Add NATS JetStream support (#1866)

* Add NATS JetStream support
Update shopify/sarama

* Fix addresses

* Don't change Addresses in Defaults

* Update saramajetstream

* Add missing error check

Keep typing events for at least one minute

* Use all configured NATS addresses

* Update saramajetstream

* Try setting up with NATS

* Make sure NATS uses own persistent directory (TODO: make this configurable)

* Update go.mod/go.sum

* Jetstream package

* Various other refactoring

* Build fixes

* Config tweaks, make random jetstream storage path for CI

* Disable interest policies

* Try to sane default on jetstream base path

* Try to use in-memory for CI

* Restore storage/retention

* Update nats.go dependency

* Adapt changes to config

* Remove unneeded TopicFor

* Dep update

* Revert "Remove unneeded TopicFor"

This reverts commit f5a4e4a339.

* Revert changes made to streams

* Fix build problems

* Update nats-server

* Update go.mod/go.sum

* Roomserver input API queuing using NATS

* Fix topic naming

* Prometheus metrics

* More refactoring to remove saramajetstream

* Add missing topic

* Don't try to populate map that doesn't exist

* Roomserver output topic

* Update go.mod/go.sum

* Message acknowledgements

* Ack tweaks

* Try to resume transaction re-sends

* Try to resume transaction re-sends

* Update to matrix-org/gomatrixserverlib@91dadfb

* Remove internal.PartitionStorer from components that don't consume keychanges

* Try to reduce re-allocations a bit in resolveConflictsV2

* Tweak delivery options on RS input

* Publish send-to-device messages into correct JetStream subject

* Async and sync roomserver input

* Update dendrite-config.yaml

* Remove roomserver tests for now (they need rewriting)

* Remove roomserver test again (was merged back in)

* Update documentation

* Docker updates

* More Docker updates

* Update Docker readme again

* Fix lint issues

* Send final event in `processEvent` synchronously (since this might stop Sytest from being so upset)

* Don't report event rejection errors via `/send`, since apparently this is upsetting tests that don't expect that

* Go 1.16 instead of Go 1.13 for upgrade tests and Complement

* Revert "Don't report event rejection errors via `/send`, since apparently this is upsetting tests that don't expect that"

This reverts commit 368675283f.

* Don't report any errors on `/send` to see what fun that creates

* Fix panics on closed channel sends

* Enforce state key matches sender

* Do the same for leave

* Various tweaks to make tests happier

Squashed commit of the following:

commit 13f9028e7a
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 15:47:14 2022 +0000

    Do the same for leave

commit e6be7f05c3
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 15:33:42 2022 +0000

    Enforce state key matches sender

commit 85ede6d64b
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 14:07:04 2022 +0000

    Fix panics on closed channel sends

commit 9755494a98
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 13:38:22 2022 +0000

    Don't report any errors on `/send` to see what fun that creates

commit 3bb4f87b5d
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 13:00:26 2022 +0000

    Revert "Don't report event rejection errors via `/send`, since apparently this is upsetting tests that don't expect that"

    This reverts commit 368675283f.

commit fe2673ed7b
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 12:09:34 2022 +0000

    Go 1.16 instead of Go 1.13 for upgrade tests and Complement

commit 368675283f
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 11:51:45 2022 +0000

    Don't report event rejection errors via `/send`, since apparently this is upsetting tests that don't expect that

commit b028dfc085
Author: Neil Alexander <neilalexander@users.noreply.github.com>
Date:   Tue Jan 4 10:29:08 2022 +0000

    Send final event in `processEvent` synchronously (since this might stop Sytest from being so upset)

* Merge in NATS Server v2.6.6 and nats.go v1.13 into the in-process connection fork

* Add `jetstream.WithJetStreamMessage` to make ack/nak-ing less messy, use process context in consumers

* Fix consumer component name in  federation API

* Add comment explaining where streams are defined

* Tweaks to roomserver input with comments

* Finish that sentence that I apparently forgot to finish in INSTALL.md

* Bump version number of config to 2

* Add comments around asynchronous sends to roomserver in processEventWithMissingState

* More useful error message when the config version does not match

* Set version in generate-config

* Fix version in config.Defaults

Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>

setup/base/base.go

@@ -81,8 +81,6 @@ type BaseDendrite struct {
 	Cfg                    *config.Dendrite
 	Caches                 *caching.Caches
 	DNSCache               *gomatrixserverlib.DNSCache
-	//	KafkaConsumer          sarama.Consumer
-	//	KafkaProducer          sarama.SyncProducer
 }
 
 const NoListener = ""

setup/config/config.go

@@ -40,7 +40,7 @@ var keyIDRegexp = regexp.MustCompile("^ed25519:[a-zA-Z0-9_]+$")
 
 // Version is the current version of the config format.
 // This will change whenever we make breaking changes to the config format.
-const Version = 1
+const Version = 2
 
 // Dendrite contains all the config used by a dendrite process.
 // Relative paths are resolved relative to the current working directory
@@ -292,7 +292,7 @@ func (config *Dendrite) Derive() error {
 
 // SetDefaults sets default config values if they are not explicitly set.
 func (c *Dendrite) Defaults(generate bool) {
-	c.Version = 1
+	c.Version = Version
 
 	c.Global.Defaults(generate)
 	c.ClientAPI.Defaults(generate)
@@ -325,6 +325,7 @@ func (c *Dendrite) Verify(configErrs *ConfigErrors, isMonolith bool) {
 }
 
 func (c *Dendrite) Wiring() {
+	c.Global.JetStream.Matrix = &c.Global
 	c.ClientAPI.Matrix = &c.Global
 	c.EDUServer.Matrix = &c.Global
 	c.FederationAPI.Matrix = &c.Global
@@ -420,7 +421,11 @@ func (config *Dendrite) check(_ bool) error { // monolithic
 
 	if config.Version != Version {
 		configErrs.Add(fmt.Sprintf(
-			"unknown config version %q, expected %q", config.Version, Version,
+			"config version is %q, expected %q - this means that the format of the configuration "+
+				"file has changed in some significant way, so please revisit the sample config "+
+				"and ensure you are not missing any important options that may have been added "+
+				"or changed recently!",
+			config.Version, Version,
 		))
 		return configErrs
 	}

setup/config/config_global.go

@@ -46,8 +46,8 @@ type Global struct {
 	// Defaults to an empty array.
 	TrustedIDServers []string `yaml:"trusted_third_party_id_servers"`
 
-	// Kafka/Naffka configuration
-	Kafka Kafka `yaml:"kafka"`
+	// JetStream configuration
+	JetStream JetStream `yaml:"jetstream"`
 
 	// Metrics configuration
 	Metrics Metrics `yaml:"metrics"`
@@ -68,7 +68,7 @@ func (c *Global) Defaults(generate bool) {
 	}
 	c.KeyValidityPeriod = time.Hour * 24 * 7
 
-	c.Kafka.Defaults(generate)
+	c.JetStream.Defaults(generate)
 	c.Metrics.Defaults(generate)
 	c.DNSCache.Defaults()
 	c.Sentry.Defaults()
@@ -78,7 +78,7 @@ func (c *Global) Verify(configErrs *ConfigErrors, isMonolith bool) {
 	checkNotEmpty(configErrs, "global.server_name", string(c.ServerName))
 	checkNotEmpty(configErrs, "global.private_key", string(c.PrivateKeyPath))
 
-	c.Kafka.Verify(configErrs, isMonolith)
+	c.JetStream.Verify(configErrs, isMonolith)
 	c.Metrics.Verify(configErrs, isMonolith)
 	c.Sentry.Verify(configErrs, isMonolith)
 	c.DNSCache.Verify(configErrs, isMonolith)

setup/config/config_jetstream.go

@@ -0,0 +1,40 @@
+package config
+
+import (
+	"fmt"
+)
+
+type JetStream struct {
+	Matrix *Global `yaml:"-"`
+
+	// Persistent directory to store JetStream streams in.
+	StoragePath Path `yaml:"storage_path"`
+	// A list of NATS addresses to connect to. If none are specified, an
+	// internal NATS server will be used when running in monolith mode only.
+	Addresses []string `yaml:"addresses"`
+	// The prefix to use for stream names for this homeserver - really only
+	// useful if running more than one Dendrite on the same NATS deployment.
+	TopicPrefix string `yaml:"topic_prefix"`
+	// Keep all storage in memory. This is mostly useful for unit tests.
+	InMemory bool `yaml:"in_memory"`
+}
+
+func (c *JetStream) TopicFor(name string) string {
+	return fmt.Sprintf("%s%s", c.TopicPrefix, name)
+}
+
+func (c *JetStream) Defaults(generate bool) {
+	c.Addresses = []string{}
+	c.TopicPrefix = "Dendrite"
+	if generate {
+		c.StoragePath = Path("./")
+	}
+}
+
+func (c *JetStream) Verify(configErrs *ConfigErrors, isMonolith bool) {
+	// If we are running in a polylith deployment then we need at least
+	// one NATS JetStream server to talk to.
+	if !isMonolith {
+		checkNotZero(configErrs, "global.jetstream.addresses", int64(len(c.Addresses)))
+	}
+}

setup/config/config_kafka.go

@@ -1,63 +0,0 @@
-package config
-
-import "fmt"
-
-// Defined Kafka topics.
-const (
-	TopicOutputTypingEvent       = "OutputTypingEvent"
-	TopicOutputSendToDeviceEvent = "OutputSendToDeviceEvent"
-	TopicOutputKeyChangeEvent    = "OutputKeyChangeEvent"
-	TopicOutputRoomEvent         = "OutputRoomEvent"
-	TopicOutputClientData        = "OutputClientData"
-	TopicOutputReceiptEvent      = "OutputReceiptEvent"
-)
-
-type Kafka struct {
-	// A list of kafka addresses to connect to.
-	Addresses []string `yaml:"addresses"`
-	// The prefix to use for Kafka topic names for this homeserver - really only
-	// useful if running more than one Dendrite on the same Kafka deployment.
-	TopicPrefix string `yaml:"topic_prefix"`
-	// Whether to use naffka instead of kafka.
-	// Naffka can only be used when running dendrite as a single monolithic server.
-	// Kafka can be used both with a monolithic server and when running the
-	// components as separate servers.
-	UseNaffka bool `yaml:"use_naffka"`
-	// The Naffka database is used internally by the naffka library, if used.
-	Database DatabaseOptions `yaml:"naffka_database"`
-	// The max size a Kafka message passed between consumer/producer can have
-	// Equals roughly max.message.bytes / fetch.message.max.bytes in Kafka
-	MaxMessageBytes *int `yaml:"max_message_bytes"`
-}
-
-func (k *Kafka) TopicFor(name string) string {
-	return fmt.Sprintf("%s%s", k.TopicPrefix, name)
-}
-
-func (c *Kafka) Defaults(generate bool) {
-	c.UseNaffka = true
-	c.Database.Defaults(10)
-	if generate {
-		c.Addresses = []string{"localhost:2181"}
-		c.Database.ConnectionString = DataSource("file:naffka.db")
-	}
-	c.TopicPrefix = "Dendrite"
-
-	maxBytes := 1024 * 1024 * 8 // about 8MB
-	c.MaxMessageBytes = &maxBytes
-}
-
-func (c *Kafka) Verify(configErrs *ConfigErrors, isMonolith bool) {
-	if c.UseNaffka {
-		if !isMonolith {
-			configErrs.Add("naffka can only be used in a monolithic server")
-		}
-		checkNotEmpty(configErrs, "global.kafka.database.connection_string", string(c.Database.ConnectionString))
-	} else {
-		// If we aren't using naffka then we need to have at least one kafka
-		// server to talk to.
-		checkNotZero(configErrs, "global.kafka.addresses", int64(len(c.Addresses)))
-	}
-	checkNotEmpty(configErrs, "global.kafka.topic_prefix", string(c.TopicPrefix))
-	checkPositive(configErrs, "global.kafka.max_message_bytes", int64(*c.MaxMessageBytes))
-}

setup/config/config_test.go

@@ -33,7 +33,7 @@ func TestLoadConfigRelative(t *testing.T) {
 }
 
 const testConfig = `
-version: 1
+version: 2
 global:
   server_name: localhost
   private_key: matrix_key.pem

setup/jetstream/helpers.go

@@ -0,0 +1,11 @@
+package jetstream
+
+import "github.com/nats-io/nats.go"
+
+func WithJetStreamMessage(msg *nats.Msg, f func(msg *nats.Msg) bool) {
+	if f(msg) {
+		_ = msg.Ack()
+	} else {
+		_ = msg.Nak()
+	}
+}

setup/jetstream/nats.go

@@ -0,0 +1,93 @@
+package jetstream
+
+import (
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/Shopify/sarama"
+	"github.com/matrix-org/dendrite/setup/config"
+	"github.com/sirupsen/logrus"
+
+	saramajs "github.com/S7evinK/saramajetstream"
+	natsserver "github.com/nats-io/nats-server/v2/server"
+	"github.com/nats-io/nats.go"
+	natsclient "github.com/nats-io/nats.go"
+)
+
+var natsServer *natsserver.Server
+var natsServerMutex sync.Mutex
+
+func Prepare(cfg *config.JetStream) (nats.JetStreamContext, sarama.Consumer, sarama.SyncProducer) {
+	// check if we need an in-process NATS Server
+	if len(cfg.Addresses) != 0 {
+		return setupNATS(cfg, nil)
+	}
+	natsServerMutex.Lock()
+	if natsServer == nil {
+		var err error
+		natsServer, err = natsserver.NewServer(&natsserver.Options{
+			ServerName:       "monolith",
+			DontListen:       true,
+			JetStream:        true,
+			StoreDir:         string(cfg.StoragePath),
+			NoSystemAccount:  true,
+			AllowNewAccounts: false,
+		})
+		if err != nil {
+			panic(err)
+		}
+		natsServer.ConfigureLogger()
+		go natsServer.Start()
+	}
+	natsServerMutex.Unlock()
+	if !natsServer.ReadyForConnections(time.Second * 10) {
+		logrus.Fatalln("NATS did not start in time")
+	}
+	nc, err := natsclient.Connect("", natsclient.InProcessServer(natsServer))
+	if err != nil {
+		logrus.Fatalln("Failed to create NATS client")
+	}
+	return setupNATS(cfg, nc)
+}
+
+func setupNATS(cfg *config.JetStream, nc *natsclient.Conn) (nats.JetStreamContext, sarama.Consumer, sarama.SyncProducer) {
+	if nc == nil {
+		var err error
+		nc, err = nats.Connect(strings.Join(cfg.Addresses, ","))
+		if err != nil {
+			logrus.WithError(err).Panic("Unable to connect to NATS")
+			return nil, nil, nil
+		}
+	}
+
+	s, err := nc.JetStream()
+	if err != nil {
+		logrus.WithError(err).Panic("Unable to get JetStream context")
+		return nil, nil, nil
+	}
+
+	for _, stream := range streams { // streams are defined in streams.go
+		name := cfg.TopicFor(stream.Name)
+		info, err := s.StreamInfo(name)
+		if err != nil && err != natsclient.ErrStreamNotFound {
+			logrus.WithError(err).Fatal("Unable to get stream info")
+		}
+		if info == nil {
+			stream.Subjects = []string{name}
+			// If we're trying to keep everything in memory (e.g. unit tests)
+			// then overwrite the storage policy.
+			if cfg.InMemory {
+				stream.Storage = nats.MemoryStorage
+			}
+
+			if _, err = s.AddStream(stream); err != nil {
+				logrus.WithError(err).WithField("stream", name).Fatal("Unable to add stream")
+			}
+		}
+	}
+
+	consumer := saramajs.NewJetStreamConsumer(nc, s, "")
+	producer := saramajs.NewJetStreamProducer(nc, s, "")
+	return s, consumer, producer
+}

setup/jetstream/streams.go

@@ -0,0 +1,61 @@
+package jetstream
+
+import (
+	"time"
+
+	"github.com/nats-io/nats.go"
+)
+
+const (
+	UserID = "user_id"
+	RoomID = "room_id"
+)
+
+var (
+	InputRoomEvent          = "InputRoomEvent"
+	OutputRoomEvent         = "OutputRoomEvent"
+	OutputSendToDeviceEvent = "OutputSendToDeviceEvent"
+	OutputKeyChangeEvent    = "OutputKeyChangeEvent"
+	OutputTypingEvent       = "OutputTypingEvent"
+	OutputClientData        = "OutputClientData"
+	OutputReceiptEvent      = "OutputReceiptEvent"
+)
+
+var streams = []*nats.StreamConfig{
+	{
+		Name:      InputRoomEvent,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.FileStorage,
+	},
+	{
+		Name:      OutputRoomEvent,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.FileStorage,
+	},
+	{
+		Name:      OutputSendToDeviceEvent,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.FileStorage,
+	},
+	{
+		Name:      OutputKeyChangeEvent,
+		Retention: nats.LimitsPolicy,
+		Storage:   nats.FileStorage,
+	},
+	{
+		Name:      OutputTypingEvent,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.MemoryStorage,
+		MaxAge:    time.Second * 60,
+	},
+	{
+		Name:      OutputClientData,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.FileStorage,
+	},
+	{
+		Name:      OutputReceiptEvent,
+		Retention: nats.InterestPolicy,
+		Storage:   nats.FileStorage,
+	},
+}

setup/kafka/kafka.go

@@ -1,58 +0,0 @@
-package kafka
-
-import (
-	"github.com/Shopify/sarama"
-	"github.com/matrix-org/dendrite/setup/config"
-	"github.com/matrix-org/naffka"
-	naffkaStorage "github.com/matrix-org/naffka/storage"
-	"github.com/sirupsen/logrus"
-)
-
-func SetupConsumerProducer(cfg *config.Kafka) (sarama.Consumer, sarama.SyncProducer) {
-	if cfg.UseNaffka {
-		return setupNaffka(cfg)
-	}
-	return setupKafka(cfg)
-}
-
-// setupKafka creates kafka consumer/producer pair from the config.
-func setupKafka(cfg *config.Kafka) (sarama.Consumer, sarama.SyncProducer) {
-	sCfg := sarama.NewConfig()
-	sCfg.Producer.MaxMessageBytes = *cfg.MaxMessageBytes
-	sCfg.Producer.Return.Successes = true
-	sCfg.Consumer.Fetch.Default = int32(*cfg.MaxMessageBytes)
-
-	consumer, err := sarama.NewConsumer(cfg.Addresses, sCfg)
-	if err != nil {
-		logrus.WithError(err).Panic("failed to start kafka consumer")
-	}
-
-	producer, err := sarama.NewSyncProducer(cfg.Addresses, sCfg)
-	if err != nil {
-		logrus.WithError(err).Panic("failed to setup kafka producers")
-	}
-
-	return consumer, producer
-}
-
-// In monolith mode with Naffka, we don't have the same constraints about
-// consuming the same topic from more than one place like we do with Kafka.
-// Therefore, we will only open one Naffka connection in case Naffka is
-// running on SQLite.
-var naffkaInstance *naffka.Naffka
-
-// setupNaffka creates kafka consumer/producer pair from the config.
-func setupNaffka(cfg *config.Kafka) (sarama.Consumer, sarama.SyncProducer) {
-	if naffkaInstance != nil {
-		return naffkaInstance, naffkaInstance
-	}
-	naffkaDB, err := naffkaStorage.NewDatabase(string(cfg.Database.ConnectionString))
-	if err != nil {
-		logrus.WithError(err).Panic("Failed to setup naffka database")
-	}
-	naffkaInstance, err = naffka.New(naffkaDB)
-	if err != nil {
-		logrus.WithError(err).Panic("Failed to setup naffka")
-	}
-	return naffkaInstance, naffkaInstance
-}

setup/mscs/msc2836/msc2836.go

@@ -649,7 +649,7 @@ func (rc *reqCtx) injectResponseToRoomserver(res *gomatrixserverlib.MSC2836Event
 		})
 	}
 	// we've got the data by this point so use a background context
-	err = roomserver.SendInputRoomEvents(context.Background(), rc.rsAPI, ires)
+	err = roomserver.SendInputRoomEvents(context.Background(), rc.rsAPI, ires, false)
 	if err != nil {
 		util.GetLogger(rc.ctx).WithError(err).Error("failed to inject MSC2836EventRelationshipsResponse into the roomserver")
 	}