@Test
public void testPGroupedTableToMultipleOutputs() throws IOException {
Pipeline pipeline = new MRPipeline(MRPipelineIT.class, tmpDir.getDefaultConfiguration());
PGroupedTable<String, String> groupedLineTable =
pipeline
.readTextFile(tmpDir.copyResourceFileName("set1.txt"))
.by(IdentityFn.<String>getInstance(), Writables.strings())
.groupByKey();
PTable<String, String> ungroupedTableA = groupedLineTable.ungroup();
PTable<String, String> ungroupedTableB = groupedLineTable.ungroup();
File outputDirA = tmpDir.getFile("output_a");
File outputDirB = tmpDir.getFile("output_b");
pipeline.writeTextFile(ungroupedTableA, outputDirA.getAbsolutePath());
pipeline.writeTextFile(ungroupedTableB, outputDirB.getAbsolutePath());
PipelineResult result = pipeline.done();
for (StageResult stageResult : result.getStageResults()) {
assertTrue(stageResult.getStageName().length() > 1);
assertTrue(stageResult.getStageId().length() > 1);
}
// Verify that output from a single PGroupedTable can be sent to multiple collections
assertTrue(new File(outputDirA, "part-r-00000").exists());
assertTrue(new File(outputDirB, "part-r-00000").exists());
}
/**
* Convert the given {@code PCollection<Pair<K, V>>} to a {@code PTable<K, V>}.
*
* @param pcollect The {@code PCollection} to convert
* @return A {@code PTable} that contains the same data as the input {@code PCollection}
*/
public static <K, V> PTable<K, V> asPTable(PCollection<Pair<K, V>> pcollect) {
PType<Pair<K, V>> pt = pcollect.getPType();
PTypeFamily ptf = pt.getFamily();
PTableType<K, V> ptt = ptf.tableOf(pt.getSubTypes().get(0), pt.getSubTypes().get(1));
DoFn<Pair<K, V>, Pair<K, V>> id = IdentityFn.getInstance();
return pcollect.parallelDo("asPTable", id, ptt);
}
public AvroType(Class<T> typeClass, Schema schema, DeepCopier<T> deepCopier, PType... ptypes) {
this(typeClass, schema, IdentityFn.getInstance(), IdentityFn.getInstance(), deepCopier, ptypes);
}
private void monitorLoop() {
status.set(Status.RUNNING);
long start = System.currentTimeMillis();
Map<PCollectionImpl<?>, Set<Target>> targetDeps = Maps.newTreeMap(DEPTH_COMPARATOR);
Set<Target> unfinished = Sets.newHashSet();
for (PCollectionImpl<?> pcollect : outputTargets.keySet()) {
targetDeps.put(pcollect, pcollect.getTargetDependencies());
unfinished.addAll(outputTargets.get(pcollect));
}
runCallables(unfinished);
while (!targetDeps.isEmpty() && doneSignal.getCount() > 0) {
Set<Target> allTargets = Sets.newHashSet();
for (PCollectionImpl<?> pcollect : targetDeps.keySet()) {
allTargets.addAll(outputTargets.get(pcollect));
}
Map<PCollectionImpl<?>, JavaRDDLike<?, ?>> pcolToRdd = Maps.newTreeMap(DEPTH_COMPARATOR);
for (PCollectionImpl<?> pcollect : targetDeps.keySet()) {
if (Sets.intersection(allTargets, targetDeps.get(pcollect)).isEmpty()) {
JavaRDDLike<?, ?> rdd = ((SparkCollection) pcollect).getJavaRDDLike(this);
pcolToRdd.put(pcollect, rdd);
}
}
distributeFiles();
for (Map.Entry<PCollectionImpl<?>, JavaRDDLike<?, ?>> e : pcolToRdd.entrySet()) {
JavaRDDLike<?, ?> rdd = e.getValue();
PType<?> ptype = e.getKey().getPType();
Set<Target> targets = outputTargets.get(e.getKey());
if (targets.size() > 1) {
rdd.rdd().cache();
}
for (Target t : targets) {
Configuration conf = new Configuration(getConfiguration());
getRuntimeContext().setConf(sparkContext.broadcast(WritableUtils.toByteArray(conf)));
if (t instanceof MapReduceTarget) { // TODO: check this earlier
Converter c = t.getConverter(ptype);
IdentityFn ident = IdentityFn.getInstance();
JavaPairRDD<?, ?> outRDD;
if (rdd instanceof JavaRDD) {
outRDD =
((JavaRDD) rdd)
.map(
new MapFunction(
c.applyPTypeTransforms() ? ptype.getOutputMapFn() : ident, ctxt))
.mapToPair(new OutputConverterFunction(c));
} else {
outRDD =
((JavaPairRDD) rdd)
.map(
new PairMapFunction(
c.applyPTypeTransforms() ? ptype.getOutputMapFn() : ident, ctxt))
.mapToPair(new OutputConverterFunction(c));
}
try {
Job job = new Job(conf);
if (t instanceof PathTarget) {
PathTarget pt = (PathTarget) t;
pt.configureForMapReduce(job, ptype, pt.getPath(), "out0");
CrunchOutputs.OutputConfig outConfig =
CrunchOutputs.getNamedOutputs(job.getConfiguration()).get("out0");
job.setOutputFormatClass(outConfig.bundle.getFormatClass());
job.setOutputKeyClass(outConfig.keyClass);
job.setOutputValueClass(outConfig.valueClass);
outConfig.bundle.configure(job.getConfiguration());
Path tmpPath = pipeline.createTempPath();
outRDD.saveAsNewAPIHadoopFile(
tmpPath.toString(),
c.getKeyClass(),
c.getValueClass(),
job.getOutputFormatClass(),
job.getConfiguration());
pt.handleOutputs(job.getConfiguration(), tmpPath, -1);
} else { // if (t instanceof MapReduceTarget) {
MapReduceTarget mrt = (MapReduceTarget) t;
mrt.configureForMapReduce(job, ptype, new Path("/tmp"), "out0");
CrunchOutputs.OutputConfig outConfig =
CrunchOutputs.getNamedOutputs(job.getConfiguration()).get("out0");
job.setOutputFormatClass(outConfig.bundle.getFormatClass());
job.setOutputKeyClass(outConfig.keyClass);
job.setOutputValueClass(outConfig.valueClass);
outRDD.saveAsHadoopDataset(new JobConf(job.getConfiguration()));
}
} catch (Exception et) {
LOG.error("Spark Exception", et);
status.set(Status.FAILED);
set(PipelineResult.EMPTY);
doneSignal.countDown();
}
}
}
unfinished.removeAll(targets);
}
if (status.get() == Status.RUNNING) {
for (PCollectionImpl<?> output : pcolToRdd.keySet()) {
if (toMaterialize.containsKey(output)) {
MaterializableIterable mi = toMaterialize.get(output);
if (mi.isSourceTarget()) {
output.materializeAt((SourceTarget) mi.getSource());
}
}
targetDeps.remove(output);
}
}
runCallables(unfinished);
}
if (status.get() != Status.FAILED || status.get() != Status.KILLED) {
status.set(Status.SUCCEEDED);
set(
new PipelineResult(
ImmutableList.of(
new PipelineResult.StageResult(
"Spark", getCounters(), start, System.currentTimeMillis())),
Status.SUCCEEDED));
} else {
set(PipelineResult.EMPTY);
}
doneSignal.countDown();
}