Factor out Consumer from the roomserver (#50)

This commit is contained in:
Kegsay 2017-03-27 10:00:28 +01:00 committed by GitHub
parent e3f3eb8f3d
commit a974b90ee9
7 changed files with 182 additions and 116 deletions

View File

@ -2,16 +2,18 @@ package main
import ( import (
"fmt" "fmt"
"github.com/matrix-org/dendrite/roomserver/input"
"github.com/matrix-org/dendrite/roomserver/query"
"github.com/matrix-org/dendrite/roomserver/storage"
"github.com/prometheus/client_golang/prometheus"
sarama "gopkg.in/Shopify/sarama.v1"
"net/http" "net/http"
_ "net/http/pprof" _ "net/http/pprof"
"os" "os"
"strconv" "strconv"
"strings" "strings"
"github.com/matrix-org/dendrite/common"
"github.com/matrix-org/dendrite/roomserver/input"
"github.com/matrix-org/dendrite/roomserver/query"
"github.com/matrix-org/dendrite/roomserver/storage"
"github.com/prometheus/client_golang/prometheus"
sarama "gopkg.in/Shopify/sarama.v1"
) )
var ( var (
@ -43,10 +45,13 @@ func main() {
} }
consumer := input.Consumer{ consumer := input.Consumer{
Consumer: kafkaConsumer, ContinualConsumer: common.ContinualConsumer{
Topic: inputRoomEventTopic,
Consumer: kafkaConsumer,
PartitionStore: db,
},
DB: db, DB: db,
Producer: kafkaProducer, Producer: kafkaProducer,
InputRoomEventTopic: inputRoomEventTopic,
OutputRoomEventTopic: outputRoomEventTopic, OutputRoomEventTopic: outputRoomEventTopic,
} }

View File

@ -0,0 +1,108 @@
package common
import (
"fmt"
sarama "gopkg.in/Shopify/sarama.v1"
)
// A PartitionOffset is the offset into a partition of the input log.
type PartitionOffset struct {
// The ID of the partition.
Partition int32
// The offset into the partition.
Offset int64
}
// A PartitionStorer has the storage APIs needed by the consumer.
type PartitionStorer interface {
// PartitionOffsets returns the offsets the consumer has reached for each partition.
PartitionOffsets(topic string) ([]PartitionOffset, error)
// SetPartitionOffset records where the consumer has reached for a partition.
SetPartitionOffset(topic string, partition int32, offset int64) error
}
// A ContinualConsumer continually consumes logs even across restarts. It requires a PartitionStorer to
// remember the offset it reached.
type ContinualConsumer struct {
// The kafkaesque topic to consume events from.
// This is the name used in kafka to identify the stream to consume events from.
Topic string
// A kafkaesque stream consumer providing the APIs for talking to the event source.
// The interface is taken from a client library for Apache Kafka.
// But any equivalent event streaming protocol could be made to implement the same interface.
Consumer sarama.Consumer
// A thing which can load and save partition offsets for a topic.
PartitionStore PartitionStorer
// ProcessMessage is a function which will be called for each message in the log. Return an error to
// stop processing messages. See ErrShutdown for specific control signals.
ProcessMessage func(msg *sarama.ConsumerMessage) error
// ShutdownCallback is called when ProcessMessage returns ErrShutdown, after the partition has been saved.
// It is optional.
ShutdownCallback func()
}
// ErrShutdown can be returned from ContinualConsumer.ProcessMessage to stop the ContinualConsumer.
var ErrShutdown = fmt.Errorf("shutdown")
// Start starts the consumer consuming.
// Starts up a goroutine for each partition in the kafka stream.
// Returns nil once all the goroutines are started.
// Returns an error if it can't start consuming for any of the partitions.
func (c *ContinualConsumer) Start() error {
offsets := map[int32]int64{}
partitions, err := c.Consumer.Partitions(c.Topic)
if err != nil {
return err
}
for _, partition := range partitions {
// Default all the offsets to the beginning of the stream.
offsets[partition] = sarama.OffsetOldest
}
storedOffsets, err := c.PartitionStore.PartitionOffsets(c.Topic)
if err != nil {
return err
}
for _, offset := range storedOffsets {
// We've already processed events from this partition so advance the offset to where we got to.
offsets[offset.Partition] = offset.Offset
}
var partitionConsumers []sarama.PartitionConsumer
for partition, offset := range offsets {
pc, err := c.Consumer.ConsumePartition(c.Topic, partition, offset)
if err != nil {
for _, p := range partitionConsumers {
p.Close()
}
return err
}
partitionConsumers = append(partitionConsumers, pc)
}
for _, pc := range partitionConsumers {
go c.consumePartition(pc)
}
return nil
}
// consumePartition consumes the room events for a single partition of the kafkaesque stream.
func (c *ContinualConsumer) consumePartition(pc sarama.PartitionConsumer) {
defer pc.Close()
for message := range pc.Messages() {
msgErr := c.ProcessMessage(message)
// Advance our position in the stream so that we will start at the right position after a restart.
if err := c.PartitionStore.SetPartitionOffset(c.Topic, message.Partition, message.Offset); err != nil {
panic(fmt.Errorf("the ContinualConsumer failed to SetPartitionOffset: %s", err))
}
// Shutdown if we were told to do so.
if msgErr == ErrShutdown {
if c.ShutdownCallback != nil {
c.ShutdownCallback()
}
return
}
}
}

View File

@ -1,9 +1,6 @@
package storage package common
import ( import "database/sql"
"database/sql"
"github.com/matrix-org/dendrite/roomserver/types"
)
const partitionOffsetsSchema = ` const partitionOffsetsSchema = `
-- The offsets that the server has processed up to. -- The offsets that the server has processed up to.
@ -26,32 +23,37 @@ const upsertPartitionOffsetsSQL = "" +
" ON CONFLICT ON CONSTRAINT topic_partition_unique" + " ON CONFLICT ON CONSTRAINT topic_partition_unique" +
" DO UPDATE SET partition_offset = $3" " DO UPDATE SET partition_offset = $3"
type partitionOffsetStatements struct { // PartitionOffsetStatements represents a set of statements that can be run on a partition_offsets table.
type PartitionOffsetStatements struct {
selectPartitionOffsetsStmt *sql.Stmt selectPartitionOffsetsStmt *sql.Stmt
upsertPartitionOffsetStmt *sql.Stmt upsertPartitionOffsetStmt *sql.Stmt
} }
func (s *partitionOffsetStatements) prepare(db *sql.DB) (err error) { // Prepare converts the raw SQL statements into prepared statements.
func (s *PartitionOffsetStatements) Prepare(db *sql.DB) (err error) {
_, err = db.Exec(partitionOffsetsSchema) _, err = db.Exec(partitionOffsetsSchema)
if err != nil { if err != nil {
return return
} }
if s.selectPartitionOffsetsStmt, err = db.Prepare(selectPartitionOffsetsSQL); err != nil {
return statementList{ return
{&s.selectPartitionOffsetsStmt, selectPartitionOffsetsSQL}, }
{&s.upsertPartitionOffsetStmt, upsertPartitionOffsetsSQL}, if s.upsertPartitionOffsetStmt, err = db.Prepare(upsertPartitionOffsetsSQL); err != nil {
}.prepare(db) return
}
return
} }
func (s *partitionOffsetStatements) selectPartitionOffsets(topic string) ([]types.PartitionOffset, error) { // SelectPartitionOffsets returns all the partition offsets for the given topic.
func (s *PartitionOffsetStatements) SelectPartitionOffsets(topic string) ([]PartitionOffset, error) {
rows, err := s.selectPartitionOffsetsStmt.Query(topic) rows, err := s.selectPartitionOffsetsStmt.Query(topic)
if err != nil { if err != nil {
return nil, err return nil, err
} }
defer rows.Close() defer rows.Close()
var results []types.PartitionOffset var results []PartitionOffset
for rows.Next() { for rows.Next() {
var offset types.PartitionOffset var offset PartitionOffset
if err := rows.Scan(&offset.Partition, &offset.Offset); err != nil { if err := rows.Scan(&offset.Partition, &offset.Offset); err != nil {
return nil, err return nil, err
} }
@ -59,7 +61,8 @@ func (s *partitionOffsetStatements) selectPartitionOffsets(topic string) ([]type
return results, nil return results, nil
} }
func (s *partitionOffsetStatements) upsertPartitionOffset(topic string, partition int32, offset int64) error { // UpsertPartitionOffset updates or inserts the partition offset for the given topic.
func (s *PartitionOffsetStatements) UpsertPartitionOffset(topic string, partition int32, offset int64) error {
_, err := s.upsertPartitionOffsetStmt.Exec(topic, partition, offset) _, err := s.upsertPartitionOffsetStmt.Exec(topic, partition, offset)
return err return err
} }

View File

@ -4,19 +4,17 @@ package input
import ( import (
"encoding/json" "encoding/json"
"fmt" "fmt"
"github.com/matrix-org/dendrite/roomserver/api"
"github.com/matrix-org/dendrite/roomserver/types"
sarama "gopkg.in/Shopify/sarama.v1"
"sync/atomic" "sync/atomic"
"github.com/matrix-org/dendrite/common"
"github.com/matrix-org/dendrite/roomserver/api"
sarama "gopkg.in/Shopify/sarama.v1"
) )
// A ConsumerDatabase has the storage APIs needed by the consumer. // A ConsumerDatabase has the storage APIs needed by the consumer.
type ConsumerDatabase interface { type ConsumerDatabase interface {
RoomEventDatabase RoomEventDatabase
// PartitionOffsets returns the offsets the consumer has reached for each partition. common.PartitionStorer
PartitionOffsets(topic string) ([]types.PartitionOffset, error)
// SetPartitionOffset records where the consumer has reached for a partition.
SetPartitionOffset(topic string, partition int32, offset int64) error
} }
// An ErrorLogger handles the errors encountered by the consumer. // An ErrorLogger handles the errors encountered by the consumer.
@ -31,16 +29,10 @@ type ErrorLogger interface {
// The events needed to construct the state at the event should already be stored on the roomserver. // The events needed to construct the state at the event should already be stored on the roomserver.
// If the event is not valid then it will be discarded and an error will be logged. // If the event is not valid then it will be discarded and an error will be logged.
type Consumer struct { type Consumer struct {
// A kafkaesque stream consumer providing the APIs for talking to the event source. ContinualConsumer common.ContinualConsumer
// The interface is taken from a client library for Apache Kafka.
// But any equivalent event streaming protocol could be made to implement the same interface.
Consumer sarama.Consumer
// The database used to store the room events. // The database used to store the room events.
DB ConsumerDatabase DB ConsumerDatabase
Producer sarama.SyncProducer Producer sarama.SyncProducer
// The kafkaesque topic to consume room events from.
// This is the name used in kafka to identify the stream to consume events from.
InputRoomEventTopic string
// The kafkaesque topic to output new room events to. // The kafkaesque topic to output new room events to.
// This is the name used in kafka to identify the stream to write events to. // This is the name used in kafka to identify the stream to write events to.
OutputRoomEventTopic string OutputRoomEventTopic string
@ -75,79 +67,42 @@ func (c *Consumer) WriteOutputRoomEvent(output api.OutputRoomEvent) error {
// Returns nil once all the goroutines are started. // Returns nil once all the goroutines are started.
// Returns an error if it can't start consuming for any of the partitions. // Returns an error if it can't start consuming for any of the partitions.
func (c *Consumer) Start() error { func (c *Consumer) Start() error {
offsets := map[int32]int64{} c.ContinualConsumer.ProcessMessage = c.processMessage
c.ContinualConsumer.ShutdownCallback = c.shutdown
return c.ContinualConsumer.Start()
}
partitions, err := c.Consumer.Partitions(c.InputRoomEventTopic) func (c *Consumer) processMessage(message *sarama.ConsumerMessage) error {
if err != nil { var input api.InputRoomEvent
return err if err := json.Unmarshal(message.Value, &input); err != nil {
} // If the message is invalid then log it and move onto the next message in the stream.
for _, partition := range partitions { c.logError(message, err)
// Default all the offsets to the beginning of the stream. } else {
offsets[partition] = sarama.OffsetOldest if err := processRoomEvent(c.DB, c, input); err != nil {
} // If there was an error processing the message then log it and
// move onto the next message in the stream.
storedOffsets, err := c.DB.PartitionOffsets(c.InputRoomEventTopic) // TODO: If the error was due to a problem talking to the database
if err != nil { // then we shouldn't move onto the next message and we should either
return err // retry processing the message, or panic and kill ourselves.
} c.logError(message, err)
for _, offset := range storedOffsets {
// We've already processed events from this partition so advance the offset to where we got to.
offsets[offset.Partition] = offset.Offset
}
var partitionConsumers []sarama.PartitionConsumer
for partition, offset := range offsets {
pc, err := c.Consumer.ConsumePartition(c.InputRoomEventTopic, partition, offset)
if err != nil {
for _, p := range partitionConsumers {
p.Close()
}
return err
} }
partitionConsumers = append(partitionConsumers, pc)
} }
for _, pc := range partitionConsumers { // Update the number of processed messages using atomic addition because it is accessed from multiple goroutines.
go c.consumePartition(pc) processed := atomic.AddInt64(&c.processed, 1)
// Check if we should stop processing.
// Note that since we have multiple goroutines it's quite likely that we'll overshoot by a few messages.
// If we try to stop processing after M message and we have N goroutines then we will process somewhere
// between M and (N + M) messages because the N goroutines could all try to process what they think will be the
// last message. We could be more careful here but this is good enough for getting rough benchmarks.
if c.StopProcessingAfter != nil && processed >= int64(*c.StopProcessingAfter) {
return common.ErrShutdown
} }
return nil return nil
} }
// consumePartition consumes the room events for a single partition of the kafkaesque stream. func (c *Consumer) shutdown() {
func (c *Consumer) consumePartition(pc sarama.PartitionConsumer) { if c.ShutdownCallback != nil {
defer pc.Close() c.ShutdownCallback(fmt.Sprintf("Stopping processing after %d messages", c.processed))
for message := range pc.Messages() {
var input api.InputRoomEvent
if err := json.Unmarshal(message.Value, &input); err != nil {
// If the message is invalid then log it and move onto the next message in the stream.
c.logError(message, err)
} else {
if err := processRoomEvent(c.DB, c, input); err != nil {
// If there was an error processing the message then log it and
// move onto the next message in the stream.
// TODO: If the error was due to a problem talking to the database
// then we shouldn't move onto the next message and we should either
// retry processing the message, or panic and kill ourselves.
c.logError(message, err)
}
}
// Advance our position in the stream so that we will start at the right position after a restart.
if err := c.DB.SetPartitionOffset(c.InputRoomEventTopic, message.Partition, message.Offset); err != nil {
c.logError(message, err)
}
// Update the number of processed messages using atomic addition because it is accessed from multiple goroutines.
processed := atomic.AddInt64(&c.processed, 1)
// Check if we should stop processing.
// Note that since we have multiple goroutines it's quite likely that we'll overshoot by a few messages.
// If we try to stop processing after M message and we have N goroutines then we will process somewhere
// between M and (N + M) messages because the N goroutines could all try to process what they think will be the
// last message. We could be more careful here but this is good enough for getting rough benchmarks.
if c.StopProcessingAfter != nil && processed >= int64(*c.StopProcessingAfter) {
if c.ShutdownCallback != nil {
c.ShutdownCallback(fmt.Sprintf("Stopping processing after %d messages", c.processed))
}
return
}
} }
} }

View File

@ -2,10 +2,12 @@ package storage
import ( import (
"database/sql" "database/sql"
"github.com/matrix-org/dendrite/common"
) )
type statements struct { type statements struct {
partitionOffsetStatements common.PartitionOffsetStatements
eventTypeStatements eventTypeStatements
eventStateKeyStatements eventStateKeyStatements
roomStatements roomStatements
@ -19,7 +21,7 @@ type statements struct {
func (s *statements) prepare(db *sql.DB) error { func (s *statements) prepare(db *sql.DB) error {
var err error var err error
if err = s.partitionOffsetStatements.prepare(db); err != nil { if err = s.PartitionOffsetStatements.Prepare(db); err != nil {
return err return err
} }

View File

@ -4,6 +4,7 @@ import (
"database/sql" "database/sql"
// Import the postgres database driver. // Import the postgres database driver.
_ "github.com/lib/pq" _ "github.com/lib/pq"
"github.com/matrix-org/dendrite/common"
"github.com/matrix-org/dendrite/roomserver/types" "github.com/matrix-org/dendrite/roomserver/types"
"github.com/matrix-org/gomatrixserverlib" "github.com/matrix-org/gomatrixserverlib"
) )
@ -28,13 +29,13 @@ func Open(dataSourceName string) (*Database, error) {
} }
// PartitionOffsets implements input.ConsumerDatabase // PartitionOffsets implements input.ConsumerDatabase
func (d *Database) PartitionOffsets(topic string) ([]types.PartitionOffset, error) { func (d *Database) PartitionOffsets(topic string) ([]common.PartitionOffset, error) {
return d.statements.selectPartitionOffsets(topic) return d.statements.SelectPartitionOffsets(topic)
} }
// SetPartitionOffset implements input.ConsumerDatabase // SetPartitionOffset implements input.ConsumerDatabase
func (d *Database) SetPartitionOffset(topic string, partition int32, offset int64) error { func (d *Database) SetPartitionOffset(topic string, partition int32, offset int64) error {
return d.statements.upsertPartitionOffset(topic, partition, offset) return d.statements.UpsertPartitionOffset(topic, partition, offset)
} }
// StoreEvent implements input.EventDatabase // StoreEvent implements input.EventDatabase

View File

@ -5,14 +5,6 @@ import (
"github.com/matrix-org/gomatrixserverlib" "github.com/matrix-org/gomatrixserverlib"
) )
// A PartitionOffset is the offset into a partition of the input log.
type PartitionOffset struct {
// The ID of the partition.
Partition int32
// The offset into the partition.
Offset int64
}
// EventTypeNID is a numeric ID for an event type. // EventTypeNID is a numeric ID for an event type.
type EventTypeNID int64 type EventTypeNID int64