zhihai xu created YARN-3727: ------------------------------- Summary: For better error recovery, check if the directory exists before using it for localization. Key: YARN-3727 URL: https://issues.apache.org/jira/browse/YARN-3727 Project: Hadoop YARN Issue Type: Improvement Components: nodemanager Affects Versions: 2.7.0 Reporter: zhihai xu Assignee: zhihai xu
For better error recovery, check if the directory exists before using it for localization. We saw the following localization failure happened due to existing cache directories. {code} 2015-05-11 18:59:59,756 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.ResourceLocalizationService: DEBUG: FAILED { hdfs://XXXX/XXXXX/libjars/1234.jar, 1431395961545, FILE, null }, Rename cannot overwrite non empty destination directory /XXXX/8/yarn/nm/usercache/XXXX/filecache/21637 2015-05-11 18:59:59,756 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalizedResource: Resource hdfs://XXXX/XXXXX/libjars/1234.jar(->/XXXX/8/yarn/nm/usercache/XXXX/filecache/21637/1234.jar) transitioned from DOWNLOADING to FAILED {code} The real cause for this failure may be disk failure, LevelDB operation failure for {{startResourceLocalization}}/{{finishResourceLocalization}} or others. I wonder whether we can add error recovery code to avoid the localization failure by not using the existing cache directories for localization. The exception happened at {{files.rename(dst_work, destDirPath, Rename.OVERWRITE)}} in FSDownload#call. Based on the following code, after the exception, the existing cache directory used by {{LocalizedResource}} will be deleted. {{code}} try { ......... files.rename(dst_work, destDirPath, Rename.OVERWRITE); } catch (Exception e) { try { files.delete(destDirPath, true); } catch (IOException ignore) { } throw e; } finally { {{code}} Since the conflicting local directory will be deleted after localization failure, I think it will be better to check if the directory exists before using it for localization to avoid the localization failure. -- This message was sent by Atlassian JIRA (v6.3.4#6332)