[
https://issues.apache.org/jira/browse/HDFS-9949?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Harsh J updated HDFS-9949:
--------------------------
Description:
In the following scenario, in releases without HDFS-8211, the DN may regenerate
its UUIDs unintentionally.
0. Consider a DN with two disks {{/data1/dfs/dn,/data2/dfs/dn}}
1. Stop DN
2. Unmount the second disk, {{/data2/dfs/dn}}
3. Create (in the scenario, this was an accident) /data2/dfs/dn on the root path
4. Start DN
5. DN now considers /data2/dfs/dn empty so formats it, but during the format it
uses {{datanode.getDatanodeUuid()}} which is null until register() is called.
6. As a result, after the directory loading, {{datanode.checkDatanodUuid()}}
gets called with successful condition, and it causes a new generation of UUID
which is written to all disks {{/data1/dfs/dn/current/VERSION}} and
{{/data2/dfs/dn/current/VERSION}}.
7. Stop DN (in the scenario, this was when the mistake of unmounted disk was
realised)
8. Mount the second disk back again {{/data2/dfs/dn}}, causing the {{VERSION}}
file to be the original one again on it (mounting masks the root path that we
last generated upon).
9. DN fails to start up cause it finds mismatched UUID between the two disks,
with an error similar to:
{code}WARN org.apache.hadoop.hdfs.server.common.Storage:
{{org.apache.hadoop.hdfs.server.common.InconsistentFSStateException: Directory
/data/2/dfs/dn is in an inconsistent state: Root /data/2/dfs/dn:
DatanodeUuid=fe3a848f-beb8-4fcb-9581-c6fb1c701cc4, does not match
8ea9493c-7097-4ee3-96a3-0cc4dfc1d6ac from other StorageDirectory.{code}
The DN should not generate a new UUID if one of the storage disks already have
the older one.
HDFS-8211 unintentionally fixes this by changing the
{{datanode.getDatanodeUuid()}} function to rely on the {{DataStorage}}
representation of the UUID vs. the {{DatanodeID}} object which only gets
available (non-null) _after_ the registration.
It'd still be good to add a direct test case to the above scenario that passes
on trunk and branch-2, but fails on branch-2.7 and lower, so we can catch a
regression around this in future.
was:
In the following scenario, in releases without HDFS-8211, the DN may regenerate
its UUIDs unintentionally.
0. Consider a DN with two disks {{/data1/dfs/dn,/data2/dfs/dn}}
1. Stop DN
2. Unmount the second disk, {{/data2/dfs/dn}}
3. Create (in the scenario, this was an accident) /data2/dfs/dn on the root path
4. Start DN
5. DN now considers /data2/dfs/dn empty so formats it, but during the format it
uses {{datanode.getDatanodeUuid()}} which is null until register() is called.
6. As a result, after the directory loading, {{datanode.checkDatanodUuid()}}
gets called with successful condition, and it causes a new generation of UUID
which is written to all disks {{/data1/dfs/dn/current/VERSION}} and
{{/data2/dfs/dn/current/VERSION}}.
7. Stop DN (in the scenario, this was when the mistake of unmounted disk was
realised)
8. Mount the second disk back again {{/data2/dfs/dn}}, causing the {{VERSION}}
file to be the original one again on it (mounting masks the root path that we
last generated upon).
9. DN fails to start up cause it finds mismatched UUID between the two disks
The DN should not generate a new UUID if one of the storage disks already have
the older one.
HDFS-8211 unintentionally fixes this by changing the
{{datanode.getDatanodeUuid()}} function to rely on the {{DataStorage}}
representation of the UUID vs. the {{DatanodeID}} object which only gets
available (non-null) _after_ the registration.
It'd still be good to add a direct test case to the above scenario that passes
on trunk and branch-2, but fails on branch-2.7 and lower, so we can catch a
regression around this in future.
> Testcase for catching DN UUID regeneration regression
> -----------------------------------------------------
>
> Key: HDFS-9949
> URL: https://issues.apache.org/jira/browse/HDFS-9949
> Project: Hadoop HDFS
> Issue Type: Test
> Affects Versions: 2.6.0
> Reporter: Harsh J
> Assignee: Harsh J
> Priority: Minor
> Attachments: HDFS-9949.000.branch-2.7.not-for-commit.patch,
> HDFS-9949.000.patch, HDFS-9949.001.patch
>
>
> In the following scenario, in releases without HDFS-8211, the DN may
> regenerate its UUIDs unintentionally.
> 0. Consider a DN with two disks {{/data1/dfs/dn,/data2/dfs/dn}}
> 1. Stop DN
> 2. Unmount the second disk, {{/data2/dfs/dn}}
> 3. Create (in the scenario, this was an accident) /data2/dfs/dn on the root
> path
> 4. Start DN
> 5. DN now considers /data2/dfs/dn empty so formats it, but during the format
> it uses {{datanode.getDatanodeUuid()}} which is null until register() is
> called.
> 6. As a result, after the directory loading, {{datanode.checkDatanodUuid()}}
> gets called with successful condition, and it causes a new generation of UUID
> which is written to all disks {{/data1/dfs/dn/current/VERSION}} and
> {{/data2/dfs/dn/current/VERSION}}.
> 7. Stop DN (in the scenario, this was when the mistake of unmounted disk was
> realised)
> 8. Mount the second disk back again {{/data2/dfs/dn}}, causing the
> {{VERSION}} file to be the original one again on it (mounting masks the root
> path that we last generated upon).
> 9. DN fails to start up cause it finds mismatched UUID between the two disks,
> with an error similar to:
> {code}WARN org.apache.hadoop.hdfs.server.common.Storage:
> {{org.apache.hadoop.hdfs.server.common.InconsistentFSStateException:
> Directory /data/2/dfs/dn is in an inconsistent state: Root /data/2/dfs/dn:
> DatanodeUuid=fe3a848f-beb8-4fcb-9581-c6fb1c701cc4, does not match
> 8ea9493c-7097-4ee3-96a3-0cc4dfc1d6ac from other StorageDirectory.{code}
> The DN should not generate a new UUID if one of the storage disks already
> have the older one.
> HDFS-8211 unintentionally fixes this by changing the
> {{datanode.getDatanodeUuid()}} function to rely on the {{DataStorage}}
> representation of the UUID vs. the {{DatanodeID}} object which only gets
> available (non-null) _after_ the registration.
> It'd still be good to add a direct test case to the above scenario that
> passes on trunk and branch-2, but fails on branch-2.7 and lower, so we can
> catch a regression around this in future.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
