dbus

      $$\       $$\                                       
      $$ |      $$ |                                      
 $$$$$$$ |      $$$$$$$\        $$\   $$\        $$$$$$$\ 
$$  __$$ |      $$  __$$\       $$ |  $$ |      $$  _____|
$$ /  $$ |      $$ |  $$ |      $$ |  $$ |      \$$$$$$\  
$$ |  $$ |      $$ |  $$ |      $$ |  $$ |       \____$$\ 
\$$$$$$$ |      $$$$$$$  |      \$$$$$$  |      $$$$$$$  |
 \_______|      \_______/        \______/       \_______/ 

What is dbus?

dbus = distributed data bus

It is yet another lightweight versatile databus system that transfer/transform pipeline data between plugins.

dbus works by building a DAG of structured data out of the different plugins: from data input, via filter(optional), to the output.

Similar projects

• logstash
• flume
• nifi
• camel
• beats
• kettle
• zapier
• google cloud dataflow
• canal
• storm
• yahoo pipes (dead)

Status

dbus is not yet a 1.0. We're writing more tests, fixing bugs, working on TODOs.

Use Case

• mysql binlog dispatcher
• multiple DC kafka mirror

Features

dbus supports powerful and scalable directed graphs of data routing, transformation and system mediation logic.

• Designed for extension
  • plugin architecture
  • build your own plugins and more
  • enables rapid development and effective testing
• Data Provenance
  • track dataflow from beginning to end
  • visualized dataflow
  • rich metrics feed into tsdb
  • online manual mediation of the dataflow
  • RESTful API
  • monitoring with alert
• Distributed Deployment
  • shard/balance/auto rebalance
  • linear scale
• Delivery Guarantee
  • loss tolerant
  • high throuput vs low latency
  • back pressure
• Robustness
  • race condition detected
  • edge cases fully covered
  • network jitter tested
  • dependent components failure tested
• Systemic Quality
  • hot reload
  • dryrun throughput 1.9M packets/s
• Cluster Support
  • modelling borrowed from helix+kafka controller
  • currently only leader/standby with sharding, without replica
  • easy to write a distributed plugin

Getting Started

1. Installing

To start using dbus, install Go and run go get:

$ go get -u github.com/funkygao/dbus

2. Create config file

Please find sample config files in etc/ directory.

3. Run the server

$ $GOPATH/dbusd -conf $myfile

Dependencies

dbus uses zookeeper for sharding/balance/election.

Plugins

More plugins are listed under dbus-plugin.

Input

• MysqlbinlogInput
• KafkaInput
• MockInput
• StreamInput

Filter

• MysqlbinlogFilter
• MockFilter

Output

• KafkaOutput
• ESOutput
• MockOutput
• StreamOutput

Configuration

• KafkaOutput async mode with batch=1024/500ms, ack=WaitForAll
• KafkaOutput retry=3, retry.backoff=350ms
• Mysql binlog positioner commit every 1s, channal buffer 100

FAQ

mysql binlog

• is it totally data loss tolerant?

  if the binlog exceeds 1MB, it will be discarded(lost)

• ERROR 1236 (HY000): Could not find first log file name in binary log index file

  the checkpointed binlog position is gone on master, reset the zk znode and replication will automatically resume

• how to migrate mysql database?

  mysqldump --user=root --master-data --single-transaction --skip-lock-tables --compact --skip-opt --quick --no-create-info --skip-extended-insert --all-databases

  parse CHANGE MASTER TO MASTER_LOG_FILE='mysql.000005', MASTER_LOG_POS=80955497;

  MysqlbinlogInput load position for incremental loading

why not canal?

• no Delivery Guarantee
• no Data Provenance
• no integration with kafka
• only hot standby deployment mode, we need sharding load
• dbus is a dataflow engine, while canal only support mysql binlog pipeline

compared with logstash

• logstash has better ecosystem
• dbus is cluster aware, provides delivery guarantee, data provenance

can there be more than 1 leaders at the same time?

Yes.

For example, 3 participants with 1 being the leader. Then 1 is network partitioned and zk session expires, [2, 3] found this event and re-elect 2 as new leader. Before 1 regain new zk session, [1] and [2] are leaders both. If [1] and [2] both found resources changes, they will both rebalance the cluster.

dbus uses epoch to solve this issue.

what if

• zookeeper crash

  dbus continues to work, but Ack will not be able to persist

TODO

• resource group
• FIXME access denied leads to orphan resource
• myslave should have no checkpoint, placed in Input
• enhance Decision.Equals to avoid thundering herd
• myslave server_id uniq across the cluster
• tweak of batcher yield
• add Operator for Filter
  • count, filter, regex, sort, split, rename
• RowsEvent avro
• controller
  • a participant is electing, then shutdown took a long time(blocked by CreateLiveNode)
  • 2 phase rebalance: close participants then notify new resources
  • what if RPC fails
  • leader.onBecomingLeader is parallal: should be sequential
  • hot reload raises cluster herd: participant changes too much
  • when leader make decision, it persists to zk before RPC for leader failover
  • owner of resource
  • leader RPC has epoch info
  • if Ack fails(zk crash), resort to local disk(load on startup)
  • engine shutdown, controller still send rpc
  • test cases
    • sharded resources
    • brain split
    • zk dies or kill -9, use cache to continue work
    • kill -9 participant/leader, and reschedule
    • cluster chaos monkey
• kafka producer qos
• batcher only retries after full batch ack'ed, add timer?
• KafkaConsumer might not be able to Stop
• pack.Payload reuse memory, json.NewEncoder(os.Stdout)
• kguard integration
• router finding matcher is slow
• hot reload on config file changed
• each Input have its own recycle chan, one block will not block others
• when Input stops, Output might still need its OnAck
• KafkaInput plugin
• use scheme to distinguish type of DSN
• plugins Run has no way of panic
• (replication.go:117) [zabbix] invalid table id 2968, no correspond table map event
• make canal, high cpu usage
  • because CAS backoff 1us, cpu busy
• ugly design of Input/Output ack mechanism
  • we might learn from storm bolt ack
• some goroutine leakage
• telemetry mysql.binlog.lag/tps tag name should be input name
• pipeline
  • 1 input, multiple output
  • filter to dispatch dbs of a single binlog to different output
• kill Packet.input field
• visualized flow throughput like nifi
  • dump uses dag pkg
  • 
• router metrics
• dbusd api server
• logging
• share zkzone instance
• presence and standby mode
• graceful shutdown
• master must drain before leave cluster
• KafkaOutput metrics
  • binlog tps
  • kafka tps
  • lag
• hub is shared, what if a plugin blocks others
  • currently, I have no idea how to solve this issue
• Batcher padding
• shutdown kafka
• zk checkpoint vs kafka checkpoint
• kafka follower stops replication
• can a mysql instance with miltiple databases have multiple Log/Position?
• kafka sync produce in batch
• DDL binlog
  • drop table y;
• trace async producer Successes channel and mark as processed
• metrics
• telemetry and alert
• what if replication conn broken
• position will be stored in zk
• play with binlog_row_image
• project feature for multi-tenant
• bug fix
  • kill dbusd, dbusd-slave did not leave cluster
  • next log position leads to failure after resume
  • KafkaOutput only support 1 partition topic for MysqlbinlogInput
  • table id issue
  • what if invalid position
  • router stat wrong Total:142,535,625 0.00B speed:22,671/s 0.00B/s max: 0.00B/0.00B
  • ffjson marshalled bytes has NL before the ending bracket
• test cases
  • restart mysql master
  • mysql kill process
  • race detection
  • tc drop network packets and high latency
  • mysql binlog zk session expire
  • reset binlog pos, and check kafka didn't recv dup events
  • MysqlbinlogInput max_event_length
  • min.insync.replicas=2, shutdown 1 kafka broker then start
• GTID
  • place config to central zk znode and watch changes
• Known issues
  • Binlog Dump thread not close github/gh-ost#292
• Roadmap
  • pubsub audit reporter
  • universal kafka listener and outputer

Memo

• mysqlbinlog input peak with mock output

  • 140k event per second
  • 30k row event per second
  • 260Mb network bandwidth
  • KafkaOutput 35K msg per second
  • it takes 2h25m to zero lag for platform of 2d lag

• dryrun MockInput -> MockOutput

  • 2.1M packet/s