, err
}
}
resp.AssignedUids = allocIds
if gq == nil || (gq.UID == 0 && len(gq.XID) == 0) {
return resp, err
}
sg, err := query.ToSubGraph(ctx, gq)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while conversion to internal format"))
return resp, err
}
l.Parsing = time.Since(l.Start)
x.Trace(ctx, "Query parsed")
rch := make(chan error)
go query.ProcessGraph(ctx, sg, nil, rch)
err = <-rch
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while executing query"))
return resp, err
}
l.Processing = time.Since(l.Start) - l.Parsing
x.Trace(ctx, "Graph processed")
node, err := sg.ToProtocolBuffer(&l)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while converting to ProtocolBuffer"))
return resp, err
}
resp.N = node
gl := new(graph.Latency)
gl.Parsing, gl.Processing, gl.Pb = l.Parsing.String(), l.Processing.String(),
l.ProtocolBuffer.String()
resp.L = gl
return resp, err
}
func queryHandler(w http.ResponseWriter, r *http.Request) {
// Add a limit on how many pending queries can be run in the system.
pendingQueries <- struct{}{}
defer func() { <-pendingQueries }()
addCorsHeaders(w)
if r.Method == "OPTIONS" {
return
}
if r.Method != "POST" {
x.SetStatus(w, x.ErrorInvalidMethod, "Invalid method")
return
}
ctx, cancel := context.WithTimeout(context.Background(), time.Minute)
defer cancel()
if rand.Float64() < *tracing {
tr := trace.New("Dgraph", "Query")
defer tr.Finish()
ctx = trace.NewContext(ctx, tr)
}
var l query.Latency
l.Start = time.Now()
defer r.Body.Close()
req, err := ioutil.ReadAll(r.Body)
q := string(req)
if err != nil || len(q) == 0 {
x.TraceError(ctx, x.Wrapf(err, "Error while reading query"))
x.SetStatus(w, x.ErrorInvalidRequest, "Invalid request encountered.")
return
}
x.Trace(ctx, "Query received: %v", q)
gq, mu, err := gql.Parse(q)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while parsing query"))
x.SetStatus(w, x.ErrorInvalidRequest, err.Error())
return
}
var allocIds map[string]uint64
var allocIdsStr map[string]string
// If we have mutations, run them first.
if mu != nil && (len(mu.Set) > 0 || len(mu.Del) > 0) {
if allocIds, err = mutationHandler(ctx, mu); err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while handling mutations"))
x.SetStatus(w, x.Error, err.Error())
return
}
// convert the new UIDs to hex string.
allocIdsStr = make(map[string]string)
for k, v := range allocIds {
allocIdsStr[k] = fmt.Sprintf("%#x", v)
}
}
if gq == nil || (gq.UID == 0 && gq.Func == nil && len(gq.XID) == 0) {
mp := map[string]interface{}{
"code": x.ErrorOk,
"message": "Done",
"uids": allocIdsStr,
}
if js, err := json.Marshal(mp); err == nil {
w.Write(js)
} else {
x.SetStatus(w, "Error", "Unable to marshal map")
}
return
}
sg, err := query.ToSubGraph(ctx, gq)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while conversion o internal format"))
x.SetStatus(w, x.ErrorInvalidRequest, err.Error())
return
}
l.Parsing = time.Since(l.Start)
x.Trace(ctx, "Query parsed")
rch := make(chan error)
go query.ProcessGraph(ctx, sg, nil, rch)
err = <-rch
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while executing query"))
x.SetStatus(w, x.Error, err.Error())
return
}
l.Processing = time.Since(l.Start) - l.Parsing
x.Trace(ctx, "Graph processed")
if len(*dumpSubgraph) > 0 {
x.Checkf(os.MkdirAll(*dumpSubgraph, 0700), *dumpSubgraph)
s := time.Now().Format("20060102.150405.000000.gob")
filename := path.Join(*dumpSubgraph, s)
f, err := os.Create(filename)
x.Checkf(err, filename)
enc := gob.NewEncoder(f)
x.Check(enc.Encode(sg))
x.Checkf(f.Close(), filename)
}
js, err := sg.ToJSON(&l)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while converting to Json"))
x.SetStatus(w, x.Error, err.Error())
return
}
x.Trace(ctx, "Latencies: Total: %v Parsing: %v Process: %v Json: %v",
time.Since(l.Start), l.Parsing, l.Processing, l.Json)
w.Header().Set("Content-Type", "application/json")
w.Write(js)
}
// ProcessGraph processes the SubGraph instance accumulating result for the query
// from different instances. Note: taskQuery is nil for root node.
func ProcessGraph(ctx context.Context, sg, parent *SubGraph, rch chan error) {
var err error
if len(sg.Attr) == 0 {
// If we have a filter SubGraph which only contains an operator,
// it won't have any attribute to work on.
// This is to allow providing SrcUIDs to the filter children.
sg.DestUIDs = sg.SrcUIDs
} else if parent == nil && len(sg.SrcFunc) == 0 {
// I am root. I don't have any function to execute, and my
// result has been prepared for me already.
sg.DestUIDs = algo.MergeSorted(sg.uidMatrix) // Could also be = sg.SrcUIDs
} else {
taskQuery := createTaskQuery(sg)
result, err := worker.ProcessTaskOverNetwork(ctx, taskQuery)
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while processing task"))
rch <- err
return
}
sg.uidMatrix = result.UidMatrix
sg.values = result.Values
if len(sg.values) > 0 {
v := sg.values[0]
x.Trace(ctx, "Sample value for attr: %v Val: %v", sg.Attr, string(v.Val))
}
sg.counts = result.Counts
if sg.Params.DoCount && len(sg.Filters) == 0 {
// If there is a filter, we need to do more work to get the actual count.
x.Trace(ctx, "Zero uids. Only count requested")
rch <- nil
return
}
if result.IntersectDest {
sg.DestUIDs = algo.IntersectSorted(result.UidMatrix)
} else {
sg.DestUIDs = algo.MergeSorted(result.UidMatrix)
}
}
if len(sg.DestUIDs.Uids) == 0 {
// Looks like we're done here. Be careful with nil srcUIDs!
x.Trace(ctx, "Zero uids for %q. Num attr children: %v", sg.Attr, len(sg.Children))
rch <- nil
return
}
// Apply filters if any.
if len(sg.Filters) > 0 {
// Run all filters in parallel.
filterChan := make(chan error, len(sg.Filters))
for _, filter := range sg.Filters {
filter.SrcUIDs = sg.DestUIDs
go ProcessGraph(ctx, filter, sg, filterChan)
}
for _ = range sg.Filters {
select {
case err = <-filterChan:
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while processing filter task"))
rch <- err
return
}
case <-ctx.Done():
x.TraceError(ctx, x.Wrapf(ctx.Err(), "Context done before full execution"))
rch <- ctx.Err()
return
}
}
// Now apply the results from filter.
var lists []*task.List
for _, filter := range sg.Filters {
lists = append(lists, filter.DestUIDs)
}
if sg.FilterOp == "|" {
sg.DestUIDs = algo.MergeSorted(lists)
} else {
sg.DestUIDs = algo.IntersectSorted(lists)
}
}
if len(sg.Params.Order) == 0 {
// There is no ordering. Just apply pagination and return.
if err = sg.applyPagination(ctx); err != nil {
rch <- err
return
}
} else {
// We need to sort first before pagination.
if err = sg.applyOrderAndPagination(ctx); err != nil {
rch <- err
return
}
}
// Here we consider handling _count_ with filtering. We do this after
// pagination because otherwise, we need to do the count with pagination
// taken into account. For example, a PL might have only 50 entries but the
// user wants to skip 100 entries and return 10 entries. In this case, you
// should return a count of 0, not 10.
if sg.Params.DoCount {
x.AssertTrue(len(sg.Filters) > 0)
sg.counts = make([]uint32, len(sg.uidMatrix))
for i, ul := range sg.uidMatrix {
// A possible optimization is to return the size of the intersection
// without forming the intersection.
algo.IntersectWith(ul, sg.DestUIDs)
sg.counts[i] = uint32(len(ul.Uids))
}
rch <- nil
return
}
childChan := make(chan error, len(sg.Children))
for i := 0; i < len(sg.Children); i++ {
child := sg.Children[i]
child.SrcUIDs = sg.DestUIDs // Make the connection.
go ProcessGraph(ctx, child, sg, childChan)
}
// Now get all the results back.
for _ = range sg.Children {
select {
case err = <-childChan:
if err != nil {
x.TraceError(ctx, x.Wrapf(err, "Error while processing child task"))
rch <- err
return
}
case <-ctx.Done():
x.TraceError(ctx, x.Wrapf(ctx.Err(), "Context done before full execution"))
rch <- ctx.Err()
return
}
}
rch <- nil
}
func handleBackupForGroup(ctx context.Context, reqId uint64, gid uint32) *BackupPayload {
n := groups().Node(gid)
if n.AmLeader() {
x.Trace(ctx, "Leader of group: %d. Running backup.", gid)
if err := backup(gid, *backupPath); err != nil {
x.TraceError(ctx, err)
return &BackupPayload{
ReqId: reqId,
Status: BackupPayload_FAILED,
}
}
x.Trace(ctx, "Backup done for group: %d.", gid)
return &BackupPayload{
ReqId: reqId,
Status: BackupPayload_SUCCESS,
GroupId: gid,
}
}
// I'm not the leader. Relay to someone who I think is.
var addrs []string
{
// Try in order: leader of given group, any server from given group, leader of group zero.
_, addr := groups().Leader(gid)
addrs = append(addrs, addr)
addrs = append(addrs, groups().AnyServer(gid))
_, addr = groups().Leader(0)
addrs = append(addrs, addr)
}
var conn *grpc.ClientConn
for _, addr := range addrs {
pl := pools().get(addr)
var err error
conn, err = pl.Get()
if err == nil {
x.Trace(ctx, "Relaying backup request for group %d to %q", gid, pl.Addr)
defer pl.Put(conn)
break
}
x.TraceError(ctx, err)
}
// Unable to find any connection to any of these servers. This should be exceedingly rare.
// But probably not worthy of crashing the server. We can just skip the backup.
if conn == nil {
x.Trace(ctx, fmt.Sprintf("Unable to find a server to backup group: %d", gid))
return &BackupPayload{
ReqId: reqId,
Status: BackupPayload_FAILED,
GroupId: gid,
}
}
c := NewWorkerClient(conn)
nr := &BackupPayload{
ReqId: reqId,
GroupId: gid,
}
nrep, err := c.Backup(ctx, nr)
if err != nil {
x.TraceError(ctx, err)
return &BackupPayload{
ReqId: reqId,
Status: BackupPayload_FAILED,
GroupId: gid,
}
}
return nrep
}
