0
0
Fork 0
mirror of https://github.com/matrix-org/dendrite synced 2024-11-17 07:10:51 +01:00
dendrite/syncapi/streams/stream_pdu.go
Till eb29a31550
Optimize /sync and history visibility (#2961)
Should fix the following issues or make a lot less worse when using
Postgres:

The main issue behind #2911: The client gives up after a certain time,
causing a cascade of context errors, because the response couldn't be
built up fast enough. This mostly happens on accounts with many rooms,
due to the inefficient way we're getting recent events and current state

For #2777: The queries for getting the membership events for history
visibility were being executed for each room (I think 185?), resulting
in a whooping 2k queries for membership events. (Getting the
statesnapshot -> block nids -> actual wanted membership event)

Both should now be better by:
- Using a LATERAL join to get all recent events for all joined rooms in
one go (TODO: maybe do the same for room summary and current state etc)
- If we're lazy loading on initial syncs, we're now not getting the
whole current state, just to drop the majority of it because we're lazy
loading members - we add a filter to exclude membership events on the
first call to `CurrentState`.
- Using an optimized query to get the membership events needed to
calculate history visibility

---------

Co-authored-by: kegsay <kegan@matrix.org>
2023-02-07 14:31:23 +01:00

654 lines
22 KiB
Go

package streams
import (
"context"
"database/sql"
"fmt"
"time"
"github.com/matrix-org/dendrite/internal/caching"
roomserverAPI "github.com/matrix-org/dendrite/roomserver/api"
"github.com/matrix-org/dendrite/syncapi/internal"
"github.com/matrix-org/dendrite/syncapi/storage"
"github.com/matrix-org/dendrite/syncapi/types"
userapi "github.com/matrix-org/dendrite/userapi/api"
"github.com/matrix-org/dendrite/syncapi/notifier"
"github.com/matrix-org/gomatrixserverlib"
"github.com/sirupsen/logrus"
)
// The max number of per-room goroutines to have running.
// Too high and this will consume lots of CPU, too low and complete
// sync responses will take longer to process.
const PDU_STREAM_WORKERS = 256
// The maximum number of tasks that can be queued in total before
// backpressure will build up and the rests will start to block.
const PDU_STREAM_QUEUESIZE = PDU_STREAM_WORKERS * 8
type PDUStreamProvider struct {
DefaultStreamProvider
// userID+deviceID -> lazy loading cache
lazyLoadCache caching.LazyLoadCache
rsAPI roomserverAPI.SyncRoomserverAPI
notifier *notifier.Notifier
}
func (p *PDUStreamProvider) Setup(
ctx context.Context, snapshot storage.DatabaseTransaction,
) {
p.DefaultStreamProvider.Setup(ctx, snapshot)
p.latestMutex.Lock()
defer p.latestMutex.Unlock()
id, err := snapshot.MaxStreamPositionForPDUs(ctx)
if err != nil {
panic(err)
}
p.latest = id
}
func (p *PDUStreamProvider) CompleteSync(
ctx context.Context,
snapshot storage.DatabaseTransaction,
req *types.SyncRequest,
) types.StreamPosition {
from := types.StreamPosition(0)
to := p.LatestPosition(ctx)
// Get the current sync position which we will base the sync response on.
// For complete syncs, we want to start at the most recent events and work
// backwards, so that we show the most recent events in the room.
r := types.Range{
From: to,
To: 0,
Backwards: true,
}
// Extract room state and recent events for all rooms the user is joined to.
joinedRoomIDs, err := snapshot.RoomIDsWithMembership(ctx, req.Device.UserID, gomatrixserverlib.Join)
if err != nil {
req.Log.WithError(err).Error("p.DB.RoomIDsWithMembership failed")
return from
}
stateFilter := req.Filter.Room.State
eventFilter := req.Filter.Room.Timeline
if err = p.addIgnoredUsersToFilter(ctx, snapshot, req, &eventFilter); err != nil {
req.Log.WithError(err).Error("unable to update event filter with ignored users")
}
recentEvents, err := snapshot.RecentEvents(ctx, joinedRoomIDs, r, &eventFilter, true, true)
if err != nil {
return from
}
// Build up a /sync response. Add joined rooms.
for _, roomID := range joinedRoomIDs {
events := recentEvents[roomID]
// Invalidate the lazyLoadCache, otherwise we end up with missing displaynames/avatars
// TODO: This might be inefficient, when joined to many and/or large rooms.
joinedUsers := p.notifier.JoinedUsers(roomID)
for _, sharedUser := range joinedUsers {
p.lazyLoadCache.InvalidateLazyLoadedUser(req.Device, roomID, sharedUser)
}
// get the join response for each room
jr, jerr := p.getJoinResponseForCompleteSync(
ctx, snapshot, roomID, &stateFilter, req.WantFullState, req.Device, false,
events.Events, events.Limited,
)
if jerr != nil {
req.Log.WithError(jerr).Error("p.getJoinResponseForCompleteSync failed")
if ctxErr := req.Context.Err(); ctxErr != nil || jerr == sql.ErrTxDone {
return from
}
continue
}
req.Response.Rooms.Join[roomID] = jr
req.Rooms[roomID] = gomatrixserverlib.Join
}
// Add peeked rooms.
peeks, err := snapshot.PeeksInRange(ctx, req.Device.UserID, req.Device.ID, r)
if err != nil {
req.Log.WithError(err).Error("p.DB.PeeksInRange failed")
return from
}
if len(peeks) > 0 {
peekRooms := make([]string, 0, len(peeks))
for _, peek := range peeks {
if !peek.Deleted {
peekRooms = append(peekRooms, peek.RoomID)
}
}
recentEvents, err = snapshot.RecentEvents(ctx, peekRooms, r, &eventFilter, true, true)
if err != nil {
return from
}
for _, roomID := range peekRooms {
var jr *types.JoinResponse
events := recentEvents[roomID]
jr, err = p.getJoinResponseForCompleteSync(
ctx, snapshot, roomID, &stateFilter, req.WantFullState, req.Device, true,
events.Events, events.Limited,
)
if err != nil {
req.Log.WithError(err).Error("p.getJoinResponseForCompleteSync failed")
if err == context.DeadlineExceeded || err == context.Canceled || err == sql.ErrTxDone {
return from
}
continue
}
req.Response.Rooms.Peek[roomID] = jr
}
}
return to
}
func (p *PDUStreamProvider) IncrementalSync(
ctx context.Context,
snapshot storage.DatabaseTransaction,
req *types.SyncRequest,
from, to types.StreamPosition,
) (newPos types.StreamPosition) {
r := types.Range{
From: from,
To: to,
Backwards: from > to,
}
var err error
var stateDeltas []types.StateDelta
var syncJoinedRooms []string
stateFilter := req.Filter.Room.State
eventFilter := req.Filter.Room.Timeline
if req.WantFullState {
if stateDeltas, syncJoinedRooms, err = snapshot.GetStateDeltasForFullStateSync(ctx, req.Device, r, req.Device.UserID, &stateFilter); err != nil {
req.Log.WithError(err).Error("p.DB.GetStateDeltasForFullStateSync failed")
return from
}
} else {
if stateDeltas, syncJoinedRooms, err = snapshot.GetStateDeltas(ctx, req.Device, r, req.Device.UserID, &stateFilter); err != nil {
req.Log.WithError(err).Error("p.DB.GetStateDeltas failed")
return from
}
}
for _, roomID := range syncJoinedRooms {
req.Rooms[roomID] = gomatrixserverlib.Join
}
if len(stateDeltas) == 0 {
return to
}
if err = p.addIgnoredUsersToFilter(ctx, snapshot, req, &eventFilter); err != nil {
req.Log.WithError(err).Error("unable to update event filter with ignored users")
}
newPos = from
for _, delta := range stateDeltas {
newRange := r
// If this room was joined in this sync, try to fetch
// as much timeline events as allowed by the filter.
if delta.NewlyJoined {
// Reverse the range, so we get the most recent first.
// This will be limited by the eventFilter.
newRange = types.Range{
From: r.To,
To: 0,
Backwards: true,
}
}
var pos types.StreamPosition
if pos, err = p.addRoomDeltaToResponse(ctx, snapshot, req.Device, newRange, delta, &eventFilter, &stateFilter, req); err != nil {
req.Log.WithError(err).Error("d.addRoomDeltaToResponse failed")
if err == context.DeadlineExceeded || err == context.Canceled || err == sql.ErrTxDone {
return newPos
}
continue
}
// Reset the position, as it is only for the special case of newly joined rooms
if delta.NewlyJoined {
pos = newRange.From
}
switch {
case r.Backwards && pos < newPos:
fallthrough
case !r.Backwards && pos > newPos:
newPos = pos
}
}
return newPos
}
// Limit the recent events to X when going backwards
const recentEventBackwardsLimit = 100
// nolint:gocyclo
func (p *PDUStreamProvider) addRoomDeltaToResponse(
ctx context.Context,
snapshot storage.DatabaseTransaction,
device *userapi.Device,
r types.Range,
delta types.StateDelta,
eventFilter *gomatrixserverlib.RoomEventFilter,
stateFilter *gomatrixserverlib.StateFilter,
req *types.SyncRequest,
) (types.StreamPosition, error) {
var err error
originalLimit := eventFilter.Limit
// If we're going backwards, grep at least X events, this is mostly to satisfy Sytest
if r.Backwards && originalLimit < recentEventBackwardsLimit {
eventFilter.Limit = recentEventBackwardsLimit // TODO: Figure out a better way
diff := r.From - r.To
if diff > 0 && diff < recentEventBackwardsLimit {
eventFilter.Limit = int(diff)
}
}
dbEvents, err := snapshot.RecentEvents(
ctx, []string{delta.RoomID}, r,
eventFilter, true, true,
)
if err != nil {
if err == sql.ErrNoRows {
return r.To, nil
}
return r.From, fmt.Errorf("p.DB.RecentEvents: %w", err)
}
recentStreamEvents := dbEvents[delta.RoomID].Events
limited := dbEvents[delta.RoomID].Limited
recentEvents := gomatrixserverlib.HeaderedReverseTopologicalOrdering(
snapshot.StreamEventsToEvents(device, recentStreamEvents),
gomatrixserverlib.TopologicalOrderByPrevEvents,
)
// If we didn't return any events at all then don't bother doing anything else.
if len(recentEvents) == 0 && len(delta.StateEvents) == 0 {
return r.To, nil
}
// Work out what the highest stream position is for all of the events in this
// room that were returned.
latestPosition := r.To
if r.Backwards {
latestPosition = r.From
}
updateLatestPosition := func(mostRecentEventID string) {
var pos types.StreamPosition
if _, pos, err = snapshot.PositionInTopology(ctx, mostRecentEventID); err == nil {
switch {
case r.Backwards && pos < latestPosition:
fallthrough
case !r.Backwards && pos > latestPosition:
latestPosition = pos
}
}
}
if stateFilter.LazyLoadMembers {
delta.StateEvents, err = p.lazyLoadMembers(
ctx, snapshot, delta.RoomID, true, limited, stateFilter,
device, recentEvents, delta.StateEvents,
)
if err != nil && err != sql.ErrNoRows {
return r.From, fmt.Errorf("p.lazyLoadMembers: %w", err)
}
}
hasMembershipChange := false
for _, recentEvent := range recentStreamEvents {
if recentEvent.Type() == gomatrixserverlib.MRoomMember && recentEvent.StateKey() != nil {
if membership, _ := recentEvent.Membership(); membership == gomatrixserverlib.Join {
req.MembershipChanges[*recentEvent.StateKey()] = struct{}{}
}
hasMembershipChange = true
}
}
// Applies the history visibility rules
events, err := applyHistoryVisibilityFilter(ctx, snapshot, p.rsAPI, delta.RoomID, device.UserID, recentEvents)
if err != nil {
logrus.WithError(err).Error("unable to apply history visibility filter")
}
if r.Backwards && len(events) > originalLimit {
// We're going backwards and the events are ordered chronologically, so take the last `limit` events
events = events[len(events)-originalLimit:]
limited = true
}
prevBatch, err := snapshot.GetBackwardTopologyPos(ctx, events)
if err != nil {
return r.From, fmt.Errorf("p.DB.GetBackwardTopologyPos: %w", err)
}
// Now that we've filtered the timeline, work out which state events are still
// left. Anything that appears in the filtered timeline will be removed from the
// "state" section and kept in "timeline".
delta.StateEvents = gomatrixserverlib.HeaderedReverseTopologicalOrdering(
removeDuplicates(delta.StateEvents, events),
gomatrixserverlib.TopologicalOrderByAuthEvents,
)
if len(delta.StateEvents) > 0 {
if last := delta.StateEvents[len(delta.StateEvents)-1]; last != nil {
updateLatestPosition(last.EventID())
}
}
if len(events) > 0 {
if last := events[len(events)-1]; last != nil {
updateLatestPosition(last.EventID())
}
}
switch delta.Membership {
case gomatrixserverlib.Join:
jr := types.NewJoinResponse()
if hasMembershipChange {
jr.Summary, err = snapshot.GetRoomSummary(ctx, delta.RoomID, device.UserID)
if err != nil {
logrus.WithError(err).Warn("failed to get room summary")
}
}
jr.Timeline.PrevBatch = &prevBatch
jr.Timeline.Events = gomatrixserverlib.HeaderedToClientEvents(events, gomatrixserverlib.FormatSync)
// If we are limited by the filter AND the history visibility filter
// didn't "remove" events, return that the response is limited.
jr.Timeline.Limited = (limited && len(events) == len(recentEvents)) || delta.NewlyJoined
jr.State.Events = gomatrixserverlib.HeaderedToClientEvents(delta.StateEvents, gomatrixserverlib.FormatSync)
req.Response.Rooms.Join[delta.RoomID] = jr
case gomatrixserverlib.Peek:
jr := types.NewJoinResponse()
jr.Timeline.PrevBatch = &prevBatch
// TODO: Apply history visibility on peeked rooms
jr.Timeline.Events = gomatrixserverlib.HeaderedToClientEvents(recentEvents, gomatrixserverlib.FormatSync)
jr.Timeline.Limited = limited
jr.State.Events = gomatrixserverlib.HeaderedToClientEvents(delta.StateEvents, gomatrixserverlib.FormatSync)
req.Response.Rooms.Peek[delta.RoomID] = jr
case gomatrixserverlib.Leave:
fallthrough // transitions to leave are the same as ban
case gomatrixserverlib.Ban:
lr := types.NewLeaveResponse()
lr.Timeline.PrevBatch = &prevBatch
lr.Timeline.Events = gomatrixserverlib.HeaderedToClientEvents(events, gomatrixserverlib.FormatSync)
// If we are limited by the filter AND the history visibility filter
// didn't "remove" events, return that the response is limited.
lr.Timeline.Limited = limited && len(events) == len(recentEvents)
lr.State.Events = gomatrixserverlib.HeaderedToClientEvents(delta.StateEvents, gomatrixserverlib.FormatSync)
req.Response.Rooms.Leave[delta.RoomID] = lr
}
return latestPosition, nil
}
// applyHistoryVisibilityFilter gets the current room state and supplies it to ApplyHistoryVisibilityFilter, to make
// sure we always return the required events in the timeline.
func applyHistoryVisibilityFilter(
ctx context.Context,
snapshot storage.DatabaseTransaction,
rsAPI roomserverAPI.SyncRoomserverAPI,
roomID, userID string,
recentEvents []*gomatrixserverlib.HeaderedEvent,
) ([]*gomatrixserverlib.HeaderedEvent, error) {
// We need to make sure we always include the latest state events, if they are in the timeline.
alwaysIncludeIDs := make(map[string]struct{})
var stateTypes []string
var senders []string
for _, ev := range recentEvents {
if ev.StateKey() != nil {
stateTypes = append(stateTypes, ev.Type())
senders = append(senders, ev.Sender())
}
}
// Only get the state again if there are state events in the timeline
if len(stateTypes) > 0 {
filter := gomatrixserverlib.DefaultStateFilter()
filter.Types = &stateTypes
filter.Senders = &senders
stateEvents, err := snapshot.CurrentState(ctx, roomID, &filter, nil)
if err != nil {
return nil, fmt.Errorf("failed to get current room state for history visibility calculation: %w", err)
}
for _, ev := range stateEvents {
alwaysIncludeIDs[ev.EventID()] = struct{}{}
}
}
startTime := time.Now()
events, err := internal.ApplyHistoryVisibilityFilter(ctx, snapshot, rsAPI, recentEvents, alwaysIncludeIDs, userID, "sync")
if err != nil {
return nil, err
}
logrus.WithFields(logrus.Fields{
"duration": time.Since(startTime),
"room_id": roomID,
"before": len(recentEvents),
"after": len(events),
}).Debugf("Applied history visibility (sync)")
return events, nil
}
func (p *PDUStreamProvider) getJoinResponseForCompleteSync(
ctx context.Context,
snapshot storage.DatabaseTransaction,
roomID string,
stateFilter *gomatrixserverlib.StateFilter,
wantFullState bool,
device *userapi.Device,
isPeek bool,
recentStreamEvents []types.StreamEvent,
limited bool,
) (jr *types.JoinResponse, err error) {
jr = types.NewJoinResponse()
// TODO: When filters are added, we may need to call this multiple times to get enough events.
// See: https://github.com/matrix-org/synapse/blob/v0.19.3/synapse/handlers/sync.py#L316
// Work our way through the timeline events and pick out the event IDs
// of any state events that appear in the timeline. We'll specifically
// exclude them at the next step, so that we don't get duplicate state
// events in both `recentStreamEvents` and `stateEvents`.
var excludingEventIDs []string
if !wantFullState {
excludingEventIDs = make([]string, 0, len(recentStreamEvents))
for _, event := range recentStreamEvents {
if event.StateKey() != nil {
excludingEventIDs = append(excludingEventIDs, event.EventID())
}
}
}
stateEvents, err := snapshot.CurrentState(ctx, roomID, stateFilter, excludingEventIDs)
if err != nil {
return
}
jr.Summary, err = snapshot.GetRoomSummary(ctx, roomID, device.UserID)
if err != nil {
logrus.WithError(err).Warn("failed to get room summary")
}
// We don't include a device here as we don't need to send down
// transaction IDs for complete syncs, but we do it anyway because Sytest demands it for:
// "Can sync a room with a message with a transaction id" - which does a complete sync to check.
recentEvents := snapshot.StreamEventsToEvents(device, recentStreamEvents)
events := recentEvents
// Only apply history visibility checks if the response is for joined rooms
if !isPeek {
events, err = applyHistoryVisibilityFilter(ctx, snapshot, p.rsAPI, roomID, device.UserID, recentEvents)
if err != nil {
logrus.WithError(err).Error("unable to apply history visibility filter")
}
}
// If we are limited by the filter AND the history visibility filter
// didn't "remove" events, return that the response is limited.
limited = limited && len(events) == len(recentEvents)
stateEvents = removeDuplicates(stateEvents, events)
if stateFilter.LazyLoadMembers {
if err != nil {
return nil, err
}
stateEvents, err = p.lazyLoadMembers(
ctx, snapshot, roomID,
false, limited, stateFilter,
device, recentEvents, stateEvents,
)
if err != nil && err != sql.ErrNoRows {
return nil, err
}
}
// Retrieve the backward topology position, i.e. the position of the
// oldest event in the room's topology.
var prevBatch *types.TopologyToken
if len(events) > 0 {
var backwardTopologyPos, backwardStreamPos types.StreamPosition
event := events[0]
// If this is the beginning of the room, we can't go back further. We're going to return
// the TopologyToken from the last event instead. (Synapse returns the /sync next_Batch)
if event.Type() == gomatrixserverlib.MRoomCreate && event.StateKeyEquals("") {
event = events[len(events)-1]
}
backwardTopologyPos, backwardStreamPos, err = snapshot.PositionInTopology(ctx, event.EventID())
if err != nil {
return
}
prevBatch = &types.TopologyToken{
Depth: backwardTopologyPos,
PDUPosition: backwardStreamPos,
}
prevBatch.Decrement()
}
jr.Timeline.PrevBatch = prevBatch
jr.Timeline.Events = gomatrixserverlib.HeaderedToClientEvents(events, gomatrixserverlib.FormatSync)
// If we are limited by the filter AND the history visibility filter
// didn't "remove" events, return that the response is limited.
jr.Timeline.Limited = limited && len(events) == len(recentEvents)
jr.State.Events = gomatrixserverlib.HeaderedToClientEvents(stateEvents, gomatrixserverlib.FormatSync)
return jr, nil
}
func (p *PDUStreamProvider) lazyLoadMembers(
ctx context.Context, snapshot storage.DatabaseTransaction, roomID string,
incremental, limited bool, stateFilter *gomatrixserverlib.StateFilter,
device *userapi.Device,
timelineEvents, stateEvents []*gomatrixserverlib.HeaderedEvent,
) ([]*gomatrixserverlib.HeaderedEvent, error) {
if len(timelineEvents) == 0 {
return stateEvents, nil
}
// Work out which memberships to include
timelineUsers := make(map[string]struct{})
if !incremental {
timelineUsers[device.UserID] = struct{}{}
}
// Add all users the client doesn't know about yet to a list
for _, event := range timelineEvents {
// Membership is not yet cached, add it to the list
if _, ok := p.lazyLoadCache.IsLazyLoadedUserCached(device, roomID, event.Sender()); !ok {
timelineUsers[event.Sender()] = struct{}{}
}
}
// Preallocate with the same amount, even if it will end up with fewer values
newStateEvents := make([]*gomatrixserverlib.HeaderedEvent, 0, len(stateEvents))
// Remove existing membership events we don't care about, e.g. users not in the timeline.events
for _, event := range stateEvents {
if event.Type() == gomatrixserverlib.MRoomMember && event.StateKey() != nil {
// If this is a gapped incremental sync, we still want this membership
isGappedIncremental := limited && incremental
// We want this users membership event, keep it in the list
stateKey := *event.StateKey()
if _, ok := timelineUsers[stateKey]; ok || isGappedIncremental || stateKey == device.UserID {
newStateEvents = append(newStateEvents, event)
if !stateFilter.IncludeRedundantMembers {
p.lazyLoadCache.StoreLazyLoadedUser(device, roomID, stateKey, event.EventID())
}
delete(timelineUsers, stateKey)
}
} else {
newStateEvents = append(newStateEvents, event)
}
}
wantUsers := make([]string, 0, len(timelineUsers))
for userID := range timelineUsers {
wantUsers = append(wantUsers, userID)
}
// Query missing membership events
filter := gomatrixserverlib.DefaultStateFilter()
filter.Senders = &wantUsers
filter.Types = &[]string{gomatrixserverlib.MRoomMember}
memberships, err := snapshot.GetStateEventsForRoom(ctx, roomID, &filter)
if err != nil {
return stateEvents, err
}
// cache the membership events
for _, membership := range memberships {
p.lazyLoadCache.StoreLazyLoadedUser(device, roomID, *membership.StateKey(), membership.EventID())
}
stateEvents = append(newStateEvents, memberships...)
return stateEvents, nil
}
// addIgnoredUsersToFilter adds ignored users to the eventfilter and
// the syncreq itself for further use in streams.
func (p *PDUStreamProvider) addIgnoredUsersToFilter(ctx context.Context, snapshot storage.DatabaseTransaction, req *types.SyncRequest, eventFilter *gomatrixserverlib.RoomEventFilter) error {
ignores, err := snapshot.IgnoresForUser(ctx, req.Device.UserID)
if err != nil {
if err == sql.ErrNoRows {
return nil
}
return err
}
req.IgnoredUsers = *ignores
userList := make([]string, 0, len(ignores.List))
for userID := range ignores.List {
userList = append(userList, userID)
}
if len(userList) > 0 {
eventFilter.NotSenders = &userList
}
return nil
}
func removeDuplicates(stateEvents, recentEvents []*gomatrixserverlib.HeaderedEvent) []*gomatrixserverlib.HeaderedEvent {
for _, recentEv := range recentEvents {
if recentEv.StateKey() == nil {
continue // not a state event
}
// TODO: This is a linear scan over all the current state events in this room. This will
// be slow for big rooms. We should instead sort the state events by event ID (ORDER BY)
// then do a binary search to find matching events, similar to what roomserver does.
for j := 0; j < len(stateEvents); j++ {
if stateEvents[j].EventID() == recentEv.EventID() {
// overwrite the element to remove with the last element then pop the last element.
// This is orders of magnitude faster than re-slicing, but doesn't preserve ordering
// (we don't care about the order of stateEvents)
stateEvents[j] = stateEvents[len(stateEvents)-1]
stateEvents = stateEvents[:len(stateEvents)-1]
break // there shouldn't be multiple events with the same event ID
}
}
}
return stateEvents
}