Examples with Flatten com.tdunning.plume.local.lazy.op.Flatten used on opensource projects

Example 1 with Flatten

use of com.tdunning.plume.local.lazy.op.Flatten in project Plume by tdunning.

the class MSCRMapper method map.

@SuppressWarnings("unchecked")
protected void map(WritableComparable key, WritableComparable value, final Mapper<WritableComparable, WritableComparable, PlumeObject, PlumeObject>.Context<WritableComparable, WritableComparable, PlumeObject, PlumeObject> context) throws IOException, InterruptedException {
    LazyCollection<?> l = null;
    FileSplit fS = FileInputSplitWrapper.getFileInputSplit(context);
    // Get LazyCollection for this input (according to FileSplit)
    for (PCollection<?> input : mscr.getInputs()) {
        LazyCollection<?> thisL = (LazyCollection<?>) input;
        if (thisL.getFile() == null) {
            // Convention for intermediate results
            thisL.setFile(tmpFolder + "/" + thisL.getPlumeId());
        }
        if (fS.getPath().toString().startsWith(thisL.getFile()) || fS.getPath().toString().startsWith("file:" + thisL.getFile())) {
            l = thisL;
            break;
        }
    }
    if (l == null) {
        throw new RuntimeException("Unable to match input split with any MSCR input");
    }
    // If this collection is a table -> process Pair, otherwise process value
    PCollectionType type = l.getType();
    Object toProcess = value;
    if (type instanceof PTableType) {
        toProcess = Pair.create(key, value);
    }
    for (DeferredOp op : l.getDownOps()) {
        if (op instanceof MultipleParallelDo) {
            MultipleParallelDo mPDo = ((MultipleParallelDo) op);
            for (Object entry : mPDo.getDests().entrySet()) {
                Map.Entry<PCollection, DoFn> en = (Map.Entry<PCollection, DoFn>) entry;
                LazyCollection<?> lCol = (LazyCollection<?>) en.getKey();
                DeferredOp childOp = null;
                if (lCol.getDownOps() != null && lCol.getDownOps().size() > 0) {
                    childOp = lCol.getDownOps().get(0);
                }
                final Integer channel;
                if (childOp != null && childOp instanceof Flatten) {
                    channel = mscr.getNumberedChannels().get(((Flatten) childOp).getDest());
                } else if (childOp != null && childOp instanceof GroupByKey) {
                    channel = mscr.getNumberedChannels().get(((GroupByKey) childOp).getOrigin());
                } else {
                    // bypass channel?
                    channel = mscr.getNumberedChannels().get(en.getKey());
                }
                if (channel == null) {
                    // This is not for this MSCR - just skip it
                    return;
                }
                // Call parallelDo function
                en.getValue().process(toProcess, new EmitFn() {

                    @Override
                    public void emit(Object v) {
                        try {
                            if (v instanceof Pair) {
                                Pair p = (Pair) v;
                                context.write(new PlumeObject((WritableComparable) p.getKey(), channel), new PlumeObject((WritableComparable) p.getValue(), channel));
                            } else {
                                context.write(new PlumeObject((WritableComparable) v, channel), new PlumeObject((WritableComparable) v, channel));
                            }
                        } catch (Exception e) {
                            // TODO How to report this
                            e.printStackTrace();
                        }
                    }
                });
            }
        } else {
            if (op instanceof Flatten) {
                l = (LazyCollection) ((Flatten) op).getDest();
            }
            int channel = mscr.getNumberedChannels().get(l);
            if (toProcess instanceof Pair) {
                context.write(new PlumeObject(key, channel), new PlumeObject(value, channel));
            } else {
                context.write(new PlumeObject(value, channel), new PlumeObject(value, channel));
            }
        }
    }
}

Example 2 with Flatten

use of com.tdunning.plume.local.lazy.op.Flatten in project Plume by tdunning.

the class BasicOptimizerTest method testSinkFlattens.

/**
 * This test has two inputs, one flatten and then one ParallelDo.
 * After sinking flattens, the tree should be as: two inputs, one ParallelDo after each input and one final Flatten.
 */
@Test
public void testSinkFlattens() {
    // Get Plume runtime
    LazyPlume plume = new LazyPlume();
    // Create simple data
    PCollection<Integer> input1 = plume.fromJava(Lists.newArrayList(1, 2, 3));
    PCollection<Integer> input2 = plume.fromJava(Lists.newArrayList(4, 5, 6));
    PCollection<Integer> output = plume.flatten(input1, input2).map(plusOne, null);
    LazyCollection<Integer> lOutput = (LazyCollection<Integer>) output;
    assertTrue(lOutput.getDeferredOp() instanceof ParallelDo);
    // Execute and assert the result before optimizing
    executeAndAssert((LazyCollection<Integer>) output, new Integer[] { 2, 3, 4, 5, 6, 7 });
    // Get an Optimizer
    Optimizer optimizer = new Optimizer();
    optimizer.sinkFlattens(output);
    // Execute and assert the result after optimizing
    executeAndAssert((LazyCollection<Integer>) output, new Integer[] { 2, 3, 4, 5, 6, 7 });
    // Check that optimizer did what it's supposed to do
    assertTrue(lOutput.getDeferredOp() instanceof Flatten);
    Flatten flatten = (Flatten) lOutput.getDeferredOp();
    assertEquals(flatten.getOrigins().size(), 2);
    for (int i = 0; i < 2; i++) {
        LazyCollection<Integer> origin = (LazyCollection<Integer>) flatten.getOrigins().get(i);
        ParallelDo newPDo = (ParallelDo) origin.getDeferredOp();
        assertEquals(newPDo.getFunction(), plusOne);
        assertTrue(newPDo.getOrigin() == input1 || newPDo.getOrigin() == input2);
    }
}

Example 3 with Flatten

use of com.tdunning.plume.local.lazy.op.Flatten in project Plume by tdunning.

the class MSCRMapper method map.

@SuppressWarnings("unchecked")
protected void map(WritableComparable key, WritableComparable value, final Mapper<WritableComparable, WritableComparable, PlumeObject, PlumeObject>.Context context) throws IOException, InterruptedException {
    LazyCollection<?> l = null;
    FileSplit fS = FileInputSplitWrapper.getFileInputSplit(context);
    // Get LazyCollection for this input (according to FileSplit)
    for (PCollection<?> input : mscr.getInputs()) {
        LazyCollection<?> thisL = (LazyCollection<?>) input;
        if (thisL.getFile() == null) {
            // Convention for intermediate results
            thisL.setFile(tmpFolder + "/" + thisL.getPlumeId());
        }
        if (fS.getPath().toString().startsWith(thisL.getFile()) || fS.getPath().toString().startsWith("file:" + thisL.getFile())) {
            l = thisL;
            break;
        }
    }
    if (l == null) {
        throw new RuntimeException("Unable to match input split with any MSCR input");
    }
    // If this collection is a table -> process Pair, otherwise process value
    PCollectionType type = l.getType();
    Object toProcess = value;
    if (type instanceof PTableType) {
        toProcess = Pair.create(key, value);
    }
    for (DeferredOp op : l.getDownOps()) {
        if (op instanceof MultipleParallelDo) {
            MultipleParallelDo mPDo = ((MultipleParallelDo) op);
            for (Object entry : mPDo.getDests().entrySet()) {
                Map.Entry<PCollection, DoFn> en = (Map.Entry<PCollection, DoFn>) entry;
                LazyCollection<?> lCol = (LazyCollection<?>) en.getKey();
                DeferredOp childOp = null;
                if (lCol.getDownOps() != null && lCol.getDownOps().size() > 0) {
                    childOp = lCol.getDownOps().get(0);
                }
                final Integer channel;
                if (childOp != null && childOp instanceof Flatten) {
                    channel = mscr.getNumberedChannels().get(((Flatten) childOp).getDest());
                } else if (childOp != null && childOp instanceof GroupByKey) {
                    channel = mscr.getNumberedChannels().get(((GroupByKey) childOp).getOrigin());
                } else {
                    // bypass channel?
                    channel = mscr.getNumberedChannels().get(en.getKey());
                }
                if (channel == null) {
                    // This is not for this MSCR - just skip it
                    return;
                }
                // Call parallelDo function
                en.getValue().process(toProcess, new EmitFn() {

                    @Override
                    public void emit(Object v) {
                        try {
                            if (v instanceof Pair) {
                                Pair p = (Pair) v;
                                context.write(new PlumeObject((WritableComparable) p.getKey(), channel), new PlumeObject((WritableComparable) p.getValue(), channel));
                            } else {
                                context.write(new PlumeObject((WritableComparable) v, channel), new PlumeObject((WritableComparable) v, channel));
                            }
                        } catch (Exception e) {
                            // TODO How to report this
                            e.printStackTrace();
                        }
                    }
                });
            }
        } else {
            if (op instanceof Flatten) {
                l = (LazyCollection) ((Flatten) op).getDest();
            }
            int channel = mscr.getNumberedChannels().get(l);
            if (toProcess instanceof Pair) {
                context.write(new PlumeObject(key, channel), new PlumeObject(value, channel));
            } else {
                context.write(new PlumeObject(value, channel), new PlumeObject(value, channel));
            }
        }
    }
}

Example 4 with Flatten

use of com.tdunning.plume.local.lazy.op.Flatten in project Plume by tdunning.

the class OptimizerTools method getMSCRBlocks.

/**
 * This utility returns all the different MSCR blocks that can be created from this plan
 *
 * (pere) As of Oct/2010, I think this code can be simplified to be more like addRemainingTrivialMSCRs(), so a possible TODO would be
 *  to refactor it and make it more understandable. An opened question is whether there is an easy way of coding finding all possible
 *  MSCRs (including trivial, not related to GroupByKey operations ones) in a single and elegant loop.
 */
@SuppressWarnings({ "rawtypes", "unchecked" })
static Set<MSCR> getMSCRBlocks(List<PCollection> outputs) {
    // Get all GroupByKeys from the tree
    List<DeferredOp> groupBys = OptimizerTools.getAll(outputs, GroupByKey.class);
    int mscrId = 1;
    Set<MSCR> mscrs = new HashSet<MSCR>();
    // For all found GroupByKey blocks
    for (DeferredOp gBK : groupBys) {
        GroupByKey groupBy = (GroupByKey<?, ?>) gBK;
        // Gather all information needed for MSCR from this GBK
        Set<PCollection<?>> inputs = new HashSet<PCollection<?>>();
        Set<GroupByKey<?, ?>> outputChannels = new HashSet<GroupByKey<?, ?>>();
        Set<Flatten<?>> unGroupedOutputChannels = new HashSet<Flatten<?>>();
        Set<PCollection<?>> bypassChannels = new HashSet<PCollection<?>>();
        Stack<LazyCollection<?>> toVisit = new Stack<LazyCollection<?>>();
        Set<LazyCollection<?>> visited = new HashSet<LazyCollection<?>>();
        LazyCollection<?> origin = (LazyCollection<?>) groupBy.getOrigin();
        toVisit.push(origin);
        outputChannels.add(groupBy);
        while (!toVisit.isEmpty()) {
            LazyCollection<?> current = toVisit.pop();
            visited.add(current);
            if (current.isMaterialized()) {
                // condition for being a materialized input. This may change.
                inputs.add(current);
                continue;
            }
            DeferredOp op = current.getDeferredOp();
            if (op instanceof MultipleParallelDo) {
                // second condition for being an input
                MultipleParallelDo<?> mPDo = (MultipleParallelDo) current.getDeferredOp();
                if (((LazyCollection<?>) mPDo.getOrigin()).isMaterialized()) {
                    // will be done in Mapper
                    inputs.add(mPDo.getOrigin());
                } else if (op instanceof ParallelDo) {
                    // will be done in Reducer
                    inputs.add(current);
                } else {
                    // will be done in Mapper
                    inputs.add(mPDo.getOrigin());
                }
                // Check for bypass channels & output channels with no group-by
                for (Map.Entry entry : mPDo.getDests().entrySet()) {
                    LazyCollection coll = (LazyCollection) entry.getKey();
                    if (coll.getDownOps() == null || coll.getDownOps().size() == 0) {
                        // leaf node
                        bypassChannels.add(coll);
                    } else if (coll.getDownOps().get(0) instanceof MultipleParallelDo) {
                        bypassChannels.add(coll);
                    /*
             * Case of an output channel that Flattens with no Group By
             */
                    } else if (coll.getDownOps().get(0) instanceof Flatten) {
                        Flatten<?> thisFlatten = (Flatten<?>) coll.getDownOps().get(0);
                        LazyCollection ldest = (LazyCollection) thisFlatten.getDest();
                        if (ldest.getDownOps() == null || ldest.getDownOps().size() == 0 || ldest.getDownOps().get(0) instanceof MultipleParallelDo) {
                            unGroupedOutputChannels.add(thisFlatten);
                            // Add the rest of this flatten's origins to the stack in order to possibly discover more output channels
                            for (PCollection<?> col : thisFlatten.getOrigins()) {
                                if (!visited.contains(col)) {
                                    toVisit.push((LazyCollection<?>) col);
                                }
                            }
                        }
                    }
                }
                continue;
            }
            if (op instanceof GroupByKey) {
                // third condition for being an input - rare case when one GBK follows another
                inputs.add(current);
                continue;
            }
            if (op instanceof Flatten) {
                Flatten<?> flatten = (Flatten<?>) op;
                for (PCollection<?> input : flatten.getOrigins()) {
                    LazyCollection<?> in = (LazyCollection<?>) input;
                    if (!visited.contains(in)) {
                        toVisit.push(in);
                    }
                }
                continue;
            }
            if (op instanceof OneToOneOp) {
                LazyCollection<?> input = (LazyCollection<?>) ((OneToOneOp<?, ?>) op).getOrigin();
                if (!visited.contains(input)) {
                    toVisit.push(input);
                }
                continue;
            }
        }
        MSCR mscrToAdd = null;
        // Check if there is already one MSCR with at least one of this inputs
        for (MSCR mscr : mscrs) {
            for (PCollection<?> input : inputs) {
                if (mscr.hasInput(input)) {
                    mscrToAdd = mscr;
                    break;
                }
            }
        }
        if (mscrToAdd == null) {
            // otherwise create new MSCR
            mscrToAdd = new MSCR(mscrId);
            mscrId++;
        }
        // Add all missing input channels to current MSCR
        for (PCollection<?> input : inputs) {
            if (!mscrToAdd.hasInput(input)) {
                mscrToAdd.addInput(input);
            }
        }
        // Add all missing bypass outputs to current MSCR
        for (PCollection<?> col : bypassChannels) {
            if (!mscrToAdd.hasOutputChannel(col)) {
                // Create new by-pass channel
                MSCR.OutputChannel oC = new MSCR.OutputChannel(col);
                mscrToAdd.addOutputChannel(oC);
            }
        }
        // Add all missing flatten-with-no-groupby outputs to current MSCR
        for (Flatten flatten : unGroupedOutputChannels) {
            if (!mscrToAdd.hasOutputChannel(flatten.getDest())) {
                // Create new channel with flatten and nothing else
                MSCR.OutputChannel oC = new MSCR.OutputChannel(flatten.getDest());
                oC.output = flatten.getDest();
                oC.flatten = flatten;
                mscrToAdd.addOutputChannel(oC);
            }
        }
        // Add all missing output channels to current MSCR
        for (GroupByKey groupByKey : outputChannels) {
            if (!mscrToAdd.hasOutputChannel(groupByKey.getOrigin())) {
                // Create new channel with group by key. It might have combiner and reducer as well.
                MSCR.OutputChannel oC = new MSCR.OutputChannel(groupByKey);
                oC.output = groupByKey.getDest();
                if (groupByKey.getOrigin().getDeferredOp() instanceof Flatten) {
                    oC.flatten = (Flatten) groupByKey.getOrigin().getDeferredOp();
                }
                if (groupByKey.getDest().getDownOps() != null && groupByKey.getDest().getDownOps().size() == 1) {
                    DeferredOp op = (DeferredOp) groupByKey.getDest().getDownOps().get(0);
                    if (op instanceof CombineValues) {
                        oC.combiner = (CombineValues) op;
                        oC.output = oC.combiner.getDest();
                        LazyCollection dest = (LazyCollection) oC.combiner.getDest();
                        if (dest.getDownOps() != null && dest.getDownOps().size() == 1) {
                            op = (DeferredOp) dest.getDownOps().get(0);
                        }
                    }
                    if (op instanceof ParallelDo) {
                        oC.reducer = (ParallelDo) op;
                        oC.output = oC.reducer.getDest();
                    }
                }
                mscrToAdd.addOutputChannel(oC);
            }
        }
        // Add if needed
        mscrs.add(mscrToAdd);
    }
    return addRemainingTrivialMSCRs(outputs, mscrId, mscrs);
}

Example 5 with Flatten

use of com.tdunning.plume.local.lazy.op.Flatten in project Plume by tdunning.

the class OptimizerTools method addRemainingTrivialMSCRs.

/**
 * This utility returns all the MSCRs that are not related to a GroupByKey -
 *  the remaining trivial cases as described in FlumeJava paper
 *
 *  These cases will be either:
 *  - Flattens that are followed by either a)MultipleParallelDo or b)nothing
 *
 *    (These ones can have correlated inputs and be parallelized just like the ones with GroupByKey)
 *
 *  - The trivial Input->ParalleDo|MultipleParalleDo->Output case
 *
 * @param outputs
 * @return
 */
@SuppressWarnings({ "unchecked", "rawtypes" })
static Set<MSCR> addRemainingTrivialMSCRs(List<PCollection> outputs, int currentMscrId, Set<MSCR> currentMSCRs) {
    // Get all Flatten from the tree
    List<DeferredOp> flattens = OptimizerTools.getAll(outputs, Flatten.class);
    List<MSCR> trivialMSCRS = new LinkedList<MSCR>();
    Iterator<DeferredOp> it = flattens.iterator();
    mainLoop: while (it.hasNext()) {
        Flatten<?> flatten = (Flatten<?>) it.next();
        // Process only remaining flattens that are not in any other MSCR
        for (MSCR mscr : currentMSCRs) {
            for (Map.Entry<PCollection<?>, MSCR.OutputChannel<?, ?, ?>> entry : mscr.getOutputChannels().entrySet()) {
                if (entry.getValue().flatten != null && entry.getValue().flatten == flatten) {
                    // skip this flatten
                    continue mainLoop;
                }
            }
        }
        // Create new trivial MSCR
        MSCR mscr = new MSCR(currentMscrId);
        currentMscrId++;
        // add single output channel
        MSCR.OutputChannel oC = new MSCR.OutputChannel(flatten.getDest());
        oC.output = flatten.getDest();
        oC.flatten = flatten;
        mscr.addOutputChannel(oC);
        // add inputs
        for (PCollection coll : flatten.getOrigins()) {
            LazyCollection lCol = (LazyCollection) coll;
            if (lCol.isMaterialized()) {
                mscr.addInput(coll);
            } else if (lCol.deferredOp instanceof ParallelDo) {
                ParallelDo pDo = (ParallelDo) lCol.deferredOp;
                if (((LazyCollection) pDo.getOrigin()).isMaterialized()) {
                    mscr.addInput(pDo.getOrigin());
                } else if (pDo instanceof MultipleParallelDo) {
                    mscr.addInput(pDo.getOrigin());
                } else {
                    mscr.addInput(coll);
                }
            } else {
                mscr.addInput(coll);
            }
        }
        Iterator<MSCR> tIt = trivialMSCRS.iterator();
        // Now we'll see if this trivial MSCR can be fused to another previous trivial MSCR
        boolean canBeFused = false;
        while (tIt.hasNext() && !canBeFused) {
            MSCR trivialMSCR = tIt.next();
            for (PCollection input : trivialMSCR.getInputs()) {
                if (mscr.getInputs().contains(input)) {
                    canBeFused = true;
                    break;
                }
            }
            if (canBeFused) {
                // add current output channel
                trivialMSCR.addOutputChannel(oC);
                for (PCollection input : mscr.getInputs()) {
                    if (!trivialMSCR.getInputs().contains(input)) {
                        // add each input that is not already contained
                        trivialMSCR.addInput(input);
                    }
                }
            }
        }
        // We have a new trivial MSCR only if it could not be fused with previous ones
        if (!canBeFused) {
            trivialMSCRS.add(mscr);
        }
    }
    currentMSCRs.addAll(trivialMSCRS);
    return currentMSCRs;
}