[ https://issues.apache.org/jira/browse/HBASE-5210?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13190537#comment-13190537 ]
Lawrence Simpson commented on HBASE-5210: ----------------------------------------- @Andrew: I looked at your suggested patch. It may not work as well as you hope since it depends on System.nanoTime() changing rapidly which it may not do on all systems. There are discussions in other blogs about this. I believe that java.util.Random uses System.nanoTime() to do default seeding, and it has not been terribly successful in my case. @Zhiyong: I believe that you are correct in that this is the appropriate place to make a change. However, I did not suggest this as a MAPREDUCE change because the existing behavior may be correct for applications other than generating HFiles. I can imagine situations in which one would want the latest version of a file produced by several reducers to be the only one left at the end of a map/reduce job. However, it's definitely not appropriate when producing HFiles for an incremental load. My own solution which I am testing now is to clone FileOutputCommitter and add logic to the moveTaskOutputs() method that creates a new name for any conflicting files. FileOutputCommitter had too many components that were private to make it easy to subclass. I subclassed FileOutputFormat to use the new output committer class and then I used that subclass in my map/reduce job. I included a logging statement in the new moveTaskOutputs() method so that I can tell when the rename logic is triggered. It may take awhile to see if the logic is successful since the two occurrences that I was able to track down occurred two months apart in a job that runs nightly. > HFiles are missing from an incremental load > ------------------------------------------- > > Key: HBASE-5210 > URL: https://issues.apache.org/jira/browse/HBASE-5210 > Project: HBase > Issue Type: Bug > Components: mapreduce > Affects Versions: 0.90.2 > Environment: HBase 0.90.2 with Hadoop-0.20.2 (with durable sync). > RHEL 2.6.18-164.15.1.el5. 4 node cluster (1 master, 3 slaves) > Reporter: Lawrence Simpson > Attachments: HBASE-5210-crazy-new-getRandomFilename.patch > > > We run an overnight map/reduce job that loads data from an external source > and adds that data to an existing HBase table. The input files have been > loaded into hdfs. The map/reduce job uses the HFileOutputFormat (and the > TotalOrderPartitioner) to create HFiles which are subsequently added to the > HBase table. On at least two separate occasions (that we know of), a range > of output would be missing for a given day. The range of keys for the > missing values corresponded to those of a particular region. This implied > that a complete HFile somehow went missing from the job. Further > investigation revealed the following: > * Two different reducers (running in separate JVMs and thus separate class > loaders) > * in the same server can end up using the same file names for their > * HFiles. The scenario is as follows: > * 1. Both reducers start near the same time. > * 2. The first reducer reaches the point where it wants to write its > first file. > * 3. It uses the StoreFile class which contains a static Random > object > * which is initialized by default using a timestamp. > * 4. The file name is generated using the random number generator. > * 5. The file name is checked against other existing files. > * 6. The file is written into temporary files in a directory named > * after the reducer attempt. > * 7. The second reduce task reaches the same point, but its > StoreClass > * (which is now in the file system's cache) gets loaded within the > * time resolution of the OS and thus initializes its Random() > * object with the same seed as the first task. > * 8. The second task also checks for an existing file with the name > * generated by the random number generator and finds no conflict > * because each task is writing files in its own temporary folder. > * 9. The first task finishes and gets its temporary files committed > * to the "real" folder specified for output of the HFiles. > * 10. The second task then reaches its own conclusion and commits its > * files (moveTaskOutputs). The released Hadoop code just > overwrites > * any files with the same name. No warning messages or anything. > * The first task's HFiles just go missing. > * > * Note: The reducers here are NOT different attempts at the same > * reduce task. They are different reduce tasks so data is > * really lost. > I am currently testing a fix in which I have added code to the Hadoop > FileOutputCommitter.moveTaskOutputs method to check for a conflict with > an existing file in the final output folder and to rename the HFile if > needed. This may not be appropriate for all uses of FileOutputFormat. > So I have put this into a new class which is then used by a subclass of > HFileOutputFormat. Subclassing of FileOutputCommitter itself was a bit > more of a problem due to private declarations. > I don't know if my approach is the best fix for the problem. If someone > more knowledgeable than myself deems that it is, I will be happy to share > what I have done and by that time I may have some information on the > results. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira