[
https://issues.apache.org/jira/browse/PIG-1637?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12915943#action_12915943 ]
Xuefu Zhang commented on PIG-1637:
----------------------------------
+1
Patch looks good, except that we don't have to require that all output expressions in the first foreach contain only simple projection. As long as the output expression in the first foreach that is referenced multiple times in the second foreach contains only simple projection, the merge can proceed. Doing this, the following two loops may be better merged to one.
@@ -93,14 +93,17 @@
// Otherwise, we may do expression calculation more than once, defeat the benefit of this
// optimization
Set<Integer> inputs = new HashSet<Integer>();
+ boolean duplicateInputs = false;
for (Operator op : foreach2.getInnerPlan().getSources()) {
// If the source is not LOInnerLoad, then it must be LOGenerate. This happens when
// the 1st ForEach does not rely on any input of 2nd ForEach
if (op instanceof LOInnerLoad) {
LOInnerLoad innerLoad = (LOInnerLoad)op;
int input = innerLoad.getProjection().getColNum();
- if (inputs.contains(input))
- return false;
+ if (inputs.contains(input)) {
+ duplicateInputs = true;
+ break;
+ }
else
inputs.add(input);
@@ -109,6 +112,27 @@
}
}
+ // Duplicate inputs in the case first foreach only containing LOInnerLoad and
+ // LOGenerate is allowed, and output plan is simple projection
+ if (duplicateInputs) {
+ Iterator<Operator> it1 = foreach1.getInnerPlan().getOperators();
+ while( it1.hasNext() ) {
+ Operator op = it1.next();
+ if(!(op instanceof LOGenerate) && !(op instanceof LOInnerLoad))
+ return false;
+ if (op instanceof LOGenerate) {
+ List<LogicalExpressionPlan> outputPlans = ((LOGenerate)op).getOutputPlans();
+ for (LogicalExpressionPlan outputPlan : outputPlans) {
+ Iterator<Operator> iter = outputPlan.getOperators();
+ while (iter.hasNext()) {
+ if (!(iter.next() instanceof ProjectExpression))
+ return false;
+ }
+ }
+ }
+ }
+ }
Combiner not use because optimizor inserts a foreach between group and algebric function
----------------------------------------------------------------------------------------
Key: PIG-1637
URL:
https://issues.apache.org/jira/browse/PIG-1637Project: Pig
Issue Type: Bug
Affects Versions: 0.8.0
Reporter: Daniel Dai
Assignee: Daniel Dai
Fix For: 0.8.0
Attachments: PIG-1637-1.patch, PIG-1637-2.patch
The following script does not use combiner after new optimization change.
{code}
A = load ':INPATH:/pigmix/page_views' using org.apache.pig.test.udf.storefunc.PigPerformanceLoader()
as (user, action, timespent, query_term, ip_addr, timestamp, estimated_revenue, page_info, page_links);
B = foreach A generate user, (int)timespent as timespent, (double)estimated_revenue as estimated_revenue;
C = group B all;
D = foreach C generate SUM(B.timespent), AVG(B.estimated_revenue);
store D into ':OUTPATH:';
{code}
This is because after group, optimizer detect group key is not used afterward, it add a foreach statement after C. This is how it looks like after optimization:
{code}
A = load ':INPATH:/pigmix/page_views' using org.apache.pig.test.udf.storefunc.PigPerformanceLoader()
as (user, action, timespent, query_term, ip_addr, timestamp, estimated_revenue, page_info, page_links);
B = foreach A generate user, (int)timespent as timespent, (double)estimated_revenue as estimated_revenue;
C = group B all;
C1 = foreach C generate B;
D = foreach C1 generate SUM(B.timespent), AVG(B.estimated_revenue);
store D into ':OUTPATH:';
{code}
That cancel the combiner optimization for D.
The way to solve the issue is to merge the C1 we inserted and D. Currently, we do not merge these two foreach. The reason is that one output of the first foreach (B) is referred twice in D, and currently rule assume after merge, we need to calculate B twice in D. Actually, C1 is only doing projection, no calculation of B. Merging C1 and D will not result calculating B twice. So C1 and D should be merged.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
