[
https://issues.apache.org/jira/browse/DATAFU-98?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15605952#comment-15605952
]
Eyal Allweil commented on DATAFU-98:
Hi Russell.
First of all, I want to apologize for the time it's taken us to get to your
contribution. I think it could be quite useful. Having said that, I wonder if
the current version - without counters - gives us enough of an advantage over
vanilla Pig. I think the following code (modified from your unit test) gives us
nearly the same functionality as the UDF in the patch:
{noformat}
data_in = LOAD 'input' as (val:int);
-- data_in: "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "20"
intermediate_data = FOREACH data_in GENERATE val, (val / 5 * 5) AS binStart;
data_out = FOREACH (GROUP intermediate_data BY binStart) GENERATE group AS
binStart, COUNT(intermediate_data) AS binCount;
-- data_out: (0,5),(5,5),(10,2),(20,1)
{noformat}
Unlike your UDF, missing bins are not included. But while including missing
bins can be useful, I do wonder if a single skewed value can cause problems,
especially with small bin sizes and long values. (as a performance-related
aside, I would try to have FrequencyCounter.toBag() called only in the Final
implementations, instead of the first two stages of the algebraic
implementation, to minimize the data copied).
So it seems to me the current UDF has the advantage of having the missing bins,
and it's obviously more readable and convenient than rewriting the Pig code I
wrote above. Did you (or you, [~andrew.musselman]) run any performance tests?
Maybe the Algebraic implementation runs faster than the vanilla Pig code by
virtue of the combiner use.
Last (but not least!) the version you mentioned with counters sounds like it
could be really great.
> New UDF for Histogram / Frequency counting
> --
>
> Key: DATAFU-98
> URL: https://issues.apache.org/jira/browse/DATAFU-98
> Project: DataFu
> Issue Type: New Feature
>Reporter: Russell Melick
> Attachments: DATAFU-98.patch
>
>
> I was thinking of creating a new UDF to compute histograms / frequency counts
> of input bags. It seems like it would make sense to support ints, longs,
> float, and doubles.
> I tried looking around to see if this was already implemented, but
> ValueHistogram and AggregateWordHistogram were about the only things I found.
> They seem to exist as an example job, and only work for Strings.
> https://hadoop.apache.org/docs/current/api/org/apache/hadoop/mapred/lib/aggregate/ValueHistogram.html
> https://hadoop.apache.org/docs/r1.2.1/api/org/apache/hadoop/examples/AggregateWordHistogram.html
> Should the user specify the bin size or the number of bins? Specifying bin
> size probably makes the implementation simpler since you can bin things
> without having seen all of the data.
> I think it would make sense to implement a version of this that didn't need
> any reducers. It could use counters to keep track of the counts per bin
> without sending any data to a reducer. You would be able to call this
> without a preceding GROUP BY as well.
> Here's my proposal for the two udfs. This assumes the input data is two
> columns, memberId and numConnections.
> {code}
> DEFINE BinnedFrequency datafu.pig.stats.BinnedFrequency('min=0;binSize=50')
> connections = LOAD 'connections' AS memberId, numConnections;
> connectionHistogram = FOREACH (GROUP connections ALL) GENERATE
> BinnedFrequency(connections.numConnections);
> {code}
> The output here would be a bag with the frequency counts
> {code}
> {('0-49', 5), ('50-99', 0), ('100-149', 10)}
> {code}
> {code}
> DEFINE BinnedFrequencyCounter
> datafu.pig.stats.BinnedFrequencyCounter('min=0;binSize=50;name=numConnectionsHistogram')
> connections = LOAD 'connections' AS memberId, numConnections;
> connections = FOREACH connections GENERATE
> BinnedFrequencyCounter(numConnections);
> {code}
> The output here would just be a counter for each bin, all sharing the same
> group of numConnectionsHistogram. It would look something like
> numConnectionsHistogram.'0-49' = 5
> numConnectionsHistogram.'50-99' = 0
> numConnectionsHistogram.'100-149' = 10
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)