Processing stream in Flink

Question

I need to process messages as follows:

each message must pass unchanged through all processes.
if the process detects a pattern match, then a copy of the message must be passed to another thread.

streamIn -> process1 -> process2 -> .... -> processN
                 \         |                /
                  \        |               /
                     sink to kafka

I tried using OutputTag

 OutputTag<SysmonPartial> filtersOutput = new OutputTag<SysmonPartial>("FiltersOutput"){};

 DataStream<SysmonPartial> kafkaSource = env.addSource(consumer);

        DataStream<SysmonPartial> source = kafkaSource.rebalance();

        SingleOutputStreamOperator<SysmonPartial> s = source
                .process(lambda1).name("lambda1").startNewChain()
                .process(lambda2).name("lambda2").startNewChain()
                .process(lambda3).startNewChain()
                .process(lambda4).startNewChain()
                .process(lambda5).startNewChain()
                .process(lambda6).startNewChain()
                .process(lambda7).startNewChain();

        SingleOutputStreamOperator<Any> output = s
                .getSideOutput(filtersOutput)
                .process(filterProcessFunction).setParallelism(1).startNewChain();

        output.addSink(sink).setParallelism(1);

        env.execute(jobName);

where lambda 1, 2, etc. are condition checking functions, for example

public class Bypass_WS_01_03 extends FilterTagFuction<SysmonPartial, SysmonPartial, SysmonPartial> {
    private static final Pattern p_1 = Pattern.compile("pattern1");
    private static final Pattern p_0 = Pattern.compile("pattern2");

    @Override
    public void processElement(SysmonPartial t, Context ctx, Collector<SysmonPartial> out) throws Exception {
        out.collect(t);
        if (
                "1".equals(t.B_VendorEventID) &&
                        t.CommandLine != null && t.CommandLine.length() != 0 && p_0.matcher(t.CommandLine).find() &&
                        t.ImageName != null && t.ImageName.length() != 0 && p_1.matcher(t.ImageName).find()
        ) {
            t.RuleId = "Bypass_WS_01_03";
            ctx.output(getOutputTag(), t);
        }
    }
}

But for some reason it doesn't work for me, maybe there are some other methods? As I understand from the documentation(https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/side_output.html), OutputTag is used to create additional messages marked with Tag. Or am I wrong?

David Anderson David Anderson · Accepted Answer · 2020-09-18T12:57:16

The way you've wired up your job graph means that toward the end of the job where you access the side output

SingleOutputStreamOperator<Any> output = s
    .getSideOutput(filtersOutput)

you are only getting whatever the last process function put on the side output -- i.e., only the events emitted by lambda7.

What I believe you intend to do can be expressed as

SingleOutputStreamOperator<...> s1 = source.process(lambda1).name("lambda1");
SingleOutputStreamOperator<...> s2 = s1.process(lambda2).name("lambda2");
SingleOutputStreamOperator<...> s3 = s2.process(lambda3);
...

DataStream<...> side1 = s1.getSideOutput(filtersOutput);
DataStream<...> side2 = s2.getSideOutput(filtersOutput);
DataStream<...> side3 = s3.getSideOutput(filtersOutput);
...

SingleOutputStreamOperator<Any> output = side1.union(side2, side3, ...)
    .process(filterProcessFunction)
    ...

Also, I don't believe you should be using startNewChain() (which disables operator chaining). Operator chaining is a valuable optimization, and should only be disabled in special circumstances.

Processing stream in Flink

1 Answers