NameNode should exclude excessive replicas when counting live replicas for a block

• Key: HADOOP-4643
• Type: Bug
• Status: Closed
• Resolution: Fixed
• Priority: Major
• Assignee: Hairong Kuang
• Reporter: Hairong Kuang
• Fix Version/s: 0.18.3

Description

Currently NameNode include excessive replicas in blockMap and count them as live replicas. Although excessive replicas have marked as invalid, scheduling deletion may be delayed and also datanode does not send deletion confirmation until the next block report. As a result, excessive replicas may stay in blocksMap for quite a while. This may cause underReplicated blocks undetected in NameNode.

For example, assume that block b is at datanode d1, d2, and d3. We have the following scenario
1. d1 loses heartbeat, NN will replicate b to another datanode, assuming it is d4.
2. d1 comes back. NN finds out b is over-replicated therefore choosing one replica, assuming d4, as a excessive replica and marking it as invalid. Now b has 3 valid replicas d1, d2, d3 and 1 excessive (invalid) replica d4, all in blocksMap.
3. d2 loses heartbeat. d2 gets removed from blocksMap. Block b has 2 valid replicas d1 and d3, and 1 excessive invalid replica d4. Block b is under-replicated; But NN still counts block b has 3 live replicas so does not take any action to replicate block b.

Comments

[Hairong Kuang]

A patch to HADOOP-4540 will fix the problem described in this jira. But while I am working on HADOOP-4540, I realize that the policy change proposed in HADOOP-4540 bring subtle dfs semantics change which needs more discussion. So I plan to mark HADOOP-4540 to be fixed in 0.20 and having a simple solution in this jira to be fixed in 0.18. This jira is a part of the solution to avoid data loss as described in HADOOP-4556.

[Hairong Kuang]

Here is a patch that does not count excess replicas as live replicas.

[Raghu Angadi]

+1. Patch looks good.

[Hairong Kuang]

patch for branch 0.18.

[Hairong Kuang]

Patch for the trunk again.

[Hairong Kuang]

Ant test-core build passed on my local machine.

Ant test-patch result:

[exec] +1 overall.

[exec] +1 @author. The patch does not contain any @author tags.

[exec] +1 tests included. The patch appears to include 10 new ormodified tests.

[exec] +1 javadoc. The javadoc tool did not generate any warningmessages.

[exec] +1 javac. The applied patch does not increase the total number of javac compiler warnings.

[exec] +1 findbugs. The patch does not introduce any new Findbugs warnings.

[exec] +1 Eclipse classpath. The patch retains Eclipse classpath integrity.

[Hairong Kuang]

I just committed this.

[Hudson]

Integrated in Hadoop-trunk #661 (See http://hudson.zones.apache.org/hudson/job/Hadoop-trunk/661/)
. NameNode should exclude excessive replicas when counting live replicas for a block. Contributed by Hairong Kuang.