HDFS is almost full 90% but data node disks are around 50%

[#1]

hi all

we have ambari cluster version 2.6.1 & HDP version 2.6.4

from the dashboard we can see that HDFS DISK Usage is almost 90%

but all data-node disk are around 90%

so why HDFS show 90% , while datanode disk are only 50%

/dev/sdc                   20G   11G  8.7G  56% /data/sdc
/dev/sde                   20G   11G  8.7G  56% /data/sde
/dev/sdd                   20G   11G  9.0G  55% /data/sdd
/dev/sdb                   20G  8.9G   11G  46% /data/sdb

is it problem of fine-tune ? or else

we also performed re-balance from the ambari GUI but this isn't help

[#2]

@Michael Bronson

The HDFS dashboard metrics widget "HDFS Disk Usage" shows: \The percentage of distributed file system (DFS) used, which is a combination of DFS and non-DFS used.

So can you just put your mouse over the "HDFS Disk Usage" widget and then see what is the different values do you see there for "DFS Used" , "non DFS Used" and Remaining. You should see something like following:

.

[#3]

@Jay , this is what we got: ( so this is like 88% used ) , and regarding my question , how it can be 88% when disk are ~50%

[#4]

@Michael Bronson

What do you see when you run the following command?

# su - hdfs -c " hdfs dfsadmin -report | grep 'DFS Used%'

(OR)

Please check the "DFS Used" shown in the NameNode UI to verify if ambari is shoiwng the same data or different? : http://$ACTIVE_NAMENODE:50070/dfshealth.html#tab-overview

[#5]

@Jay , I want to note that we have 4 datenode machines ( 4 workers machines ) and each worker have 4 disks with 20G

[#6]

@jay , we got the following results

 su - hdfs -c " hdfs dfsadmin -report | grep 'DFS Used%' "
DFS Used%: 87.38%
DFS Used%: 88.48%
DFS Used%: 87.00%
DFS Used%: 84.70%
DFS Used%: 87.93%

[#7]

@Jay another info from my datanode about the disk use

datanode1
47% disk1
64% disk2
48% disk3
49% disk4


datanode2
44%
53%
61%
44%


datanode3
55%
46%
45%
91%


datanode4
63%
45%
49%
46%

[#8]

@Jay do you need other info?

[#9]

@Jay , according to http://$ACTIVE_NAMENODE:50070/dfshealth.html#tab-overview

we get

[#10]

@Michael Bronson

As "HDFS Disk Usage" shows: The percentage of distributed file system (DFS) used, which is a combination of DFS and non-DFS used.

The NameNode commands/UI shows that the DFS Used is around 87.06% and Non DFS Used is 0%

So which is almost same which ambari is showing like almost 88% (DFS + Non DFS Usage) so there seems to be no contradiction to me.

Please let us know what is the value you are expecting.

[#11]

@jay , yes I agree , but how it can be the datanode disk are with capacity of 50% and HDFS show 88% ? , and why we not used all the size of datanode disks ? , I am really not understant that

[#12]

hi

per your request , this is the file

      <name>dfs.datanode.data.dir</name>
      <value>/data/sdb/hadoop/hdfs/data,/data/sdc/hadoop/hdfs/data,/data/sdd/hadoop/hdfs/data,/data/sde/hadoop/hdfs/data</value>
--
      <name>dfs.datanode.data.dir.perm</name>
      <value>750</value>

[#13]

@Michael Bronson

Looks good to me. Just do one more check, what was the config that getting loaded into NN in-memory?

http://<active nn host>:50070/conf

and find it "dfs.datanode.data.dir".

You must share us the logs. No point in going with assumptions. 🙂

[#14]

this is the relevant info from the file ( long file , but I think you want to look on the relevant disks )

<value>
/data/sdb/hadoop/hdfs/data,/data/sdc/hadoop/hdfs/data,/data/sdd/hadoop/hdfs/data,/data/sde/hadoop/hdfs/data
</value>
<source>hdfs-site.xml</source>
</property>


<value>
/data/sdb/hadoop/hdfs/data,/data/sdc/hadoop/hdfs/data,/data/sdd/hadoop/hdfs/data,/data/sde/hadoop/hdfs/data
</value>
<source>hdfs-site.xml</source>
</property>

[#15]

about the logs , please remind me what the logs that you want to look for ?

[#16]

Thanks for the confirmation. I need namenode and datanode log after HDFS service restart.

[#17]

because the logs are huge , do you want to search specific sting in the logs ?

[#18]

You can do tail in namenode and datanode log, also you can redirect output to dummy log file during restart.

#tailf <namenode log> >/tmp/namenode-`hostname`.log

#tailf <datanode log> >/tmp/datanode-`hostname`.log

[#19]

HDFS originally splits and stores the data in blocks. Each block is 64MB or 128MB by default based on your HDFS version. Consider a file which is of size (2MB) is stored in a block. The remaining 62MB (considering the default size to be 64MB) is not used by HDFS. Which means here the HDFS used space is 64MB but the actual hard disk used space is 2MB.

Hope this is relevant to what you are asking.

[#20]

@Rabbit , so in our case based on that 88% is HDFS uses , the only one option to be with more HDFS space is to add adisk in each datanode ? am I right ?

[#21]

Yes possibly @Michael Bronson. You could also check "Trash" files size as @Geoffrey Shelton Okot suggests.

In general it is advised not to store too many small files in HDFS. In general, HDFS is good for storing large files.