// Tests Single input case for both blocking and non-blocking
// with both map and reduce phases
@Test
public void testSim1() throws Exception {
PhysicalPlan php = new PhysicalPlan();
POLoad ld = GenPhyOp.topLoadOp();
php.add(ld);
PhysicalPlan grpChain1 = GenPhyOp.grpChain();
php.merge(grpChain1);
php.connect(ld, grpChain1.getRoots().get(0));
PhysicalOperator leaf = php.getLeaves().get(0);
PhysicalPlan grpChain2 = GenPhyOp.grpChain();
php.merge(grpChain2);
php.connect(leaf, grpChain2.getRoots().get(0));
leaf = php.getLeaves().get(0);
POFilter fl = GenPhyOp.topFilterOp();
php.add(fl);
php.connect(leaf, fl);
POStore st = GenPhyOp.topStoreOp();
php.add(st);
php.connect(fl, st);
run(php, "test/org/apache/pig/test/data/GoldenFiles/MRC1.gld");
}
/** Configures the Reduce plan, the POPackage operator and the reporter thread */
@SuppressWarnings("unchecked")
@Override
protected void setup(Context context) throws IOException, InterruptedException {
super.setup(context);
inIllustrator = inIllustrator(context);
if (inIllustrator) pack = getPack(context);
Configuration jConf = context.getConfiguration();
SpillableMemoryManager.configure(ConfigurationUtil.toProperties(jConf));
context
.getConfiguration()
.set(
PigConstants.TASK_INDEX,
Integer.toString(context.getTaskAttemptID().getTaskID().getId()));
sJobContext = context;
sJobConfInternal.set(context.getConfiguration());
sJobConf = context.getConfiguration();
try {
PigContext.setPackageImportList(
(ArrayList<String>) ObjectSerializer.deserialize(jConf.get("udf.import.list")));
pigContext = (PigContext) ObjectSerializer.deserialize(jConf.get("pig.pigContext"));
// This attempts to fetch all of the generated code from the distributed cache, and resolve
// it
SchemaTupleBackend.initialize(jConf, pigContext);
if (rp == null)
rp = (PhysicalPlan) ObjectSerializer.deserialize(jConf.get("pig.reducePlan"));
stores = PlanHelper.getPhysicalOperators(rp, POStore.class);
if (!inIllustrator)
pack = (POPackage) ObjectSerializer.deserialize(jConf.get("pig.reduce.package"));
// To be removed
if (rp.isEmpty()) log.debug("Reduce Plan empty!");
else {
ByteArrayOutputStream baos = new ByteArrayOutputStream();
rp.explain(baos);
log.debug(baos.toString());
}
pigReporter = new ProgressableReporter();
if (!(rp.isEmpty())) {
roots = rp.getRoots().toArray(new PhysicalOperator[1]);
leaf = rp.getLeaves().get(0);
}
// Get the UDF specific context
MapRedUtil.setupUDFContext(jConf);
} catch (IOException ioe) {
String msg = "Problem while configuring reduce plan.";
throw new RuntimeException(msg, ioe);
}
log.info(
"Aliases being processed per job phase (AliasName[line,offset]): "
+ jConf.get("pig.alias.location"));
Utils.setDefaultTimeZone(PigMapReduce.sJobConfInternal.get());
}
@Test(expected = MRCompilerException.class)
public void testMRCompilerErr() throws Exception {
String query = "a = load 'input';" + "b = filter a by $0 > 5;" + "store b into 'output';";
PhysicalPlan pp = Util.buildPp(pigServer, query);
pp.remove(pp.getRoots().get(0));
try {
Util.buildMRPlan(new PhysicalPlan(), pc);
} catch (MRCompilerException mrce) {
assertEquals(2053, mrce.getErrorCode());
throw mrce;
}
}
@Test
public void testMergeJoin() throws Exception {
String query =
"a = load '/tmp/input1';"
+ "b = load '/tmp/input2';"
+ "c = join a by $0, b by $0 using 'merge';"
+ "store c into '/tmp/output1';";
PhysicalPlan pp = Util.buildPp(pigServer, query);
MRCompiler comp = new MRCompiler(pp, pc);
comp.compile();
MROperPlan mrp = comp.getMRPlan();
assertTrue(mrp.size() == 2);
MapReduceOper mrOp0 = mrp.getRoots().get(0);
assertTrue(mrOp0.mapPlan.size() == 2);
PhysicalOperator load0 = mrOp0.mapPlan.getRoots().get(0);
MergeJoinIndexer func =
(MergeJoinIndexer)
PigContext.instantiateFuncFromSpec(((POLoad) load0).getLFile().getFuncSpec());
Field lrField = MergeJoinIndexer.class.getDeclaredField("lr");
lrField.setAccessible(true);
POLocalRearrange lr = (POLocalRearrange) lrField.get(func);
List<PhysicalPlan> innerPlans = lr.getPlans();
PhysicalOperator localrearrange0 = mrOp0.mapPlan.getSuccessors(load0).get(0);
assertTrue(localrearrange0 instanceof POLocalRearrange);
assertTrue(mrOp0.reducePlan.size() == 3);
PhysicalOperator pack0 = mrOp0.reducePlan.getRoots().get(0);
assertTrue(pack0 instanceof POPackage);
PhysicalOperator foreach0 = mrOp0.reducePlan.getSuccessors(pack0).get(0);
assertTrue(foreach0 instanceof POForEach);
PhysicalOperator store0 = mrOp0.reducePlan.getSuccessors(foreach0).get(0);
assertTrue(store0 instanceof POStore);
assertTrue(innerPlans.size() == 1);
PhysicalPlan innerPlan = innerPlans.get(0);
assertTrue(innerPlan.size() == 1);
PhysicalOperator project = innerPlan.getRoots().get(0);
assertTrue(project instanceof POProject);
assertTrue(((POProject) project).getColumn() == 0);
MapReduceOper mrOp1 = mrp.getSuccessors(mrOp0).get(0);
assertTrue(mrOp1.mapPlan.size() == 3);
PhysicalOperator load1 = mrOp1.mapPlan.getRoots().get(0);
assertTrue(load1 instanceof POLoad);
PhysicalOperator mergejoin1 = mrOp1.mapPlan.getSuccessors(load1).get(0);
assertTrue(mergejoin1 instanceof POMergeJoin);
PhysicalOperator store1 = mrOp1.mapPlan.getSuccessors(mergejoin1).get(0);
assertTrue(store1 instanceof POStore);
assertTrue(mrOp1.reducePlan.isEmpty());
}
@Test
public void testMRCompilerErr() throws Exception {
planTester.buildPlan("a = load 'input';");
LogicalPlan lp = planTester.buildPlan("b = filter a by $0 > 5;");
PhysicalPlan pp = Util.buildPhysicalPlan(lp, pc);
pp.remove(pp.getRoots().get(0));
try {
Util.buildMRPlan(new PhysicalPlan(), pc);
fail("Expected failure.");
} catch (MRCompilerException mrce) {
assertTrue(mrce.getErrorCode() == 2053);
}
}
public void testDistinct1() throws Exception {
PhysicalPlan php = new PhysicalPlan();
PhysicalPlan ldFil1 = GenPhyOp.loadedFilter();
php.merge(ldFil1);
PODistinct op = new PODistinct(new OperatorKey("", r.nextLong()), -1, null);
php.addAsLeaf(op);
PhysicalPlan grpChain1 = GenPhyOp.grpChain();
php.merge(grpChain1);
php.connect(op, grpChain1.getRoots().get(0));
PODistinct op1 = new PODistinct(new OperatorKey("", r.nextLong()), -1, null);
php.addAsLeaf(op1);
POStore st = GenPhyOp.topStoreOp();
php.addAsLeaf(st);
run(php, "test/org/apache/pig/test/data/GoldenFiles/MRC16.gld");
}
@Test
public void testSortUDF1() throws Exception {
PhysicalPlan php = new PhysicalPlan();
PhysicalPlan ldFil1 = GenPhyOp.loadedFilter();
php.merge(ldFil1);
// set up order by *
String funcName = WeirdComparator.class.getName();
POUserComparisonFunc comparator =
new POUserComparisonFunc(
new OperatorKey("", r.nextLong()), -1, null, new FuncSpec(funcName));
POSort sort =
new POSort(
new OperatorKey("", r.nextLong()),
-1,
ldFil1.getLeaves(),
null,
new ArrayList<Boolean>(),
comparator);
sort.setRequestedParallelism(20);
PhysicalPlan nesSortPlan = new PhysicalPlan();
POProject topPrj = new POProject(new OperatorKey("", r.nextLong()));
topPrj.setColumn(1);
topPrj.setOverloaded(true);
topPrj.setResultType(DataType.TUPLE);
nesSortPlan.add(topPrj);
POProject prjStar2 = new POProject(new OperatorKey("", r.nextLong()));
prjStar2.setResultType(DataType.TUPLE);
prjStar2.setStar(true);
nesSortPlan.add(prjStar2);
nesSortPlan.connect(topPrj, prjStar2);
List<PhysicalPlan> nesSortPlanLst = new ArrayList<PhysicalPlan>();
nesSortPlanLst.add(nesSortPlan);
sort.setSortPlans(nesSortPlanLst);
php.add(sort);
php.connect(ldFil1.getLeaves().get(0), sort);
// have a foreach which takes the sort output
// and send it two two udfs
List<String> udfs = new ArrayList<String>();
udfs.add(COUNT.class.getName());
udfs.add(SUM.class.getName());
POForEach fe3 = GenPhyOp.topForEachOPWithUDF(udfs);
php.add(fe3);
php.connect(sort, fe3);
// add a group above the foreach
PhysicalPlan grpChain1 = GenPhyOp.grpChain();
php.merge(grpChain1);
php.connect(fe3, grpChain1.getRoots().get(0));
udfs.clear();
udfs.add(AVG.class.getName());
POForEach fe4 = GenPhyOp.topForEachOPWithUDF(udfs);
php.addAsLeaf(fe4);
PhysicalPlan grpChain2 = GenPhyOp.grpChain();
php.merge(grpChain2);
php.connect(fe4, grpChain2.getRoots().get(0));
udfs.clear();
udfs.add(GFCross.class.getName() + "('1')");
POForEach fe5 = GenPhyOp.topForEachOPWithUDF(udfs);
php.addAsLeaf(fe5);
POStore st = GenPhyOp.topStoreOp();
php.addAsLeaf(st);
run(php, "test/org/apache/pig/test/data/GoldenFiles/MRC15.gld");
}
/**
* Algebraic functions and distinct in nested plan of a foreach are partially computed in the map
* and combine phase. A new foreach statement with initial and intermediate forms of algebraic
* functions are added to map and combine plans respectively.
*
* <p>If bag portion of group-by result is projected or a non algebraic expression/udf has bag as
* input, combiner will not be used. This is because the use of combiner in such case is likely to
* degrade performance as there will not be much reduction in data size in combine stage to offset
* the cost of the additional number of times (de)serialization is done.
*
* <p>Major areas for enhancement: 1. use of combiner in cogroup 2. queries with order-by, limit
* or sort in a nested foreach after group-by 3. case where group-by is followed by filter that
* has algebraic expression
*/
public static void addCombiner(
PhysicalPlan mapPlan,
PhysicalPlan reducePlan,
PhysicalPlan combinePlan,
CompilationMessageCollector messageCollector,
boolean doMapAgg)
throws VisitorException {
// part one - check if this MR job represents a group-by + foreach. Find
// the POLocalRearrange in the map. I'll need it later.
List<PhysicalOperator> mapLeaves = mapPlan.getLeaves();
if (mapLeaves == null || mapLeaves.size() != 1) {
messageCollector.collect(
"Expected map to have single leaf", MessageType.Warning, PigWarning.MULTI_LEAF_MAP);
return;
}
PhysicalOperator mapLeaf = mapLeaves.get(0);
if (!(mapLeaf instanceof POLocalRearrange)) {
return;
}
POLocalRearrange rearrange = (POLocalRearrange) mapLeaf;
List<PhysicalOperator> reduceRoots = reducePlan.getRoots();
if (reduceRoots.size() != 1) {
messageCollector.collect(
"Expected reduce to have single root", MessageType.Warning, PigWarning.MULTI_ROOT_REDUCE);
return;
}
// I expect that the first root should always be a POPackage. If not, I
// don't know what's going on, so I'm out of here.
PhysicalOperator root = reduceRoots.get(0);
if (!(root instanceof POPackage)) {
messageCollector.collect(
"Expected reduce root to be a POPackage",
MessageType.Warning,
PigWarning.NON_PACKAGE_REDUCE_PLAN_ROOT);
return;
}
POPackage pack = (POPackage) root;
List<PhysicalOperator> packSuccessors = reducePlan.getSuccessors(root);
if (packSuccessors == null || packSuccessors.size() != 1) {
return;
}
PhysicalOperator successor = packSuccessors.get(0);
if (successor instanceof POLimit) {
// POLimit is acceptable, as long has it has a single foreach as
// successor
List<PhysicalOperator> limitSucs = reducePlan.getSuccessors(successor);
if (limitSucs != null && limitSucs.size() == 1 && limitSucs.get(0) instanceof POForEach) {
// the code below will now further examine the foreach
successor = limitSucs.get(0);
}
}
if (successor instanceof POForEach) {
POForEach foreach = (POForEach) successor;
List<PhysicalPlan> feInners = foreach.getInputPlans();
// find algebraic operators and also check if the foreach statement
// is suitable for combiner use
List<Pair<PhysicalOperator, PhysicalPlan>> algebraicOps = findAlgebraicOps(feInners);
if (algebraicOps == null || algebraicOps.size() == 0) {
// the plan is not combinable or there is nothing to combine
// we're done
return;
}
if (combinePlan != null && combinePlan.getRoots().size() != 0) {
messageCollector.collect(
"Wasn't expecting to find anything already " + "in the combiner!",
MessageType.Warning,
PigWarning.NON_EMPTY_COMBINE_PLAN);
return;
}
LOG.info("Choosing to move algebraic foreach to combiner");
try {
// replace PODistinct->Project[*] with distinct udf (which is Algebraic)
for (Pair<PhysicalOperator, PhysicalPlan> op2plan : algebraicOps) {
if (!(op2plan.first instanceof PODistinct)) {
continue;
}
DistinctPatcher distinctPatcher = new DistinctPatcher(op2plan.second);
distinctPatcher.visit();
if (distinctPatcher.getDistinct() == null) {
int errCode = 2073;
String msg =
"Problem with replacing distinct operator with distinct built-in function.";
throw new PlanException(msg, errCode, PigException.BUG);
}
op2plan.first = distinctPatcher.getDistinct();
}
// create new map foreach
POForEach mfe = createForEachWithGrpProj(foreach, rearrange.getKeyType());
Map<PhysicalOperator, Integer> op2newpos = Maps.newHashMap();
Integer pos = 1;
// create plan for each algebraic udf and add as inner plan in map-foreach
for (Pair<PhysicalOperator, PhysicalPlan> op2plan : algebraicOps) {
PhysicalPlan udfPlan = createPlanWithPredecessors(op2plan.first, op2plan.second);
mfe.addInputPlan(udfPlan, false);
op2newpos.put(op2plan.first, pos++);
}
changeFunc(mfe, POUserFunc.INITIAL);
// since we will only be creating SingleTupleBag as input to
// the map foreach, we should flag the POProjects in the map
// foreach inner plans to also use SingleTupleBag
for (PhysicalPlan mpl : mfe.getInputPlans()) {
try {
new fixMapProjects(mpl).visit();
} catch (VisitorException e) {
int errCode = 2089;
String msg = "Unable to flag project operator to use single tuple bag.";
throw new PlanException(msg, errCode, PigException.BUG, e);
}
}
// create new combine foreach
POForEach cfe = createForEachWithGrpProj(foreach, rearrange.getKeyType());
// add algebraic functions with appropriate projection
addAlgebraicFuncToCombineFE(cfe, op2newpos);
changeFunc(cfe, POUserFunc.INTERMEDIATE);
// fix projection and function time for algebraic functions in reduce foreach
for (Pair<PhysicalOperator, PhysicalPlan> op2plan : algebraicOps) {
setProjectInput(op2plan.first, op2plan.second, op2newpos.get(op2plan.first));
byte resultType = op2plan.first.getResultType();
((POUserFunc) op2plan.first).setAlgebraicFunction(POUserFunc.FINAL);
op2plan.first.setResultType(resultType);
}
// we have modified the foreach inner plans - so set them again
// for the foreach so that foreach can do any re-initialization
// around them.
// FIXME - this is a necessary evil right now because the leaves
// are explicitly stored in the POForeach as a list rather than
// computed each time at run time from the plans for
// optimization. Do we want to have the Foreach compute the
// leaves each time and have Java optimize it (will Java
// optimize?)?
mfe.setInputPlans(mfe.getInputPlans());
cfe.setInputPlans(cfe.getInputPlans());
foreach.setInputPlans(foreach.getInputPlans());
// tell POCombinerPackage which fields need projected and which
// placed in bags. First field is simple project rest need to go
// into bags
int numFields = algebraicOps.size() + 1; // algebraic funcs + group key
boolean[] bags = new boolean[numFields];
bags[0] = false;
for (int i = 1; i < numFields; i++) {
bags[i] = true;
}
// Use the POCombiner package in the combine plan
// as it needs to act differently than the regular
// package operator.
CombinerPackager pkgr = new CombinerPackager(pack.getPkgr(), bags);
POPackage combinePack = pack.clone();
combinePack.setPkgr(pkgr);
combinePack.setParentPlan(null);
combinePlan.add(combinePack);
combinePlan.add(cfe);
combinePlan.connect(combinePack, cfe);
// No need to connect projections in cfe to cp, because
// PigCombiner directly attaches output from package to
// root of remaining plan.
POLocalRearrange mlr = getNewRearrange(rearrange);
POPartialAgg mapAgg = null;
if (doMapAgg) {
mapAgg = createPartialAgg(cfe);
}
// A specialized local rearrange operator will replace
// the normal local rearrange in the map plan. This behaves
// like the regular local rearrange in the getNext()
// as far as getting its input and constructing the
// "key" out of the input. It then returns a tuple with
// two fields - the key in the first position and the
// "value" inside a bag in the second position. This output
// format resembles the format out of a Package. This output
// will feed to the map foreach which expects this format.
// If the key field isn't in the project of the combiner or map foreach,
// it is added to the end (This is required so that we can
// set up the inner plan of the new Local Rearrange leaf in the map
// and combine plan to contain just the project of the key).
patchUpMap(mapPlan, getPreCombinerLR(rearrange), mfe, mapAgg, mlr);
POLocalRearrange clr = getNewRearrange(rearrange);
clr.setParentPlan(null);
combinePlan.add(clr);
combinePlan.connect(cfe, clr);
// Change the package operator in the reduce plan to
// be the POCombiner package, as it needs to act
// differently than the regular package operator.
pack.setPkgr(pkgr.clone());
} catch (Exception e) {
int errCode = 2018;
String msg = "Internal error. Unable to introduce the combiner for optimization.";
throw new OptimizerException(msg, errCode, PigException.BUG, e);
}
}
}
/**
* find algebraic operators and also check if the foreach statement is suitable for combiner use
*
* @param feInners inner plans of foreach
* @return null if plan is not combinable, otherwise list of combinable operators
* @throws VisitorException
*/
private static List<Pair<PhysicalOperator, PhysicalPlan>> findAlgebraicOps(
List<PhysicalPlan> feInners) throws VisitorException {
List<Pair<PhysicalOperator, PhysicalPlan>> algebraicOps = Lists.newArrayList();
// check each foreach inner plan
for (PhysicalPlan pplan : feInners) {
// check for presence of non combinable operators
AlgebraicPlanChecker algChecker = new AlgebraicPlanChecker(pplan);
algChecker.visit();
if (algChecker.sawNonAlgebraic) {
return null;
}
// if we found a combinable distinct add that to list
if (algChecker.sawDistinctAgg) {
algebraicOps.add(new Pair<PhysicalOperator, PhysicalPlan>(algChecker.getDistinct(), pplan));
continue;
}
List<PhysicalOperator> roots = pplan.getRoots();
// combinable operators have to be attached to POProject root(s)
// if root does not have a successor that is combinable, the project
// has to be projecting the group column. Otherwise this MR job is
// considered not combinable as we don't want to use combiner for
// cases where this foreach statement is projecting bags (likely to
// bad for performance because of additional (de)serialization
// costs)
for (PhysicalOperator root : roots) {
if (root instanceof ConstantExpression) {
continue;
}
if (!(root instanceof POProject)) {
// how can this happen? - expect root of inner plan to be
// constant or project. not combining it
return null;
}
POProject proj = (POProject) root;
POUserFunc combineUdf = getAlgebraicSuccessor(proj, pplan);
if (combineUdf == null) {
if (proj.isProjectToEnd()) {
// project-star or project to end
// not combinable
return null;
}
// Check to see if this is a projection of the grouping column.
// If so, it will be a projection of col 0
List<Integer> cols = proj.getColumns();
if (cols != null && cols.size() == 1 && cols.get(0) == 0) {
// it is project of grouping column, so the plan is
// still combinable
continue;
} else {
// not combinable
return null;
}
}
// The algebraic udf can have more than one input. Add the udf only once
boolean exist = false;
for (Pair<PhysicalOperator, PhysicalPlan> pair : algebraicOps) {
if (pair.first.equals(combineUdf)) {
exist = true;
break;
}
}
if (!exist) algebraicOps.add(new Pair<PhysicalOperator, PhysicalPlan>(combineUdf, pplan));
}
}
return algebraicOps;
}
