Examples with Writer org.apache.beam.sdk.io.FileBasedSink.Writer used on opensource projects

Example 1 with Writer

use of org.apache.beam.sdk.io.FileBasedSink.Writer in project beam by apache.

the class WriteFiles method createWrite.

/**
   * A write is performed as sequence of three {@link ParDo}'s.
   *
   * <p>This singleton collection containing the WriteOperation is then used as a side
   * input to a ParDo over the PCollection of elements to write. In this bundle-writing phase,
   * {@link WriteOperation#createWriter} is called to obtain a {@link Writer}.
   * {@link Writer#open} and {@link Writer#close} are called in
   * {@link DoFn.StartBundle} and {@link DoFn.FinishBundle}, respectively, and
   * {@link Writer#write} method is called for every element in the bundle. The output
   * of this ParDo is a PCollection of <i>writer result</i> objects (see {@link FileBasedSink}
   * for a description of writer results)-one for each bundle.
   *
   * <p>The final do-once ParDo uses a singleton collection asinput and the collection of writer
   * results as a side-input. In this ParDo, {@link WriteOperation#finalize} is called
   * to finalize the write.
   *
   * <p>If the write of any element in the PCollection fails, {@link Writer#close} will be
   * called before the exception that caused the write to fail is propagated and the write result
   * will be discarded.
   *
   * <p>Since the {@link WriteOperation} is serialized after the initialization ParDo and
   * deserialized in the bundle-writing and finalization phases, any state change to the
   * WriteOperation object that occurs during initialization is visible in the latter
   * phases. However, the WriteOperation is not serialized after the bundle-writing
   * phase. This is why implementations should guarantee that
   * {@link WriteOperation#createWriter} does not mutate WriteOperation).
   */
private PDone createWrite(PCollection<T> input) {
    Pipeline p = input.getPipeline();
    if (!windowedWrites) {
        // Re-window the data into the global window and remove any existing triggers.
        input = input.apply(Window.<T>into(new GlobalWindows()).triggering(DefaultTrigger.of()).discardingFiredPanes());
    }
    // Perform the per-bundle writes as a ParDo on the input PCollection (with the
    // WriteOperation as a side input) and collect the results of the writes in a
    // PCollection. There is a dependency between this ParDo and the first (the
    // WriteOperation PCollection as a side input), so this will happen after the
    // initial ParDo.
    PCollection<FileResult> results;
    final PCollectionView<Integer> numShardsView;
    Coder<BoundedWindow> shardedWindowCoder = (Coder<BoundedWindow>) input.getWindowingStrategy().getWindowFn().windowCoder();
    if (computeNumShards == null && numShardsProvider == null) {
        numShardsView = null;
        results = input.apply("WriteBundles", ParDo.of(windowedWrites ? new WriteWindowedBundles() : new WriteUnwindowedBundles()));
    } else {
        List<PCollectionView<?>> sideInputs = Lists.newArrayList();
        if (computeNumShards != null) {
            numShardsView = input.apply(computeNumShards);
            sideInputs.add(numShardsView);
        } else {
            numShardsView = null;
        }
        PCollection<KV<Integer, Iterable<T>>> sharded = input.apply("ApplyShardLabel", ParDo.of(new ApplyShardingKey<T>(numShardsView, (numShardsView != null) ? null : numShardsProvider)).withSideInputs(sideInputs)).apply("GroupIntoShards", GroupByKey.<Integer, T>create());
        shardedWindowCoder = (Coder<BoundedWindow>) sharded.getWindowingStrategy().getWindowFn().windowCoder();
        results = sharded.apply("WriteShardedBundles", ParDo.of(new WriteShardedBundles()));
    }
    results.setCoder(FileResultCoder.of(shardedWindowCoder));
    if (windowedWrites) {
        // When processing streaming windowed writes, results will arrive multiple times. This
        // means we can't share the below implementation that turns the results into a side input,
        // as new data arriving into a side input does not trigger the listening DoFn. Instead
        // we aggregate the result set using a singleton GroupByKey, so the DoFn will be triggered
        // whenever new data arrives.
        PCollection<KV<Void, FileResult>> keyedResults = results.apply("AttachSingletonKey", WithKeys.<Void, FileResult>of((Void) null));
        keyedResults.setCoder(KvCoder.of(VoidCoder.of(), FileResultCoder.of(shardedWindowCoder)));
        // Is the continuation trigger sufficient?
        keyedResults.apply("FinalizeGroupByKey", GroupByKey.<Void, FileResult>create()).apply("Finalize", ParDo.of(new DoFn<KV<Void, Iterable<FileResult>>, Integer>() {

            @ProcessElement
            public void processElement(ProcessContext c) throws Exception {
                LOG.info("Finalizing write operation {}.", writeOperation);
                List<FileResult> results = Lists.newArrayList(c.element().getValue());
                writeOperation.finalize(results);
                LOG.debug("Done finalizing write operation");
            }
        }));
    } else {
        final PCollectionView<Iterable<FileResult>> resultsView = results.apply(View.<FileResult>asIterable());
        ImmutableList.Builder<PCollectionView<?>> sideInputs = ImmutableList.<PCollectionView<?>>builder().add(resultsView);
        if (numShardsView != null) {
            sideInputs.add(numShardsView);
        }
        // Finalize the write in another do-once ParDo on the singleton collection containing the
        // Writer. The results from the per-bundle writes are given as an Iterable side input.
        // The WriteOperation's state is the same as after its initialization in the first
        // do-once ParDo. There is a dependency between this ParDo and the parallel write (the writer
        // results collection as a side input), so it will happen after the parallel write.
        // For the non-windowed case, we guarantee that  if no data is written but the user has
        // set numShards, then all shards will be written out as empty files. For this reason we
        // use a side input here.
        PCollection<Void> singletonCollection = p.apply(Create.of((Void) null));
        singletonCollection.apply("Finalize", ParDo.of(new DoFn<Void, Integer>() {

            @ProcessElement
            public void processElement(ProcessContext c) throws Exception {
                LOG.info("Finalizing write operation {}.", writeOperation);
                List<FileResult> results = Lists.newArrayList(c.sideInput(resultsView));
                LOG.debug("Side input initialized to finalize write operation {}.", writeOperation);
                // We must always output at least 1 shard, and honor user-specified numShards if
                // set.
                int minShardsNeeded;
                if (numShardsView != null) {
                    minShardsNeeded = c.sideInput(numShardsView);
                } else if (numShardsProvider != null) {
                    minShardsNeeded = numShardsProvider.get();
                } else {
                    minShardsNeeded = 1;
                }
                int extraShardsNeeded = minShardsNeeded - results.size();
                if (extraShardsNeeded > 0) {
                    LOG.info("Creating {} empty output shards in addition to {} written for a total of {}.", extraShardsNeeded, results.size(), minShardsNeeded);
                    for (int i = 0; i < extraShardsNeeded; ++i) {
                        Writer<T> writer = writeOperation.createWriter();
                        writer.openUnwindowed(UUID.randomUUID().toString(), UNKNOWN_SHARDNUM);
                        FileResult emptyWrite = writer.close();
                        results.add(emptyWrite);
                    }
                    LOG.debug("Done creating extra shards.");
                }
                writeOperation.finalize(results);
                LOG.debug("Done finalizing write operation {}", writeOperation);
            }
        }).withSideInputs(sideInputs.build()));
    }
    return PDone.in(input.getPipeline());
}