[
https://issues.apache.org/jira/browse/HDFS-15655?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17243784#comment-17243784
]
Ayush Saxena commented on HDFS-15655:
-------------------------------------
Thanx [~hadoop_yangyun] for the patch. Looks good overall.
Some comments :
{code:java}
+ public static final String DFS_NAMENODE_HOT_BLOCK_TIME_INTERVAL_KEY
+ = "dfs.namenode.hot.block.time.interval";
{code}
* Shouldn't the config be DFS_BALANCER* similar to
{{DFS_BALANCER_GETBLOCKS_MIN_BLOCK_SIZE_KEY}} which too is is used as part of
the same method, this config too seems quite balancer specific
{code:java}
INodeFile iFile = ((INode)getBlockCollection(blockInfo)).asFile();
{code}
* Any reasons why you aren't typecasting directly to INodeFile?
{code:java}
private boolean isHotBlock(BlockInfo blockInfo, long time) {
1641 INodeFile iFile = ((INode)getBlockCollection(blockInfo)).asFile();
1642 if(iFile == null) {
1643 return false;
{code}
* Get into this block only when the time is greater than 0, Add a if check as
the first line of the method, if the value is set to 0, no need of these
conversions, a value of 0 should set this feature off, and shouldn't impact in
any way.
{code:java}
- <name>dfs.datanode.same-disk-tiering.enabled</name>
- <value>false</value>
+ <name>dfs.namenode.hot.block.time.interval"</name>
+ <value>0</value>
{code}
* Why have you removed {{dfs.datanode.same-disk-tiering.enabled}}, Just add a
new conf without bothering anything else, and if possible please add support
for time units as well for this conf.
Couldn't check the test in detail, but :
{code:java}
+ Thread.sleep(hotInterval);
{code}
* This sleep can cause problems later, can we get away of this sleep and try
some mocking, or if that too isn't the case, then something like
{{GenericTestUtils.waitFor}} should be better than having an explicit sleep.
{code:java}
1705 for(int i = 0; i < pending.size() && totalSize < size; i++)
{
1706 curBlock = pending.get(i);
{code}
* Can you drop a comment here, explaining what it does, Just like if the total
size isn't achieved, trying to add the excluded cold blocks or something like
that.
* Give a check to the Balancer doc as well, Add the appropriate details there
as well.
> Add option to make balancer prefer to get cold blocks
> -----------------------------------------------------
>
> Key: HDFS-15655
> URL: https://issues.apache.org/jira/browse/HDFS-15655
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: balancer & mover
> Reporter: Yang Yun
> Assignee: Yang Yun
> Priority: Minor
> Attachments: HDFS-15655.001.patch, HDFS-15655.002.patch,
> HDFS-15655.003.patch, HDFS-15655.004.patch
>
>
> We met two issues when using balancer.
> # Moving hot files may cause failing of dfsclient reading.
> # Some blocks of temporary files are moved and they are deleted soon.
> Add a config key " dfs.namenode.hot.block.time.interval", the balancer
> prefer to get the blocks which are belong to the cold files created before
> this time period.
> Also add a option "-hotBlockTimeInterval" to the command line of balancer for
> inputting.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
