Flink Streaming: How to output one data stream to different outputs depending on the data?

Question

In Apache Flink I have a stream of tuples. Let's assume a really simple Tuple1<String>. The tuple can have an arbitrary value in it's value field (e.g. 'P1', 'P2', etc.). The set of possible values is finite but I don't know the full set beforehand (so there could be a 'P362'). I want to write that tuple to a certain output location depending on the value inside of the tuple. So e.g. I would like to have the following file structure:

• /output/P1
• /output/P2

In the documentation I only found possibilities to write to locations that I know beforehand (e.g. stream.writeCsv("/output/somewhere")), but no way of letting the contents of the data decide where the data is actually ending up.

I read about output splitting in the documentation but this doesn't seem to provide a way to redirect the output to different destinations the way I would like to have it (or I just don't understand how this would work).

Can this be done with the Flink API, if so, how? If not, is there maybe a third party library that can do it or would I have to build such a thing on my own?

Update

Following Matthias' suggestion I came up with a sifting sink function which determines the output path and then writes the tuple to the respective file after serializing it. I put it here for reference, maybe it is useful for someone else:

public class SiftingSinkFunction<IT> extends RichSinkFunction<IT> {

    private final OutputSelector<IT> outputSelector;
    private final MapFunction<IT, String> serializationFunction;
    private final String basePath;
    Map<String, TextOutputFormat<String>> formats = new HashMap<>();

    /**
     * @param outputSelector        the selector which determines into which output(s) a record is written.
     * @param serializationFunction a function which serializes the record to a string.
     * @param basePath              the base path for writing the records. It will be appended with the output selector.
     */
    public SiftingSinkFunction(OutputSelector<IT> outputSelector, MapFunction<IT, String> serializationFunction, String basePath) {
        this.outputSelector = outputSelector;
        this.serializationFunction = serializationFunction;
        this.basePath = basePath;
    }


    @Override
    public void invoke(IT value) throws Exception {
        // find out where to write.
        Iterable<String> selection = outputSelector.select(value);
        for (String s : selection) {
            // ensure we have a format for this.
            TextOutputFormat<String> destination = ensureDestinationExists(s);
            // then serialize and write.
            destination.writeRecord(serializationFunction.map(value));
        }
    }

    private TextOutputFormat<String> ensureDestinationExists(String selection) throws IOException {
        // if we know the destination, we just return the format.
        if (formats.containsKey(selection)) {
            return formats.get(selection);
        }

        // create a new output format and initialize it from the context.
        TextOutputFormat<String> format = new TextOutputFormat<>(new Path(basePath, selection));
        StreamingRuntimeContext context = (StreamingRuntimeContext) getRuntimeContext();
        format.configure(context.getTaskStubParameters());
        format.open(context.getIndexOfThisSubtask(), context.getNumberOfParallelSubtasks());

        // put it into our map.
        formats.put(selection, format);
        return format;
    }

    @Override
    public void close() throws IOException {
        Exception lastException = null;
        try {
            for (TextOutputFormat<String> format : formats.values()) {
                try {
                    format.close();
                } catch (Exception e) {
                    lastException = e;
                    format.tryCleanupOnError();
                }
            }
        } finally {
            formats.clear();
        }

        if (lastException != null) {
            throw new IOException("Close failed.", lastException);
        }
    }
}

Answer 1

You can implement a custom sink. Inherit from one of both:

• org.apache.flink.streaming.api.functions.sink.SinkFunction
• org.apache.flink.streaming.api.functions.sink.RichSinkFunction

In your program use:

stream.addSink(SinkFunction<T> sinkFunction);

instead of stream.writeCsv("/output/somewhere").