x/ref/services/syncbase/vsync/responder.go - roadmap.go.syncbase - Git at Google

 // Copyright 2015 The Vanadium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.

 package vsync

 import (
 	"container/heap"
 	"sort"
 	"strings"

 	"v.io/syncbase/x/ref/services/syncbase/server/interfaces"
 	"v.io/syncbase/x/ref/services/syncbase/server/util"
 	"v.io/syncbase/x/ref/services/syncbase/server/watchable"
 	"v.io/syncbase/x/ref/services/syncbase/store"
 	"v.io/v23/context"
 	"v.io/v23/verror"
 	"v.io/x/lib/vlog"
 )

 // GetDeltas implements the responder side of the GetDeltas RPC.
 func (s *syncService) GetDeltas(ctx *context.T, call interfaces.SyncGetDeltasServerCall) error {
 	recvr := call.RecvStream()
 	for recvr.Advance() {
 		req := recvr.Value()
 		// Ignoring errors since if one Database fails for any reason,
 		// it is fine to continue to the next one. In fact, sometimes
 		// the failure might be genuine. For example, the responder is
 		// no longer part of the requested SyncGroups, or the app/db is
 		// locally deleted, or a permission change has denied access.
 		rSt := newResponderState(ctx, call, s, req)
 		rSt.sendDeltasPerDatabase(ctx)
 	}

 	// TODO(hpucha): Is there a need to call finish or some such?
 	return recvr.Err()
 }

 // responderState is state accumulated per Database by the responder during an
 // initiation round.
 type responderState struct {
 	req      interfaces.DeltaReq
 	call     interfaces.SyncGetDeltasServerCall // Stream handle for the GetDeltas RPC.
 	errState error                              // Captures the error from the first two phases of the responder.
 	sync     *syncService
 	st       store.Store // Store handle to the Database.
 	diff     genRangeVector
 	outVec   interfaces.GenVector
 }

 func newResponderState(ctx *context.T, call interfaces.SyncGetDeltasServerCall, sync *syncService, req interfaces.DeltaReq) *responderState {
 	rSt := &responderState{call: call, sync: sync, req: req}
 	return rSt
 }

 // sendDeltasPerDatabase sends to an initiator all the missing generations
 // corresponding to the prefixes requested for this Database, and a genvector
 // summarizing the knowledge transferred from the responder to the
 // initiator. This happens in three phases:
 //
 // In the first phase, the initiator is checked against the SyncGroup ACLs of
 // all the SyncGroups it is requesting, and only those prefixes that belong to
 // allowed SyncGroups are carried forward.
 //
 // In the second phase, for a given set of nested prefixes from the initiator,
 // the shortest prefix in that set is extracted. The initiator's prefix
 // genvector for this shortest prefix represents the lower bound on its
 // knowledge for the entire set of nested prefixes. This prefix genvector
 // (representing the lower bound) is diffed with all the responder prefix
 // genvectors corresponding to same or deeper prefixes compared to the initiator
 // prefix. This diff produces a bound on the missing knowledge. For example, say
 // the initiator is interested in prefixes {foo, foobar}, where each prefix is
 // associated with a prefix genvector. Since the initiator strictly has as much
 // or more knowledge for prefix "foobar" as it has for prefix "foo", "foo"'s
 // prefix genvector is chosen as the lower bound for the initiator's
 // knowledge. Similarly, say the responder has knowledge on prefixes {f,
 // foobarX, foobarY, bar}. The responder diffs the prefix genvectors for
 // prefixes f, foobarX and foobarY with the initiator's prefix genvector to
 // compute a bound on missing generations (all responder's prefixes that match
 // "foo". Note that since the responder doesn't have a prefix genvector at
 // "foo", its knowledge at "f" is applicable to "foo").
 //
 // Since the second phase outputs an aggressive calculation of missing
 // generations containing more generation entries than strictly needed by the
 // initiator, in the third phase, each missing generation is sent to the
 // initiator only if the initiator is eligible for it and is not aware of
 // it. The generations are sent to the initiator in the same order as the
 // responder learned them so that the initiator can reconstruct the DAG for the
 // objects by learning older nodes first.
 func (rSt *responderState) sendDeltasPerDatabase(ctx *context.T) error {
 	// Phase 1 of sendDeltas: Authorize the initiator and respond to the
 	// caller only for the SyncGroups that allow access.
 	rSt.authorizeAndFilterSyncGroups(ctx)

 	// Phase 2 of sendDeltas: diff contains the bound on the
 	// generations missing from the initiator per device.
 	rSt.computeDeltaBound(ctx)

 	// Phase 3 of sendDeltas: Process the diff, filtering out records that
 	// are not needed, and send the remainder on the wire ordered.
 	return rSt.filterAndSendDeltas(ctx)
 }

 // authorizeAndFilterSyncGroups authorizes the initiator against the requested
 // SyncGroups and filters the initiator's prefixes to only include those from
 // allowed SyncGroups (phase 1 of sendDeltas).
 func (rSt *responderState) authorizeAndFilterSyncGroups(ctx *context.T) {
 	rSt.st, rSt.errState = rSt.sync.getDbStore(ctx, nil, rSt.req.AppName, rSt.req.DbName)
 	if rSt.errState != nil {
 		return
 	}

 	allowedPfxs := make(map[string]struct{})
 	for sgid := range rSt.req.SgIds {
 		// Check permissions for the SyncGroup.
 		var sg *interfaces.SyncGroup
 		sg, rSt.errState = getSyncGroupById(ctx, rSt.st, sgid)
 		if rSt.errState != nil {
 			return
 		}
 		rSt.errState = authorize(ctx, rSt.call.Security(), sg)
 		if verror.ErrorID(rSt.errState) == verror.ErrNoAccess.ID {
 			continue
 		} else if rSt.errState != nil {
 			return
 		}

 		for _, p := range sg.Spec.Prefixes {
 			allowedPfxs[p] = struct{}{}
 		}
 	}

 	// Filter the initiator's prefixes to what is allowed.
 	for pfx := range rSt.req.InitVec {
 		if _, ok := allowedPfxs[pfx]; ok {
 			continue
 		}
 		allowed := false
 		for p := range allowedPfxs {
 			if strings.HasPrefix(pfx, p) {
 				allowed = true
 			}
 		}

 		if !allowed {
 			delete(rSt.req.InitVec, pfx)
 		}
 	}
 	return
 }

 // computeDeltaBound computes the bound on missing generations across all
 // requested prefixes (phase 2 of sendDeltas).
 func (rSt *responderState) computeDeltaBound(ctx *context.T) {
 	// Check error from phase 1.
 	if rSt.errState != nil {
 		return
 	}

 	if len(rSt.req.InitVec) == 0 {
 		rSt.errState = verror.New(verror.ErrInternal, ctx, "empty initiator generation vector")
 		return
 	}

 	var respVec interfaces.GenVector
 	var respGen uint64
 	respVec, respGen, rSt.errState = rSt.sync.copyDbGenInfo(ctx, rSt.req.AppName, rSt.req.DbName)
 	if rSt.errState != nil {
 		return
 	}
 	respPfxs := extractAndSortPrefixes(respVec)
 	initPfxs := extractAndSortPrefixes(rSt.req.InitVec)

 	rSt.outVec = make(interfaces.GenVector)
 	rSt.diff = make(genRangeVector)
 	pfx := initPfxs[0]

 	for _, p := range initPfxs {
 		if strings.HasPrefix(p, pfx) && p != pfx {
 			continue
 		}

 		// Process this prefix as this is the start of a new set of
 		// nested prefixes.
 		pfx = p

 		// Lower bound on initiator's knowledge for this prefix set.
 		initpgv := rSt.req.InitVec[pfx]

 		// Find the relevant responder prefixes and add the corresponding knowledge.
 		var respgv interfaces.PrefixGenVector
 		var rpStart string
 		for _, rp := range respPfxs {
 			if !strings.HasPrefix(rp, pfx) && !strings.HasPrefix(pfx, rp) {
 				// No relationship with pfx.
 				continue
 			}

 			if strings.HasPrefix(pfx, rp) {
 				// If rp is a prefix of pfx, remember it because
 				// it may be a potential starting point for the
 				// responder's knowledge. The actual starting
 				// point is the deepest prefix where rp is a
 				// prefix of pfx.
 				//
 				// Say the initiator is looking for "foo", and
 				// the responder has knowledge for "f" and "fo",
 				// the responder's starting point will be the
 				// prefix genvector for "fo". Similarly, if the
 				// responder has knowledge for "foo", the
 				// starting point will be the prefix genvector
 				// for "foo".
 				rpStart = rp
 			} else {
 				// If pfx is a prefix of rp, this knowledge must
 				// be definitely sent to the initiator. Diff the
 				// prefix genvectors to adjust the delta bound and
 				// include in outVec.
 				respgv = respVec[rp]
 				rSt.diffPrefixGenVectors(respgv, initpgv)
 				rSt.outVec[rp] = respgv
 			}
 		}

 		// Deal with the starting point.
 		if rpStart == "" {
 			// No matching prefixes for pfx were found.
 			respgv = make(interfaces.PrefixGenVector)
 			respgv[rSt.sync.id] = respGen
 		} else {
 			respgv = respVec[rpStart]
 		}
 		rSt.diffPrefixGenVectors(respgv, initpgv)
 		rSt.outVec[pfx] = respgv
 	}

 	return
 }

 // filterAndSendDeltas filters the computed delta to remove records already
 // known by the initiator, and sends the resulting records to the initiator
 // (phase 3 of sendDeltas).
 func (rSt *responderState) filterAndSendDeltas(ctx *context.T) error {
 	// Always send a start and finish response so that the initiator can
 	// move on to the next Database.
 	//
 	// TODO(hpucha): Although ok for now to call SendStream once per
 	// Database, would like to make this implementation agnostic.
 	sender := rSt.call.SendStream()
 	sender.Send(interfaces.DeltaRespStart{true})
 	defer sender.Send(interfaces.DeltaRespFinish{true})

 	// Check error from phase 2.
 	if rSt.errState != nil {
 		return rSt.errState
 	}

 	// First two phases were successful. So now on to phase 3. We now visit
 	// every log record in the generation range as obtained from phase 1 in
 	// their log order. We use a heap to incrementally sort the log records
 	// as per their position in the log.
 	//
 	// Init the min heap, one entry per device in the diff.
 	mh := make(minHeap, 0, len(rSt.diff))
 	for dev, r := range rSt.diff {
 		r.cur = r.min
 		rec, err := getNextLogRec(ctx, rSt.st, dev, r)
 		if err != nil {
 			return err
 		}
 		if rec != nil {
 			mh = append(mh, rec)
 		} else {
 			delete(rSt.diff, dev)
 		}
 	}
 	heap.Init(&mh)

 	// Process the log records in order.
 	initPfxs := extractAndSortPrefixes(rSt.req.InitVec)
 	for mh.Len() > 0 {
 		rec := heap.Pop(&mh).(*localLogRec)

 		if !filterLogRec(rec, rSt.req.InitVec, initPfxs) {
 			// Send on the wire.
 			wireRec, err := makeWireLogRec(ctx, rSt.st, rec)
 			if err != nil {
 				return err
 			}
 			sender.Send(interfaces.DeltaRespRec{*wireRec})
 		}

 		// Add a new record from the same device if not done.
 		dev := rec.Metadata.Id
 		rec, err := getNextLogRec(ctx, rSt.st, dev, rSt.diff[dev])
 		if err != nil {
 			return err
 		}
 		if rec != nil {
 			heap.Push(&mh, rec)
 		} else {
 			delete(rSt.diff, dev)
 		}
 	}

 	sender.Send(interfaces.DeltaRespRespVec{rSt.outVec})
 	return nil
 }

 // genRange represents a range of generations (min and max inclusive).
 type genRange struct {
 	min uint64
 	max uint64
 	cur uint64
 }

 type genRangeVector map[uint64]*genRange

 // diffPrefixGenVectors diffs two generation vectors, belonging to the responder
 // and the initiator, and updates the range of generations per device known to
 // the responder but not known to the initiator. "gens" (generation range) is
 // passed in as an input argument so that it can be incrementally updated as the
 // range of missing generations grows when different responder prefix genvectors
 // are used to compute the diff.
 //
 // For example: Generation vector for responder is say RVec = {A:10, B:5, C:1},
 // Generation vector for initiator is say IVec = {A:5, B:10, D:2}. Diffing these
 // two vectors returns: {A:[6-10], C:[1-1]}.
 //
 // TODO(hpucha): Add reclaimVec for GCing.
 func (rSt *responderState) diffPrefixGenVectors(respPVec, initPVec interfaces.PrefixGenVector) {
 	// Compute missing generations for devices that are in both initiator's and responder's vectors.
 	for devid, gen := range initPVec {
 		rgen, ok := respPVec[devid]
 		if ok {
 			updateDevRange(devid, rgen, gen, rSt.diff)
 		}
 	}

 	// Compute missing generations for devices not in initiator's vector but in responder's vector.
 	for devid, rgen := range respPVec {
 		if _, ok := initPVec[devid]; !ok {
 			updateDevRange(devid, rgen, 0, rSt.diff)
 		}
 	}
 }

 func updateDevRange(devid, rgen, gen uint64, gens genRangeVector) {
 	if gen < rgen {
 		// Need to include all generations in the interval [gen+1,rgen], gen+1 and rgen inclusive.
 		if r, ok := gens[devid]; !ok {
 			gens[devid] = &genRange{min: gen + 1, max: rgen}
 		} else {
 			if gen+1 < r.min {
 				r.min = gen + 1
 			}
 			if rgen > r.max {
 				r.max = rgen
 			}
 		}
 	}
 }

 func extractAndSortPrefixes(vec interfaces.GenVector) []string {
 	pfxs := make([]string, len(vec))
 	i := 0
 	for p := range vec {
 		pfxs[i] = p
 		i++
 	}
 	sort.Strings(pfxs)
 	return pfxs
 }

 // TODO(hpucha): This can be optimized using a scan instead of "gets" in a for
 // loop.
 func getNextLogRec(ctx *context.T, sn store.StoreReader, dev uint64, r *genRange) (*localLogRec, error) {
 	for i := r.cur; i <= r.max; i++ {
 		rec, err := getLogRec(ctx, sn, dev, i)
 		if err == nil {
 			r.cur = i + 1
 			return rec, nil
 		}
 		if verror.ErrorID(err) != verror.ErrNoExist.ID {
 			return nil, err
 		}
 	}
 	return nil, nil
 }

 // Note: initPfxs is sorted.
 func filterLogRec(rec *localLogRec, initVec interfaces.GenVector, initPfxs []string) bool {
 	// The key starts with one of the store's reserved prefixes for managed
 	// namespaces (e.g. $row, $perms).  Remove that prefix before comparing
 	// it with the SyncGroup prefixes which are defined by the application.
 	parts := util.SplitKeyParts(rec.Metadata.ObjId)
 	if len(parts) < 2 {
 		vlog.Fatalf("filterLogRec: invalid entry key %s", rec.Metadata.ObjId)
 	}
 	key := util.JoinKeyParts(parts[1:]...)

 	filter := true
 	var maxGen uint64
 	for _, p := range initPfxs {
 		if strings.HasPrefix(key, p) {
 			// Do not filter. Initiator is interested in this
 			// prefix.
 			filter = false

 			// Track if the initiator knows of this record.
 			gen := initVec[p][rec.Metadata.Id]
 			if maxGen < gen {
 				maxGen = gen
 			}
 		}
 	}

 	// Filter this record if the initiator already has it.
 	if maxGen >= rec.Metadata.Gen {
 		filter = true
 	}

 	return filter
 }

 // makeWireLogRec creates a sync log record to send on the wire from a given
 // local sync record.
 func makeWireLogRec(ctx *context.T, st store.Store, rec *localLogRec) (*interfaces.LogRec, error) {
 	// Get the object value at the required version.
 	key, version := rec.Metadata.ObjId, rec.Metadata.CurVers
 	value, err := watchable.GetAtVersion(ctx, st, []byte(key), nil, []byte(version))
 	if err != nil {
 		return nil, err
 	}

 	wireRec := &interfaces.LogRec{Metadata: rec.Metadata, Value: value}
 	return wireRec, nil
 }

 // A minHeap implements heap.Interface and holds local log records.
 type minHeap []*localLogRec

 func (mh minHeap) Len() int { return len(mh) }

 func (mh minHeap) Less(i, j int) bool {
 	return mh[i].Pos < mh[j].Pos
 }

 func (mh minHeap) Swap(i, j int) {
 	mh[i], mh[j] = mh[j], mh[i]
 }

 func (mh *minHeap) Push(x interface{}) {
 	item := x.(*localLogRec)
 	*mh = append(*mh, item)
 }

 func (mh *minHeap) Pop() interface{} {
 	old := *mh
 	n := len(old)
 	item := old[n-1]
 	*mh = old[0 : n-1]
 	return item
 }
	// Copyright 2015 The Vanadium Authors. All rights reserved.
	// Use of this source code is governed by a BSD-style
	// license that can be found in the LICENSE file.

	package vsync

	import (
	"container/heap"
	"sort"
	"strings"

	"v.io/syncbase/x/ref/services/syncbase/server/interfaces"
	"v.io/syncbase/x/ref/services/syncbase/server/util"
	"v.io/syncbase/x/ref/services/syncbase/server/watchable"
	"v.io/syncbase/x/ref/services/syncbase/store"
	"v.io/v23/context"
	"v.io/v23/verror"
	"v.io/x/lib/vlog"
	)

	// GetDeltas implements the responder side of the GetDeltas RPC.
	func (s syncService) GetDeltas(ctx context.T, call interfaces.SyncGetDeltasServerCall) error {
	recvr := call.RecvStream()
	for recvr.Advance() {
	req := recvr.Value()
	// Ignoring errors since if one Database fails for any reason,
	// it is fine to continue to the next one. In fact, sometimes
	// the failure might be genuine. For example, the responder is
	// no longer part of the requested SyncGroups, or the app/db is
	// locally deleted, or a permission change has denied access.
	rSt := newResponderState(ctx, call, s, req)
	rSt.sendDeltasPerDatabase(ctx)
	}

	// TODO(hpucha): Is there a need to call finish or some such?
	return recvr.Err()
	}

	// responderState is state accumulated per Database by the responder during an
	// initiation round.
	type responderState struct {
	req interfaces.DeltaReq
	call interfaces.SyncGetDeltasServerCall // Stream handle for the GetDeltas RPC.
	errState error // Captures the error from the first two phases of the responder.
	sync *syncService
	st store.Store // Store handle to the Database.
	diff genRangeVector
	outVec interfaces.GenVector
	}

	func newResponderState(ctx context.T, call interfaces.SyncGetDeltasServerCall, sync syncService, req interfaces.DeltaReq) *responderState {
	rSt := &responderState{call: call, sync: sync, req: req}
	return rSt
	}

	// sendDeltasPerDatabase sends to an initiator all the missing generations
	// corresponding to the prefixes requested for this Database, and a genvector
	// summarizing the knowledge transferred from the responder to the
	// initiator. This happens in three phases:
	//
	// In the first phase, the initiator is checked against the SyncGroup ACLs of
	// all the SyncGroups it is requesting, and only those prefixes that belong to
	// allowed SyncGroups are carried forward.
	//
	// In the second phase, for a given set of nested prefixes from the initiator,
	// the shortest prefix in that set is extracted. The initiator's prefix
	// genvector for this shortest prefix represents the lower bound on its
	// knowledge for the entire set of nested prefixes. This prefix genvector
	// (representing the lower bound) is diffed with all the responder prefix
	// genvectors corresponding to same or deeper prefixes compared to the initiator
	// prefix. This diff produces a bound on the missing knowledge. For example, say
	// the initiator is interested in prefixes {foo, foobar}, where each prefix is
	// associated with a prefix genvector. Since the initiator strictly has as much
	// or more knowledge for prefix "foobar" as it has for prefix "foo", "foo"'s
	// prefix genvector is chosen as the lower bound for the initiator's
	// knowledge. Similarly, say the responder has knowledge on prefixes {f,
	// foobarX, foobarY, bar}. The responder diffs the prefix genvectors for
	// prefixes f, foobarX and foobarY with the initiator's prefix genvector to
	// compute a bound on missing generations (all responder's prefixes that match
	// "foo". Note that since the responder doesn't have a prefix genvector at
	// "foo", its knowledge at "f" is applicable to "foo").
	//
	// Since the second phase outputs an aggressive calculation of missing
	// generations containing more generation entries than strictly needed by the
	// initiator, in the third phase, each missing generation is sent to the
	// initiator only if the initiator is eligible for it and is not aware of
	// it. The generations are sent to the initiator in the same order as the
	// responder learned them so that the initiator can reconstruct the DAG for the
	// objects by learning older nodes first.
	func (rSt responderState) sendDeltasPerDatabase(ctx context.T) error {
	// Phase 1 of sendDeltas: Authorize the initiator and respond to the
	// caller only for the SyncGroups that allow access.
	rSt.authorizeAndFilterSyncGroups(ctx)

	// Phase 2 of sendDeltas: diff contains the bound on the
	// generations missing from the initiator per device.
	rSt.computeDeltaBound(ctx)

	// Phase 3 of sendDeltas: Process the diff, filtering out records that
	// are not needed, and send the remainder on the wire ordered.
	return rSt.filterAndSendDeltas(ctx)
	}

	// authorizeAndFilterSyncGroups authorizes the initiator against the requested
	// SyncGroups and filters the initiator's prefixes to only include those from
	// allowed SyncGroups (phase 1 of sendDeltas).
	func (rSt responderState) authorizeAndFilterSyncGroups(ctx context.T) {
	rSt.st, rSt.errState = rSt.sync.getDbStore(ctx, nil, rSt.req.AppName, rSt.req.DbName)
	if rSt.errState != nil {
	return
	}

	allowedPfxs := make(map[string]struct{})
	for sgid := range rSt.req.SgIds {
	// Check permissions for the SyncGroup.
	var sg *interfaces.SyncGroup
	sg, rSt.errState = getSyncGroupById(ctx, rSt.st, sgid)
	if rSt.errState != nil {
	return
	}
	rSt.errState = authorize(ctx, rSt.call.Security(), sg)
	if verror.ErrorID(rSt.errState) == verror.ErrNoAccess.ID {
	continue
	} else if rSt.errState != nil {
	return
	}

	for _, p := range sg.Spec.Prefixes {
	allowedPfxs[p] = struct{}{}
	}
	}

	// Filter the initiator's prefixes to what is allowed.
	for pfx := range rSt.req.InitVec {
	if _, ok := allowedPfxs[pfx]; ok {
	continue
	}
	allowed := false
	for p := range allowedPfxs {
	if strings.HasPrefix(pfx, p) {
	allowed = true
	}
	}

	if !allowed {
	delete(rSt.req.InitVec, pfx)
	}
	}
	return
	}

	// computeDeltaBound computes the bound on missing generations across all
	// requested prefixes (phase 2 of sendDeltas).
	func (rSt responderState) computeDeltaBound(ctx context.T) {
	// Check error from phase 1.
	if rSt.errState != nil {
	return
	}

	if len(rSt.req.InitVec) == 0 {
	rSt.errState = verror.New(verror.ErrInternal, ctx, "empty initiator generation vector")
	return
	}

	var respVec interfaces.GenVector
	var respGen uint64
	respVec, respGen, rSt.errState = rSt.sync.copyDbGenInfo(ctx, rSt.req.AppName, rSt.req.DbName)
	if rSt.errState != nil {
	return
	}
	respPfxs := extractAndSortPrefixes(respVec)
	initPfxs := extractAndSortPrefixes(rSt.req.InitVec)

	rSt.outVec = make(interfaces.GenVector)
	rSt.diff = make(genRangeVector)
	pfx := initPfxs[0]

	for _, p := range initPfxs {
	if strings.HasPrefix(p, pfx) && p != pfx {
	continue
	}

	// Process this prefix as this is the start of a new set of
	// nested prefixes.
	pfx = p

	// Lower bound on initiator's knowledge for this prefix set.
	initpgv := rSt.req.InitVec[pfx]

	// Find the relevant responder prefixes and add the corresponding knowledge.
	var respgv interfaces.PrefixGenVector
	var rpStart string
	for _, rp := range respPfxs {
	if !strings.HasPrefix(rp, pfx) && !strings.HasPrefix(pfx, rp) {
	// No relationship with pfx.
	continue
	}

	if strings.HasPrefix(pfx, rp) {
	// If rp is a prefix of pfx, remember it because
	// it may be a potential starting point for the
	// responder's knowledge. The actual starting
	// point is the deepest prefix where rp is a
	// prefix of pfx.
	//
	// Say the initiator is looking for "foo", and
	// the responder has knowledge for "f" and "fo",
	// the responder's starting point will be the
	// prefix genvector for "fo". Similarly, if the
	// responder has knowledge for "foo", the
	// starting point will be the prefix genvector
	// for "foo".
	rpStart = rp
	} else {
	// If pfx is a prefix of rp, this knowledge must
	// be definitely sent to the initiator. Diff the
	// prefix genvectors to adjust the delta bound and
	// include in outVec.
	respgv = respVec[rp]
	rSt.diffPrefixGenVectors(respgv, initpgv)
	rSt.outVec[rp] = respgv
	}
	}

	// Deal with the starting point.
	if rpStart == "" {
	// No matching prefixes for pfx were found.
	respgv = make(interfaces.PrefixGenVector)
	respgv[rSt.sync.id] = respGen
	} else {
	respgv = respVec[rpStart]
	}
	rSt.diffPrefixGenVectors(respgv, initpgv)
	rSt.outVec[pfx] = respgv
	}

	return
	}

	// filterAndSendDeltas filters the computed delta to remove records already
	// known by the initiator, and sends the resulting records to the initiator
	// (phase 3 of sendDeltas).
	func (rSt responderState) filterAndSendDeltas(ctx context.T) error {
	// Always send a start and finish response so that the initiator can
	// move on to the next Database.
	//
	// TODO(hpucha): Although ok for now to call SendStream once per
	// Database, would like to make this implementation agnostic.
	sender := rSt.call.SendStream()
	sender.Send(interfaces.DeltaRespStart{true})
	defer sender.Send(interfaces.DeltaRespFinish{true})

	// Check error from phase 2.
	if rSt.errState != nil {
	return rSt.errState
	}

	// First two phases were successful. So now on to phase 3. We now visit
	// every log record in the generation range as obtained from phase 1 in
	// their log order. We use a heap to incrementally sort the log records
	// as per their position in the log.
	//
	// Init the min heap, one entry per device in the diff.
	mh := make(minHeap, 0, len(rSt.diff))
	for dev, r := range rSt.diff {
	r.cur = r.min
	rec, err := getNextLogRec(ctx, rSt.st, dev, r)
	if err != nil {
	return err
	}
	if rec != nil {
	mh = append(mh, rec)
	} else {
	delete(rSt.diff, dev)
	}
	}
	heap.Init(&mh)

	// Process the log records in order.
	initPfxs := extractAndSortPrefixes(rSt.req.InitVec)
	for mh.Len() > 0 {
	rec := heap.Pop(&mh).(*localLogRec)

	if !filterLogRec(rec, rSt.req.InitVec, initPfxs) {
	// Send on the wire.
	wireRec, err := makeWireLogRec(ctx, rSt.st, rec)
	if err != nil {
	return err
	}
	sender.Send(interfaces.DeltaRespRec{*wireRec})
	}

	// Add a new record from the same device if not done.
	dev := rec.Metadata.Id
	rec, err := getNextLogRec(ctx, rSt.st, dev, rSt.diff[dev])
	if err != nil {
	return err
	}
	if rec != nil {
	heap.Push(&mh, rec)
	} else {
	delete(rSt.diff, dev)
	}
	}

	sender.Send(interfaces.DeltaRespRespVec{rSt.outVec})
	return nil
	}

	// genRange represents a range of generations (min and max inclusive).
	type genRange struct {
	min uint64
	max uint64
	cur uint64
	}

	type genRangeVector map[uint64]*genRange

	// diffPrefixGenVectors diffs two generation vectors, belonging to the responder
	// and the initiator, and updates the range of generations per device known to
	// the responder but not known to the initiator. "gens" (generation range) is
	// passed in as an input argument so that it can be incrementally updated as the
	// range of missing generations grows when different responder prefix genvectors
	// are used to compute the diff.
	//
	// For example: Generation vector for responder is say RVec = {A:10, B:5, C:1},
	// Generation vector for initiator is say IVec = {A:5, B:10, D:2}. Diffing these
	// two vectors returns: {A:[6-10], C:[1-1]}.
	//
	// TODO(hpucha): Add reclaimVec for GCing.
	func (rSt *responderState) diffPrefixGenVectors(respPVec, initPVec interfaces.PrefixGenVector) {
	// Compute missing generations for devices that are in both initiator's and responder's vectors.
	for devid, gen := range initPVec {
	rgen, ok := respPVec[devid]
	if ok {
	updateDevRange(devid, rgen, gen, rSt.diff)
	}
	}

	// Compute missing generations for devices not in initiator's vector but in responder's vector.
	for devid, rgen := range respPVec {
	if _, ok := initPVec[devid]; !ok {
	updateDevRange(devid, rgen, 0, rSt.diff)
	}
	}
	}

	func updateDevRange(devid, rgen, gen uint64, gens genRangeVector) {
	if gen < rgen {
	// Need to include all generations in the interval [gen+1,rgen], gen+1 and rgen inclusive.
	if r, ok := gens[devid]; !ok {
	gens[devid] = &genRange{min: gen + 1, max: rgen}
	} else {
	if gen+1 < r.min {
	r.min = gen + 1
	}
	if rgen > r.max {
	r.max = rgen
	}
	}
	}
	}

	func extractAndSortPrefixes(vec interfaces.GenVector) []string {
	pfxs := make([]string, len(vec))
	i := 0
	for p := range vec {
	pfxs[i] = p
	i++
	}
	sort.Strings(pfxs)
	return pfxs
	}

	// TODO(hpucha): This can be optimized using a scan instead of "gets" in a for
	// loop.
	func getNextLogRec(ctx context.T, sn store.StoreReader, dev uint64, r genRange) (*localLogRec, error) {
	for i := r.cur; i <= r.max; i++ {
	rec, err := getLogRec(ctx, sn, dev, i)
	if err == nil {
	r.cur = i + 1
	return rec, nil
	}
	if verror.ErrorID(err) != verror.ErrNoExist.ID {
	return nil, err
	}
	}
	return nil, nil
	}

	// Note: initPfxs is sorted.
	func filterLogRec(rec *localLogRec, initVec interfaces.GenVector, initPfxs []string) bool {
	// The key starts with one of the store's reserved prefixes for managed
	// namespaces (e.g. $row, $perms). Remove that prefix before comparing
	// it with the SyncGroup prefixes which are defined by the application.
	parts := util.SplitKeyParts(rec.Metadata.ObjId)
	if len(parts) < 2 {
	vlog.Fatalf("filterLogRec: invalid entry key %s", rec.Metadata.ObjId)
	}
	key := util.JoinKeyParts(parts[1:]...)

	filter := true
	var maxGen uint64
	for _, p := range initPfxs {
	if strings.HasPrefix(key, p) {
	// Do not filter. Initiator is interested in this
	// prefix.
	filter = false

	// Track if the initiator knows of this record.
	gen := initVec[p][rec.Metadata.Id]
	if maxGen < gen {
	maxGen = gen
	}
	}
	}

	// Filter this record if the initiator already has it.
	if maxGen >= rec.Metadata.Gen {
	filter = true
	}

	return filter
	}

	// makeWireLogRec creates a sync log record to send on the wire from a given
	// local sync record.
	func makeWireLogRec(ctx context.T, st store.Store, rec localLogRec) (*interfaces.LogRec, error) {
	// Get the object value at the required version.
	key, version := rec.Metadata.ObjId, rec.Metadata.CurVers
	value, err := watchable.GetAtVersion(ctx, st, []byte(key), nil, []byte(version))
	if err != nil {
	return nil, err
	}

	wireRec := &interfaces.LogRec{Metadata: rec.Metadata, Value: value}
	return wireRec, nil
	}

	// A minHeap implements heap.Interface and holds local log records.
	type minHeap []*localLogRec

	func (mh minHeap) Len() int { return len(mh) }

	func (mh minHeap) Less(i, j int) bool {
	return mh[i].Pos < mh[j].Pos
	}

	func (mh minHeap) Swap(i, j int) {
	mh[i], mh[j] = mh[j], mh[i]
	}

	func (mh *minHeap) Push(x interface{}) {
	item := x.(*localLogRec)
	mh = append(mh, item)
	}

	func (mh *minHeap) Pop() interface{} {
	old := *mh
	n := len(old)
	item := old[n-1]
	*mh = old[0 : n-1]
	return item
	}