hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Wang (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-10530) BlockManager reconstruction work scheduling should correctly adhere to EC block placement policy
Date Thu, 16 Mar 2017 01:56:41 GMT

    [ https://issues.apache.org/jira/browse/HDFS-10530?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15927311#comment-15927311
] 

Andrew Wang commented on HDFS-10530:
------------------------------------

Hi Manoj, thanks for working on this, hoping you can help me understand the current behavior,
also a few review comments too:

In the unit test, looks like we start a cluster with 6 racks, write a 6+3 file, add 3 more
hosts, then wait for parity blocks to be written to these 3 new nodes. 

* Are these necessarily the parity blocks, or could they be any of the blocks that are co-located
on the first 6 racks?
* Also, does this happen via EC reconstruction, or do we simply copy the blocks over to the
new racks? If it's the first, we should file a follow-on to do the second when possible.
* Is the BPP violated before entering the waitFor? If so we should assert that. This may require
pausing reconstruction work and resuming later.

A few other notes:

* "block recovery" refers to specifically the process of recovering from a DN failure while
writing. I'm guessing this is actually more like "replication" or since it's EC, "reconstruction"
work?
* Do you think TestBPPRackFaultTolerant needs any additional unit tests along these lines?

{code}
      cluster.startDataNodes(conf, 3, true, null, null,
          new String[]{"host3", "host4", "host5"}, null);
{code}

Looks like these have the same names as the initial DNs as Takanobu noted. Might be nice to
specify the racks too to be explicit.

{code}
      lb = DFSTestUtil.getAllBlocks(dfs, testFileUnsatisfied).get(0);
      blockInfo = bm.getStoredBlock(lb.getBlock().getLocalBlock());
      dumpStoragesFor(blockInfo);
      // But, there will not be any block placement changes yet
      assertFalse("Block group of testFileUnsatisfied should not be placement" +
          " policy satisfied", bm.isPlacementPolicySatisfied(blockInfo));
{code}

If we later enhance the NN to automatically fix up misplaced EC blocks, this assert will be
flaky. Maybe add a comment? Or we could pause reconstruction work, add the DNs, and resume
after the assert.

> BlockManager reconstruction work scheduling should correctly adhere to EC block placement
policy
> ------------------------------------------------------------------------------------------------
>
>                 Key: HDFS-10530
>                 URL: https://issues.apache.org/jira/browse/HDFS-10530
>             Project: Hadoop HDFS
>          Issue Type: Sub-task
>          Components: namenode
>            Reporter: Rui Gao
>            Assignee: Manoj Govindassamy
>              Labels: hdfs-ec-3.0-nice-to-have
>         Attachments: HDFS-10530.1.patch, HDFS-10530.2.patch, HDFS-10530.3.patch, HDFS-10530.4.patch
>
>
> This issue was found by [~tfukudom].
> Under RS-DEFAULT-6-3-64k EC policy, 
> 1. Create an EC file, the file was witten to all the 5 racks( 2 dns for each) of the
cluster.
> 2. Reconstruction work would be scheduled if the 6th rack is added. 
> 3. While adding the 7th rack or more racks will not trigger reconstruction work. 
> Based on default EC block placement policy defined in “BlockPlacementPolicyRackFaultTolerant.java”,
EC file should be able to be scheduled to distribute to 9 racks if possible.
> In *BlockManager#isPlacementPolicySatisfied(BlockInfo storedBlock)* , *numReplicas* of
striped blocks might should be *getRealTotalBlockNum()*, instead of *getRealDataBlockNum()*.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

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


Mime
View raw message