[
https://issues.apache.org/jira/browse/PIG-4265?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14197754#comment-14197754
]
liyunzhang_intel commented on PIG-4265:
---------------------------------------
Thanks [~xuefuz]'s comment. I made some mistakes in previous bug description.
After further investigation, i found the problem is not on "Java double
precision problems" of AlgebraicDoubleMathBase.java
but on other issues maybe.
{code}
Ruby_UDFs.pig
a = load 'studenttab10k' using PigStorage() as (name, age:int, gpa:double);
b = group a by name;
c = foreach b generate group, SUM(a.age), SUM(a.gpa);
d = foreach c generate $0, $1, (double)((int)$2*100)/100;
store d into 'RubyUDFs_10_benchmark.out';
{code}
run Ruby_UDFs.pig in spark, the sum(a.gpa) is 16.0 and
(double)((int)$2*100)/100 will be "david steinbeck 266 16.0".
run Ruby_UDFs.pig in mapreduce mode, the sum(a.gpa) is 15.999999999999998 and
(double)((int)$2*100)/100 will be "david steinbeck 266 15.0".
As [~xuefuz] said
{quote}
It's expected that a double value 16 is represented by a system as
15.99999999999 or 16.000000001. The problem seems to be the casting of the
double value to int
{quote}
I don't know why the same code by different execution engines(spark and
mapreduce) on the same os returns different results. I will investigate more.
I will rename the bug title to SUM functions returns different value in spark
and mapreduce engine.
> AlgebraicDoubleMathBase has "Java double precision problems"
> ------------------------------------------------------------
>
> Key: PIG-4265
> URL: https://issues.apache.org/jira/browse/PIG-4265
> Project: Pig
> Issue Type: Bug
> Reporter: liyunzhang_intel
> Assignee: liyunzhang_intel
> Attachments: PIG-4265.patch
>
>
> $PIG_HOME/bin/pig -x local RubyUDFs_10.pig
> #RubyUDFs_10.pig
> a = load 'studenttab10k' using PigStorage() as (name, age:int, gpa:double);
> b = group a by name;
> c = foreach b generate group, SUM(a.age), SUM(a.gpa);
> d = foreach c generate $0, $1, (double)((int)$2*100)/100;
> store d into 'local.output/RubyUDFs_10_benchmark.out';
> the result in RubyUDFs_10.out/part
> #grep "david s" RubyUDFs_10.out/part-r-00000
> david steinbeck 266 15.0
> #grep "david s" studenttab10k
> david steinbeck 21 2.44
> david steinbeck 33 1.17
> david steinbeck 42 1.94
> david steinbeck 42 1.35
> david steinbeck 31 2.77
> david steinbeck 40 2.42
> david steinbeck 57 3.91
> when you sum all the gpa of "david steinbeck" in the file "studenttab10k",
> the result is "16" while the result in RubyUDFs_10.out/part-r-00000 is "15".
> The reason is because double precision problem in
> AlgebraicDoubleMathBase.java.
> It sums all the gpa numbers to 15.999999-(double)((int)15.999999*100)/100 =
> 15.0.
> {code}
> AlgebraicDoubleMathBase.java
> private static Double doWork(Double arg1, Double arg2, KNOWN_OP op) {
> if (arg1 == null) {
> return arg2;
> } else if (arg2 == null) {
> return arg1;
> } else {
> switch (op) {
> case MAX: return Math.max(arg1, arg2);
> case MIN: return Math.min(arg1, arg2);
> case SUM: return arg1+arg2; //this line has "Java BigDecimal
> precision problem"
> default: return null;
> }
> }
> }
> {code}
> The detail Java double precision problem you can refer
> "https://community.oracle.com/thread/2448849?tstart=0";
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
