hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Daryn Sharp (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-10348) Namenode report bad block method doesn't check whether the block belongs to datanode before adding it to corrupt replicas map.
Date Wed, 04 May 2016 22:06:12 GMT

    [ https://issues.apache.org/jira/browse/HDFS-10348?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15271522#comment-15271522

Daryn Sharp commented on HDFS-10348:

The logic around this code is dicey.  There are some odd edge cases surrounding {{storageInfo#addBlock}}.
 A client can report a corrupt GS that's different than currently in the blocks map.  I question
if the those should be ignored.  Or more importantly, if the client reports a storage different
than the NN expects, the NN decides to move the block to the client reported storage.

I'll file another jira for the seemingly odd behavior of immediately invalidating a block
if reported corrupt with a younger GS, forgetting it's corrupt or even exists on the DN, DN
ignoring the invalidate because the GS is wrong, DN eventually sending BR with the forgotten

> Namenode report bad block method doesn't check whether the block belongs to datanode
before adding it to corrupt replicas map.
> ------------------------------------------------------------------------------------------------------------------------------
>                 Key: HDFS-10348
>                 URL: https://issues.apache.org/jira/browse/HDFS-10348
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: namenode
>    Affects Versions: 2.7.0
>            Reporter: Rushabh S Shah
>            Assignee: Rushabh S Shah
>         Attachments: HDFS-10348-1.patch, HDFS-10348.patch
> Namenode (via report bad block nethod) doesn't check whether the block belongs to the
datanode before it adds to corrupt replicas map.
> In one of our cluster we found that there were 3 lingering corrupt blocks.
> It happened in the following order.
> 1. Two clients called getBlockLocations for a particular file.
> 2. Client C1 tried to open the file and encountered checksum error from   node N3 and
it reported bad block (blk1) to the namenode.
> 3. Namenode added that node N3 and block blk1  to corrrupt replicas map   and ask one
of the good node (one of the 2 nodes) to replicate the block to another node N4.
> 4. After receiving the block, N4 sends an IBR (with RECEIVED_BLOCK) to namenode.
> 5. Namenode removed the block and node N3 from corrupt replicas map.
>    It also removed N3's storage from triplets and queued an invalidate request for N3.
> 6. In the mean time, Client C2 tries to open the file and the request went to node N3.
>    C2 also encountered the checksum exception and reported bad block to namenode.
> 7. Namenode added the corrupt block blk1 and node N3 to the corrupt replicas map without
confirming whether node N3 has the block or not.
> After deleting the block, N3 sends an IBR (with DELETED) and the namenode simply ignores
the report since the N3's storage is no longer in the triplets(from step 5)
> We took the node out of rotation, but still the block was present only in the corruptReplciasMap.

> Since on removing the node, we only goes through the block which are present in the triplets
for a given datanode.
> [~kshukla]'s patch fixed this bug via https://issues.apache.org/jira/browse/HDFS-9958.
> But I think the following check should be made in the BlockManager#markBlockAsCorrupt
instead of BlockManager#findAndMarkBlockAsCorrupt.
> {noformat}
> if (storage == null) {
>       storage = storedBlock.findStorageInfo(node);
>     }
>     if (storage == null) {
>       blockLog.debug("BLOCK* findAndMarkBlockAsCorrupt: {} not found on {}",
>           blk, dn);
>       return;
> }
> {noformat}

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: hdfs-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-help@hadoop.apache.org

View raw message