Can't understand TwoPhaseCommitSinkFunction lifecycle - scala

Question

I needed a sink to Postgres DB, so I started to build a custom Flink SinkFunction. As FlinkKafkaProducer implements TwoPhaseCommitSinkFunction, then I decided to do the same. As stated in O'Reilley's book Stream Processing with Apache Flink, you just need to implement the abstract methods, enable checkpointing and you're up to go. But what really happens when I run my code is that commit method is called only once, and it is called before invoke, what is totally unexpected since you shouldn't be ready to commit if your set of ready-to-commit transactions is empty. And the worst is that, after committing, invoke is called for all of the transaction lines present in my file, and then abort is called, which is even more unexpected.
When the Sink is initialized, It is of my understanding that the following should occur:
beginTransaction is called and sends an identifier to invoke
invoke adds the lines to the transaction, according to the identifier received
pre-commit makes all final modification on current transaction data
commit handles the finalized transaction of pre-commited data
So, I can't see why my program doesn't show this behaviour.
Here goes my sink code:
package PostgresConnector
import java.sql.{BatchUpdateException, DriverManager, PreparedStatement, SQLException, Timestamp}
import java.text.ParseException
import java.util.{Date, Properties, UUID}
import org.apache.flink.api.common.ExecutionConfig
import org.apache.flink.configuration.Configuration
import org.apache.flink.streaming.api.functions.sink.{SinkFunction, TwoPhaseCommitSinkFunction}
import org.apache.flink.streaming.api.scala._
import org.slf4j.{Logger, LoggerFactory}
class PostgreSink(props : Properties, config : ExecutionConfig) extends TwoPhaseCommitSinkFunction[(String,String,String,String),String,String](createTypeInformation[String].createSerializer(config),createTypeInformation[String].createSerializer(config)){
private var transactionMap : Map[String,Array[(String,String,String,String)]] = Map()
private var parsedQuery : PreparedStatement = _
private val insertionString : String = "INSERT INTO mydb (field1,field2,point) values (?,?,point(?,?))"
override def invoke(transaction: String, value: (String,String,String,String), context: SinkFunction.Context[_]): Unit = {
val LOG = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
val res = this.transactionMap.get(transaction)
if(res.isDefined){
var array = res.get
array = array ++ Array(value)
this.transactionMap += (transaction -> array)
}else{
val array = Array(value)
this.transactionMap += (transaction -> array)
}
LOG.info("\n\nPassing through invoke\n\n")
()
}
override def beginTransaction(): String = {
val LOG: Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
val identifier = UUID.randomUUID.toString
LOG.info("\n\nPassing through beginTransaction\n\n")
identifier
}
override def preCommit(transaction: String): Unit = {
val LOG = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
try{
val tuple : Option[Array[(String,String,String,String)]]= this.transactionMap.get(transaction)
if(tuple.isDefined){
tuple.get.foreach( (value : (String,String,String,String)) => {
LOG.info("\n\n"+value.toString()+"\n\n")
this.parsedQuery.setString(1,value._1)
this.parsedQuery.setString(2,value._2)
this.parsedQuery.setString(3,value._3)
this.parsedQuery.setString(4,value._4)
this.parsedQuery.addBatch()
})
}
}catch{
case e : SQLException =>
LOG.info("\n\nError when adding transaction to batch: SQLException\n\n")
case f : ParseException =>
LOG.info("\n\nError when adding transaction to batch: ParseException\n\n")
case g : NoSuchElementException =>
LOG.info("\n\nError when adding transaction to batch: NoSuchElementException\n\n")
case h : Exception =>
LOG.info("\n\nError when adding transaction to batch: Exception\n\n")
}
this.transactionMap = this.transactionMap.empty
LOG.info("\n\nPassing through preCommit...\n\n")
}
override def commit(transaction: String): Unit = {
val LOG : Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
if(this.parsedQuery != null) {
LOG.info("\n\n" + this.parsedQuery.toString+ "\n\n")
}
try{
this.parsedQuery.executeBatch
val LOG : Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
LOG.info("\n\nExecuting batch\n\n")
}catch{
case e : SQLException =>
val LOG : Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
LOG.info("\n\n"+"Error : SQLException"+"\n\n")
}
this.transactionMap = this.transactionMap.empty
LOG.info("\n\nPassing through commit...\n\n")
}
override def abort(transaction: String): Unit = {
val LOG : Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
this.transactionMap = this.transactionMap.empty
LOG.info("\n\nPassing through abort...\n\n")
}
override def open(parameters: Configuration): Unit = {
val LOG: Logger = LoggerFactory.getLogger(classOf[FlinkCEPClasses.FlinkCEPPipeline])
val driver = props.getProperty("driver")
val url = props.getProperty("url")
val user = props.getProperty("user")
val password = props.getProperty("password")
Class.forName(driver)
val connection = DriverManager.getConnection(url + "?user=" + user + "&password=" + password)
this.parsedQuery = connection.prepareStatement(insertionString)
LOG.info("\n\nConfiguring BD conection parameters\n\n")
}
}
And this is my main program:
package FlinkCEPClasses
import PostgresConnector.PostgreSink
import org.apache.flink.api.java.io.TextInputFormat
import org.apache.flink.api.java.utils.ParameterTool
import org.apache.flink.cep.PatternSelectFunction
import org.apache.flink.cep.pattern.conditions.SimpleCondition
import org.apache.flink.cep.scala.pattern.Pattern
import org.apache.flink.core.fs.{FileSystem, Path}
import org.apache.flink.streaming.api.scala._
import org.apache.flink.streaming.api.TimeCharacteristic
import org.apache.flink.cep.scala.{CEP, PatternStream}
import org.apache.flink.streaming.api.functions.source.FileProcessingMode
import org.apache.flink.streaming.api.scala.{DataStream, StreamExecutionEnvironment}
import java.util.Properties
import org.apache.flink.api.common.ExecutionConfig
import org.slf4j.{Logger, LoggerFactory}
class FlinkCEPPipeline {
val LOG: Logger = LoggerFactory.getLogger(classOf[FlinkCEPPipeline])
LOG.info("\n\nStarting the pipeline...\n\n")
var env : StreamExecutionEnvironment = StreamExecutionEnvironment.getExecutionEnvironment
env.enableCheckpointing(10)
env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime)
env.setParallelism(1)
//var input : DataStream[String] = env.readFile(new TextInputFormat(new Path("/home/luca/Desktop/lines")),"/home/luca/Desktop/lines",FileProcessingMode.PROCESS_CONTINUOUSLY,1)
var input : DataStream[String] = env.readTextFile("/home/luca/Desktop/lines").name("Raw stream")
var tupleStream : DataStream[(String,String,String,String)] = input.map(new S2PMapFunction()).name("Tuple Stream")
var properties : Properties = new Properties()
properties.setProperty("driver","org.postgresql.Driver")
properties.setProperty("url","jdbc:postgresql://localhost:5432/mydb")
properties.setProperty("user","luca")
properties.setProperty("password","root")
tupleStream.addSink(new PostgreSink(properties,env.getConfig)).name("Postgres Sink").setParallelism(1)
tupleStream.writeAsText("/home/luca/Desktop/output",FileSystem.WriteMode.OVERWRITE).name("File Sink").setParallelism(1)
env.execute()
}
My S2PMapFunction code:
package FlinkCEPClasses
import org.apache.flink.api.common.functions.MapFunction
case class S2PMapFunction() extends MapFunction[String,(String,String,String,String)] {
override def map(value: String): (String, String, String,String) = {
var tuple = value.replaceAllLiterally("(","").replaceAllLiterally(")","").split(',')
(tuple(0),tuple(1),tuple(2),tuple(3))
}
}
My pipeline works like this: I read lines from a file, map them to a tuple of strings, and use the data inside the tuples to save them in a Postgres DB
If you want to simulate the data, just create a file with lines in a format like this:
(field1,field2,pointx,pointy)
Edit
The execution order of the TwoPhaseCommitSinkFUnction's methods is the following:
Starting pipeline...
beginTransaction
preCommit
beginTransaction
commit
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
invoke
abort

Answer 1

I'm not an expert on this topic, but a couple of guesses:
preCommit is called whenever Flink begins a checkpoint, and commit is called when the checkpoint is complete. These methods are called simply because checkpointing is happening, regardless of whether the sink has received any data.
Checkpointing is happening periodically, regardless of whether any data is flowing through your pipeline. Given your very short checkpointing interval (10 msec), it does seem plausible that the first checkpoint barrier will reach the sink before the source has managed to send it any data.
It also looks like you are assuming that only one transaction will be open at a time. I'm not sure that's strictly guaranteed, but so long as maxConcurrentCheckpoints is 1 (which is the default), you should be okay.

Answer 2

So, here goes the "answer" for this question. Just to be clear: at this moment, the problem about the TwoPhaseCommitSinkFunction hasn't been solved yet. If what you're looking for is about the original problem, then you should look for another answer. If you don't care about what you'll use as a sink, then maybe I can help you with that.
As suggested by #DavidAnderson, I started to study the Table API and see if it could solve my problem, which was using Flink to insert lines in my database table.
It turned out to be really simple, as you'll see.
OBS: Beware of the version you are using. My Flink's version is 1.9.0.
Source code
package FlinkCEPClasses
import java.sql.Timestamp
import java.util.Properties
import org.apache.flink.api.common.typeinfo.{TypeInformation, Types}
import org.apache.flink.api.java.io.jdbc.JDBCAppendTableSink
import org.apache.flink.streaming.api.TimeCharacteristic
import org.apache.flink.streaming.api.scala.{DataStream, StreamExecutionEnvironment}
import org.apache.flink.table.api.{EnvironmentSettings, Table}
import org.apache.flink.table.api.scala.StreamTableEnvironment
import org.apache.flink.streaming.api.scala._
import org.apache.flink.table.sinks.TableSink
import org.postgresql.Driver
class TableAPIPipeline {
// --- normal pipeline initialization in this block ---
var env : StreamExecutionEnvironment = StreamExecutionEnvironment.getExecutionEnvironment
env.enableCheckpointing(10)
env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime)
env.setParallelism(1)
var input : DataStream[String] = env.readTextFile("/home/luca/Desktop/lines").name("Original stream")
var tupleStream : DataStream[(String,Timestamp,Double,Double)] = input.map(new S2PlacaMapFunction()).name("Tuple Stream")
var properties : Properties = new Properties()
properties.setProperty("driver","org.postgresql.Driver")
properties.setProperty("url","jdbc:postgresql://localhost:5432/mydb")
properties.setProperty("user","myuser")
properties.setProperty("password","mypassword")
// --- normal pipeline initialization in this block END ---
// These two lines create what Flink calls StreamTableEnvironment.
// It seems pretty similar to a normal stream initialization.
val settings = EnvironmentSettings.newInstance().useBlinkPlanner().inStreamingMode().build()
val tableEnv = StreamTableEnvironment.create(env,settings)
//Since I wanted to sink data into a database, I used JDBC TableSink,
//because it is very intuitive and is a exact match with my need. You may
//look for other TableSink classes that fit better in you solution.
var tableSink : JDBCAppendTableSink = JDBCAppendTableSink.builder()
.setBatchSize(1)
.setDBUrl("jdbc:postgresql://localhost:5432/mydb")
.setDrivername("org.postgresql.Driver")
.setPassword("mypassword")
.setUsername("myuser")
.setQuery("INSERT INTO mytable (data1,data2,data3) VALUES (?,?,point(?,?))")
.setParameterTypes(Types.STRING,Types.SQL_TIMESTAMP,Types.DOUBLE,Types.DOUBLE)
.build()
val fieldNames = Array("data1","data2","data3","data4")
val fieldTypes = Array[TypeInformation[_]](Types.STRING,Types.SQL_TIMESTAMP,Types.DOUBLE, Types.DOUBLE)
// This is the crucial part of the code: first, you need to register
// your table sink, informing the name, the field names, field types and
// the TableSink object.
tableEnv.registerTableSink("postgres-table-sink",
fieldNames,
fieldTypes,
tableSink
)
// Then, you transform your DataStream into a Table object.
var table = tableEnv.fromDataStream(tupleStream)
// Finally, you insert your stream data into the registered sink.
table.insertInto("postgres-table-sink")
env.execute()
}