@Test
public void testDefaultParallelInSkewJoin() throws Throwable {
// default_parallel is considered only at runtime, so here we only test requested parallel
// more thorough tests can be found in TestNumberOfReducers.java
String query =
"a = load 'input';"
+ "b = load 'input';"
+ "c = join a by $0, b by $0 using 'skewed' parallel 100;"
+ "store c into 'output';";
PigServer ps = new PigServer(cluster.getExecType(), cluster.getProperties());
PhysicalPlan pp = Util.buildPp(ps, query);
MROperPlan mrPlan = Util.buildMRPlan(pp, pc);
// Get the skew join job
Iterator<MapReduceOper> iter = mrPlan.getKeys().values().iterator();
int counter = 0;
while (iter.hasNext()) {
MapReduceOper op = iter.next();
counter++;
if (op.isSkewedJoin()) {
assertTrue(op.getRequestedParallelism() == 100);
}
}
assertEquals(3, counter);
pc.defaultParallel = -1;
}
@Test
public void testReducerNumEstimationForOrderBy() throws Exception {
// Skip the test for Tez. Tez use a different mechanism.
// Equivalent test is in TestTezAutoParallelism
Assume.assumeTrue("Skip this test for TEZ", Util.isMapredExecType(cluster.getExecType()));
// use the estimation
pc.getProperties().setProperty("pig.exec.reducers.bytes.per.reducer", "100");
pc.getProperties().setProperty("pig.exec.reducers.max", "10");
String query = "a = load '/passwd';" + "b = order a by $0;" + "store b into 'output';";
PigServer ps = new PigServer(cluster.getExecType(), cluster.getProperties());
PhysicalPlan pp = Util.buildPp(ps, query);
MROperPlan mrPlan = Util.buildMRPlanWithOptimizer(pp, pc);
Configuration conf = ConfigurationUtil.toConfiguration(pc.getProperties());
JobControlCompiler jcc = new JobControlCompiler(pc, conf);
JobControl jobControl = jcc.compile(mrPlan, query);
assertEquals(2, mrPlan.size());
// first job uses a single reducer for the sampling
Util.assertParallelValues(-1, 1, -1, 1, jobControl.getWaitingJobs().get(0).getJobConf());
// Simulate the first job having run so estimation kicks in.
MapReduceOper sort = mrPlan.getLeaves().get(0);
jcc.updateMROpPlan(jobControl.getReadyJobs());
FileLocalizer.create(sort.getQuantFile(), pc);
jobControl = jcc.compile(mrPlan, query);
sort = mrPlan.getLeaves().get(0);
long reducer =
Math.min(
(long) Math.ceil(new File("test/org/apache/pig/test/data/passwd").length() / 100.0),
10);
assertEquals(reducer, sort.getRequestedParallelism());
// the second job estimates reducers
Util.assertParallelValues(
-1, -1, reducer, reducer, jobControl.getWaitingJobs().get(0).getJobConf());
// use the PARALLEL key word, it will override the estimated reducer number
query = "a = load '/passwd';" + "b = order a by $0 PARALLEL 2;" + "store b into 'output';";
pp = Util.buildPp(ps, query);
mrPlan = Util.buildMRPlanWithOptimizer(pp, pc);
assertEquals(2, mrPlan.size());
sort = mrPlan.getLeaves().get(0);
assertEquals(2, sort.getRequestedParallelism());
// the estimation won't take effect when it apply to non-dfs or the files doesn't exist, such as
// hbase
query =
"a = load 'hbase://passwd' using org.apache.pig.backend.hadoop.hbase.HBaseStorage('c:f1 c:f2');"
+ "b = order a by $0 ;"
+ "store b into 'output';";
pp = Util.buildPp(ps, query);
mrPlan = Util.buildMRPlanWithOptimizer(pp, pc);
assertEquals(2, mrPlan.size());
sort = mrPlan.getLeaves().get(0);
// the requested parallel will be -1 if users don't set any of default_parallel, paralllel
// and the estimation doesn't take effect. MR framework will finally set it to 1.
assertEquals(-1, sort.getRequestedParallelism());
// test order by with three jobs (after optimization)
query =
"a = load '/passwd';"
+ "b = foreach a generate $0, $1, $2;"
+ "c = order b by $0;"
+ "store c into 'output';";
pp = Util.buildPp(ps, query);
mrPlan = Util.buildMRPlanWithOptimizer(pp, pc);
assertEquals(3, mrPlan.size());
// Simulate the first 2 jobs having run so estimation kicks in.
sort = mrPlan.getLeaves().get(0);
FileLocalizer.create(sort.getQuantFile(), pc);
jobControl = jcc.compile(mrPlan, query);
Util.copyFromLocalToCluster(
cluster,
"test/org/apache/pig/test/data/passwd",
((POLoad) sort.mapPlan.getRoots().get(0)).getLFile().getFileName());
// First job is just foreach with projection, mapper-only job, so estimate gets ignored
Util.assertParallelValues(-1, -1, -1, 0, jobControl.getWaitingJobs().get(0).getJobConf());
jcc.updateMROpPlan(jobControl.getReadyJobs());
jobControl = jcc.compile(mrPlan, query);
jcc.updateMROpPlan(jobControl.getReadyJobs());
// Second job is a sampler, which requests and gets 1 reducer
Util.assertParallelValues(-1, 1, -1, 1, jobControl.getWaitingJobs().get(0).getJobConf());
jobControl = jcc.compile(mrPlan, query);
sort = mrPlan.getLeaves().get(0);
assertEquals(reducer, sort.getRequestedParallelism());
// Third job is the order, which uses the estimated number of reducers
Util.assertParallelValues(
-1, -1, reducer, reducer, jobControl.getWaitingJobs().get(0).getJobConf());
}