hadoop mapreduce job on Cassandra result is wrong - hadoop

Question

I have 7 node cassandra (1.1.1) and hadoop (1.03) cluster ( tasktracker install same on every cassandra node).
and my column family use wide row pattern. 1 row contains about 200k columns (max about 300k).
My problem is when we use Hadoop to run analytic jobs ( count numbers of occurrence of a word) the result i received is wrong ( result is too lower as I expected in test records)
there 's one strange when we monitoring on job tracker is map progress task indicate wrong ( in my image below ) , And number of "Map input records" when i rerun job ( same data) is not same.
here is my init job code:
Job job = new Job(conf);
job.setJobName(this.jobname);
job.setJarByClass(BannerCount.class);
job.setMapperClass(BannerViewMapper.class);
job.setReducerClass(BannerClickReducer.class);
FileSystem fs = FileSystem.get(conf);
ConfigHelper.setInputRpcPort(job.getConfiguration(), "9160");
ConfigHelper.setInputInitialAddress(job.getConfiguration(), "192.168.23.114,192.168.23.115,192.168.23.116,192.168.23.117,192.168.23.121,192.168.23.122,192.168.23.123");
ConfigHelper.setInputPartitioner(job.getConfiguration(), "org.apache.cassandra.dht.RandomPartitioner");
ConfigHelper.setInputColumnFamily(job.getConfiguration(), KEYSPACE, COLUMN_FAMILY, true);
ConfigHelper.setRangeBatchSize(job.getConfiguration(), 500);
SlicePredicate predicate = new SlicePredicate();
SliceRange sliceRange = new SliceRange();
sliceRange.setStart(ByteBufferUtil.EMPTY_BYTE_BUFFER);
sliceRange.setFinish(ByteBufferUtil.EMPTY_BYTE_BUFFER);
sliceRange.setCount(200000);
predicate.setSlice_range(sliceRange);
ConfigHelper.setInputSlicePredicate(job.getConfiguration(), predicate);
String outPathString = "BannerViewResultV3" + COLUMN_FAMILY;
if (fs.exists(new Path(outPathString)))
fs.delete(new Path(outPathString), true);
FileOutputFormat.setOutputPath(job, new Path(outPathString));
job.setInputFormatClass(ColumnFamilyInputFormat.class);
job.setOutputFormatClass(TextOutputFormat.class);
job.setMapOutputKeyClass(Text.class);
job.setMapOutputValueClass(LongWritable.class);
job.setOutputKeyClass(Text.class);
job.setOutputValueClass(LongWritable.class);
job.setNumReduceTasks(28);
job.waitForCompletion(true);
return 1;