[ https://issues.apache.org/jira/browse/HADOOP-16721?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Steve Loughran updated HADOOP-16721: ------------------------------------ Description: h3. race condition in delete/rename overlap If you have multiple threads on a system doing rename operations, then one thread doing a delete(dest/subdir) may delete the last file under a subdir, and, before its listed and recreated any parent dir marker -other threads may conclude there's an empty dest dir and fail. This is most likely on an overloaded system with many threads executing rename operations, as with parallel copying taking place there are many threads to schedule and https connections to pool. h3. failure reporting the classic \{[rename(source, dest)}} operation returns \{{false}} on certain failures, which, while somewhat consistent with the posix APIs, turns out to be useless for identifying the cause of problems. Applications tend to have code which goes {code} if (!fs.rename(src, dest)) throw new IOException("rename failed"); {code} While ultimately the rename/3 call needs to be made public (HADOOP-11452) it would then need a adoption across applications. We can do this in the hadoop modules, but for Hive, Spark etc it will take along time. Proposed: a switch to tell S3A to stop downgrading certain failures (source is dir, dest is file, src==dest, etc) into "false". This can be turned on when trying to diagnose why things like Hive are failing. Production code: trivial * change in rename(), * new option * docs. Test code: * need to clear this option for rename contract tests * need to create a new FS with this set to verify the various failure modes trigger it. If this works we should do the same for ABFS, GCS. Hey, maybe even HDFS was: Improve rename resilience in two ways. h3. parent dir probes allow an option skip the LIST for the parent and just do HEAD object to makes sure it is not a file. h3. failure reporting the classic \{[rename(source, dest)}} operation returns \{{false}} on certain failures, which, while somewhat consistent with the posix APIs, turns out to be useless for identifying the cause of problems. Applications tend to have code which goes {code} if (!fs.rename(src, dest)) throw new IOException("rename failed"); {code} While ultimately the rename/3 call needs to be made public (HADOOP-11452) it would then need a adoption across applications. We can do this in the hadoop modules, but for Hive, Spark etc it will take along time. Proposed: a switch to tell S3A to stop downgrading certain failures (source is dir, dest is file, src==dest, etc) into "false". This can be turned on when trying to diagnose why things like Hive are failing. Production code: trivial * change in rename(), * new option * docs. Test code: * need to clear this option for rename contract tests * need to create a new FS with this set to verify the various failure modes trigger it. If this works we should do the same for ABFS, GCS. Hey, maybe even HDFS > Improve S3A rename resilience > ----------------------------- > > Key: HADOOP-16721 > URL: https://issues.apache.org/jira/browse/HADOOP-16721 > Project: Hadoop Common > Issue Type: Sub-task > Components: fs/s3 > Affects Versions: 3.2.0 > Reporter: Steve Loughran > Assignee: Steve Loughran > Priority: Minor > Labels: pull-request-available > Time Spent: 20m > Remaining Estimate: 0h > > h3. race condition in delete/rename overlap > If you have multiple threads on a system doing rename operations, then one > thread doing a delete(dest/subdir) may delete the last file under a subdir, > and, before its listed and recreated any parent dir marker -other threads may > conclude there's an empty dest dir and fail. > This is most likely on an overloaded system with many threads executing > rename operations, as with parallel copying taking place there are many > threads to schedule and https connections to pool. > h3. failure reporting > the classic \{[rename(source, dest)}} operation returns \{{false}} on certain > failures, which, while somewhat consistent with the posix APIs, turns out to > be useless for identifying the cause of problems. Applications tend to have > code which goes > {code} > if (!fs.rename(src, dest)) throw new IOException("rename failed"); > {code} > While ultimately the rename/3 call needs to be made public (HADOOP-11452) it > would then need a adoption across applications. We can do this in the hadoop > modules, but for Hive, Spark etc it will take along time. > Proposed: a switch to tell S3A to stop downgrading certain failures (source > is dir, dest is file, src==dest, etc) into "false". This can be turned on > when trying to diagnose why things like Hive are failing. > Production code: trivial > * change in rename(), > * new option > * docs. > Test code: > * need to clear this option for rename contract tests > * need to create a new FS with this set to verify the various failure modes > trigger it. > > If this works we should do the same for ABFS, GCS. Hey, maybe even HDFS -- This message was sent by Atlassian Jira (v8.3.4#803005) --------------------------------------------------------------------- To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: common-issues-h...@hadoop.apache.org