Synthetic Control actually used to work with all the clustering jobs.
The move to Hadoop 0.19 introduced intermittent problems that depend
upon optimizations done behind the scenes in Hadoop. All of the original
implementations used combiners under the assumption that they would only
run after the mapper and they would run exactly once. These assumptions
changed in 0.19. M-99 fixed K-Means but not Canopy or Mean Shift which
still have these assumptions.
Unfortunately, the combiner seems to run only once and only with the
mappers in the development mode which is used by the build and all the
unit tests. This caused the severity of the semantics change to remain
undetected until recently when users are trying to run clustering on
real Hadoop clusters.
The only solution I can imagine right now is to move the combiner
centroid summation code back into the mappers and have the mappers
output fully combined data during close(). It is not very elegant,
perhaps someone has a better solution in mind. I will take a look at it
tonight after the Hadoop Summit.
Jeff
Adil Aijaz wrote:
Hi folks,
I am new to mahout and I started exploring mahout 0.1 release by
trying to run the kmeans clustering example as described in
http://cwiki.apache.org/MAHOUT/syntheticcontroldata.html
After a bunch of runs where no matter what parameters I specified, the
output never changed I realized that:
1. KMeans was clustering all 600 points of syntheticcontroldata into
one cluster.
2. There is a bug in
examples/src/main/java/org/apache/mahout/clustering/syntheticcontrol/kmeans/Job.java
that called runJob from main function with my provided arguments
transposed. So, my convergenceDelta was interpreted as t1, t1 as t2,
and t2 as convergenceDelta. I will commit a patch as soon as I get
approval for opensource commits from my employer, however, I thought
I'd put it out there in case someone else is going through the same
issue.
As for the more serious issue#1 (kmeans clustering everything into one
cluster), I found that this is because the CanopyClusteringJob was
generating only one canopy. Digging deeper, I found that this problem
was coming from the CanopyCombiner being run in both map & reduce
phases. From there I discovered this post from december 2008:
http://tinyurl.com/l83ff4
which indicates that from hadoop 0.18 onwards the combiner will be run
in both map and reduce which is bad since the CanopyCombiner and
KMeansCombiner assume that they are executed only on map side. Now,
the suggested workaround is specific to hadoop 0.18 and it doesn't
work with mahout-0.1 since it requires hadoop 0.19. This means a code
fix is needed for this issue. From the thread Grant talks about a
patch (MAHOUT-99) that fixes the code but that patch is already part
of mahout-0.1 and so it apparently does not fix the issue.
All that to say, I haven't been able to get the kmeans clustering
example on syntheticdata to work which is a bummer. My questions are:
1) Are there any open jiras on this issue (I didn't find any) ? If no,
should I create one?
2) Any workarounds for now?
Adil
