pig-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Alex Bain" <ambclo...@gmail.com>
Subject Re: Review Request 16309: PIG-3629 Implement STREAM operator in Tez
Date Fri, 10 Jan 2014 22:35:34 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/16309/
-----------------------------------------------------------

(Updated Jan. 10, 2014, 2:35 p.m.)


Review request for pig, Cheolsoo Park, Daniel Dai, Mark Wagner, and Rohini Palaniswamy.


Changes
-------

Reopening for fixes to Rohini's issues


Bugs: PIG-3629
    https://issues.apache.org/jira/browse/PIG-3629


Repository: pig-git


Description
-------

Implement STREAM operator in Tez - https://issues.apache.org/jira/browse/PIG-3629

In this patch, I do not add resources to pig-misc.jar, I just add them individually. See my
discussion post: https://groups.google.com/forum/#!topic/pig-on-tez/8S80GMKhMaU

Basic Changes:
-Run the PhyPlanSetter and EndOfAllInputSetter to set the parent plan and the end-of-all input
flags necessary for STREAM, just like in MR Pig.
-Add a map to hold plan-specific extra local resources in TezOperPlan.java. These resources
can either come from the user's directory (e.g. SHIP('/home/abain/foo')) or from HDFS (e.g.
CACHE('/user/abain/bar') in HDFS).
-Add the new class TezPOStreamVisitor that assembles all the plan-specific local resources
that get added in TezOperPlan.java.

Resource Manager Changes:
-TezResourcManager resources were previously a map of java.net.URL -> Path in HDFS. Previously,
the URL's were all local files, e.g. file://home/abain/pig-withouthHadoop.jar. However, the
CACHE statement requires that resources already present in HDFS be able to be added as local
resources. Unfortunately java.net.URL does not support hdfs:// URL's, so I changed this data
structure to be a YARN URL instead. I also added methods to the ResourceManager to distinguish
whether you are adding a local resource or a resource already present in HDFS.
-CACHE also supports URL's with fragments at the end, which become a "shortcut" to the name,
e.g. CACHE(/input/big-data-name.gz#data.gz). I changed the resource manager to look for a
fragments and use that as the resource name (if the fragment exist). This results in the symlink
to the resource being created with the fragment name, which is what we want.

Race condition:
-I found a race condition that resulted from reusing the Result object in POSimpleTezLoad.
There are several possible solutions. After discussing in the newsgroup, we decided to change
POSimpleTezLoad for now.
-I also made a small cleanup to PhysicalOperator.java.


Diffs (updated)
-----

  src/org/apache/pig/backend/hadoop/executionengine/tez/TezJobControlCompiler.java 28a110a

  src/org/apache/pig/backend/hadoop/executionengine/tez/TezOperPlan.java 3e6ec7b 
  src/org/apache/pig/backend/hadoop/executionengine/tez/TezPlanContainer.java 7342dab 
  src/org/apache/pig/backend/hadoop/executionengine/tez/TezResourceManager.java e28de47 

Diff: https://reviews.apache.org/r/16309/diff/


Testing
-------

Added a unit test to TestTezCompiler.java
Added a new unit test e2e test to tez.conf with session reuse enabled
Ported three other e2e tests from streaming.conf to tez.conf to increase coverage

ant test-tez passes
ant test-e2e-tez passes
Manually tested with a large subset of tests from streaming.conf (the ones using features
currently supported by Pig-on-Tez), they pass


Thanks,

Alex Bain


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message