[PIG-1216] New load store design does not allow Pig to validate inputs and outputs up front

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.7.0
• Fix Version/s: 0.7.0
• Hadoop Flags: Reviewed
• Assignee: Ashutosh Chauhan
• Reporter: Alan Gates
• Created: 03/Feb/10 01:14
• Updated: 14/May/10 06:46
• Resolved: 26/Feb/10 19:35

Description

In Pig 0.6 and before, Pig attempts to verify existence of inputs and non-existence of outputs during parsing to avoid run time failures when inputs don't exist or outputs can't be overwritten. The downside to this was that Pig assumed all inputs and outputs were HDFS files, which made implementation harder for non-HDFS based load and store functions. In the load store redesign (PIG-966) this was delegated to InputFormats and OutputFormats to avoid this problem and to make use of the checks already being done in those implementations. Unfortunately, for Pig Latin scripts that run more then one MR job, this does not work well. MR does not do input/output verification on all the jobs at once. It does them one at a time. So if a Pig Latin script results in 10 MR jobs and the file to store to at the end already exists, the first 9 jobs will be run before the 10th job discovers that the whole thing was doomed from the beginning.

To avoid this a validate call needs to be added to the new LoadFunc and StoreFunc interfaces. Pig needs to pass this method enough information that the load function implementer can delegate to InputFormat.getSplits() and the store function implementer to OutputFormat.checkOutputSpecs() if s/he decides to. Since 90% of all load and store functions use HDFS and PigStorage will also need to, the Pig team should implement a default file existence check on HDFS and make it available as a static method to other Load/Store function implementers.

Issue Links

• blocks: PIG-966 Proposed rework for LoadFunc, StoreFunc, and Slice/r interfaces