[
https://issues.apache.org/jira/browse/DATAFU-119?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15500764#comment-15500764
]
Eyal Allweil commented on DATAFU-119:
-------------------------------------
I've run it on results that were in the tens of millions. I think the main
reason for using it / including it in DataFu is that if you're developing Pig
code, and running it on a cluster (or on any given environment), being able to
stay in the Pig ecosystem is convenient for fast development cycles. If your
original job can run on the given environment, a comparison job can run their
efficiently, too. And there's less copying because you leave the previous
results in the hdfs under a different name, and compare easily.
The output is human-readable, but the expected results is that most records
return null, because they're identical, and the ones that do come out are
usually edge cases that turned out different.
That's the reasoning behind having "something" like this UDF. The output type
and it's not having a schema is a different story - it would be better to have
a schema. But I'm hesitant to spend the time to do it if it isn't likely that
someone else will want to write a different output format for it.
> New UDF - TupleDiff
> -------------------
>
> Key: DATAFU-119
> URL: https://issues.apache.org/jira/browse/DATAFU-119
> Project: DataFu
> Issue Type: New Feature
> Reporter: Eyal Allweil
> Assignee: Eyal Allweil
>
> A UDF that given two tuples, prints out the differences between them in
> human-readable form. This is not meant for production - we use it in PayPal
> for regression tests, to compare the results of two runs. Differences are
> calculated based on position, but the tuples' schemas are used, if available,
> for displaying more friendly results. If no schema is available the output
> uses field numbers.
> It should be used when you want a more fine-grained description of what has
> changed, unlike
> [org.apache.pig.builtin.DIFF|https://pig.apache.org/docs/r0.14.0/func.html#diff].
> Also, because DIFF takes as its input two bags to be compared, they must fit
> in memory. This UDF only takes one pair of tuples at a time, so it can run on
> large inputs.
> We use a macro much like the following in conjunction with this UDF:
> {noformat}
> DEFINE diff_macro(diff_macro_old, diff_macro_new, diff_macro_pk,
> diff_macro_ignored_field) returns diffs {
> DEFINE TupleDiff datafu.pig.util.TupleDiff;
>
> old = FOREACH $diff_macro_old GENERATE $diff_macro_pk, TOTUPLE(*) AS
> original;
> new = FOREACH $diff_macro_new GENERATE $diff_macro_pk, TOTUPLE(*) AS
> original;
>
> join_data = JOIN new BY $diff_macro_pk full, old BY $diff_macro_pk;
>
> join_data = FOREACH join_data GENERATE TupleDiff(old::original,
> new::original, '$diff_macro_ignored_field') AS tupleDiff, old::original,
> new::original;
>
> $diffs = FILTER join_data BY tupleDiff IS NOT NULL ;
> };
> {noformat}
> Currently, the output from the macro looks like this (when comma-separated):
> {noformat}
> added,<original tuple>,
> missing,,<new tuple>
> changed field2 field4,<original tuple>,<new tuple>
> {noformat}
> The UDF takes a variable number of parameters - the two tuples to be
> compared, and any number of field names or numbers to be ignored. We use this
> to ignore fields representing execution or creation time (the macro I've
> given as an example assumes only one ignored field)
> The current implementation "drills down" into tuples, but not bags or maps -
> tuple boundaries are indicated with parentheses, like this:
> {noformat}
> changed outerEmbeddedTuple(innerEmbeddedTuple(fieldNameThatIsDifferent)
> innerEmbeddedTuple(anotherFieldThatIsDifferent))
> {noformat}
> I have a few final things left to do and then I'll put it up on reviewboard.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
