[PIG-733] Order by sampling dumps entire sample to hdfs which causes dfs "FileSystem closed" error on large input

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.2.0
• Fix Version/s: 0.3.0
• Assignee: Pradeep Kamath
• Reporter: Pradeep Kamath
• Created: 25/Mar/09 19:04
• Updated: 24/Mar/10 22:10
• Resolved: 10/Apr/09 02:36

Description

Order by has a sampling job which samples the input and creates a sorted list of sample items. CUrrently the number of items sampled is 100 per map task. So if the input is large resulting in many maps (say 50,000) the sample is big. This sorted sample is stored on dfs. The WeightedRangePartitioner computes quantile boundaries and weighted probabilities for repeating values in each map by reading the samples file from DFS. In queries with many maps (in the order of 50,000) the dfs read of the sample file fails with "FileSystem closed" error. This seems to point to a dfs issue wherein a big dfs file being read simultaneously by many dfs clients (in this case all maps) causes the clients to be closed. However on the pig side, loading the sample from each map in the final map reduce job and computing the quantile boundaries and weighted probabilities is inefficient. We should do this computation through a FindQuantiles udf in the same map reduce job which produces the sorted samples. This way lesser data is written to dfs and in the final map reduce job, the weightedRangePartitioner needs to just load the computed information.