@Test
public void testImplicitSplitInCoGroup2() throws Exception {
// this query is similar to the one reported in JIRA - PIG-537
LogicalPlanTester planTester = new LogicalPlanTester();
planTester.buildPlan("a = load 'file1' using PigStorage(':') as (name:chararray, marks:int);");
planTester.buildPlan(
"b = load 'file2' using PigStorage(':') as (name:chararray, rank:chararray);");
planTester.buildPlan("c = cogroup a by name, b by name;");
planTester.buildPlan("d = foreach c generate group, FLATTEN(a.marks) as newmarks;");
planTester.buildPlan("e = cogroup a by marks, d by newmarks;");
LogicalPlan plan =
planTester.buildPlan("f = foreach e generate group, flatten(a), flatten(d);");
// Set the logical plan values correctly in all the operators
PlanSetter ps = new PlanSetter(plan);
ps.visit();
// run through validator
CompilationMessageCollector collector = new CompilationMessageCollector();
TypeCheckingValidator typeValidator = new TypeCheckingValidator();
typeValidator.validate(plan, collector);
printMessageCollector(collector);
printTypeGraph(plan);
if (collector.hasError()) {
throw new Exception("Error during type checking");
}
// this will run ImplicitSplitInserter
TestLogicalOptimizer.optimizePlan(plan);
// get Schema of leaf and compare:
Schema expectedSchema =
Util.getSchemaFromString(
"grp: int,A::username: chararray,A::marks: int,AB::group: chararray,AB::newmarks: int");
assertTrue(Schema.equals(expectedSchema, plan.getLeaves().get(0).getSchema(), false, true));
}
/* (non-Javadoc)
* @see org.apache.pig.newplan.logical.expression.LogicalExpressionVisitor#visit(org.apache.pig.newplan.logical.expression.CastExpression)
*if input type is bytearray, find and set the corresponding load function
* that this field comes from.
* The load functions LoadCaster interface will help with conversion
* at runtime.
* If there is no corresponding load function (eg if the input is an
* output of a udf), set nothing - it assumes that bytearray is being used
* as equivalent of 'unknown' type. It will try to identify the type
* at runtime and cast it.
*/
@Override
public void visit(CastExpression cast) throws FrontendException {
byte inType = cast.getExpression().getType();
byte outType = cast.getType();
if (containsByteArrayOrEmtpyInSchema(cast.getExpression().getFieldSchema())) {
long inUid = cast.getExpression().getFieldSchema().uid;
FuncSpec inLoadFunc = uid2LoadFuncMap.get(inUid);
if (inLoadFunc == null) {
String msg =
"Cannot resolve load function to use for casting from "
+ DataType.findTypeName(inType)
+ " to "
+ DataType.findTypeName(outType)
+ ". ";
msgCollector.collect(msg, MessageType.Warning);
} else {
cast.setFuncSpec(inLoadFunc);
}
}
}
/**
* 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);
}
}
}
