Hbase | Hbase col qualifier hidden using Hbase shell cmds but visible via hbaserdd spark code - apache-spark

Question

I am stuck in a very odd situation related to Hbase design i would say.
Hbase version >> Version 2.1.0-cdh6.2.1
So, the problem statement is, in Hbase, we have a row in our table.
We perform new insert and then subsequent updates of the same Hbase row, as we receive the data from downstream.
say we received data like below
INSERT of {a=1,b=1,c=1,d=1,rowkey='row1'}
UPDATE of {b=1,c=1,d=1,rowkey='row1'}
and
say the final row is like this in our Hbase table
hbase(main):008:0> get 'test', 'row1'
COLUMN CELL
cf:b timestamp=1288380727188, value=value1
cf:c timestamp=1288380727188, value=value1
cf:d timestamp=1288380727188, value=value1
1 row(s) in 0.0400 seconds
So, cf:a, column qualifier is missing in above data as visible above when fetched via scan or get commands. But as per our ingestion flow/process, it should have been there. So, we are triaging as to where it went or what happened and so on. Still the analysis is in process and we are kind of clueless as to where it is.
Now, cut story short, we have a spark util to read the Hbase table into a Rdd, via
hbasecontext.hbaseRdd API function, convert it into a dataframe and display the tabular data. So, we ran this spark util on the same table to help locate this row and very surprisingly it returned 2 rows for the this same rowkey 'row1', where 1st row was the same as above get/scan row (above data) and the 2nd row had our missing column cf:a (surprising it had the same value which was expected). Say the output dataframe appeared something like below.
rowkey |cf:a |cf:b|cf:c|cf:d
row1 |null | 1 | 1 | 1 >> cf:a col qualifier missing (same as in Hbase shell)
row1 | 1 | 1 | 1 | 1 >> This cf:a was expected
We checked our Hbase table schema as well, so we dont have multiple versions of the cf:a in the describe or we dont do versioning on the table. The schema of the Hbase table describe has
VERSIONS => '1'
Anyways, i am clueless as to how hbaseRdd is able to read that row or missing col qualifier, but the Hbase shell cmds via get, scans does not read the missing col qualifier or row.
Any Hbase expert or suggestions please.
Fyi, i tried Hbase shell cmds as well via get - versions on the row, but it only returns the above get data and not the missing cf:a.
Is the col qualifier cf:a marked for deletion or something like that, which the Hbase shell cmd doesn't show ?
Any help would be appreciated.
Thanks !!

Answer 1

This is a strange problem, which I suspect has to do with puts with the same rowkey having different column qualifiers at different times. However, I just tried to recreate this behaviour and I don't seem to be getting this problem. But I have a regular HBase 2.x build, as opposed to yours.
One option I would recommend to explore the problem more closely is to inspect the HFiles physically, outside of hbase shell. You can use the HBase HFile utility to print the physical key-value content at the HFile level. Obviously try to do this on a small HFile! Don't forget to flush and major-compact your table before you do it though, because HBase stores all updates in memory while it can.
You can launch the utility as below, and it will print all key-values sequentially:
hbase hfile -f hdfs://HDFS-NAMENODE:9000/hbase/data/default/test/29cfaecf083bff2f8aa2289c6a078678/f/09f569670678405a9262c8dfa7af8924 -p --printkv
In the above command, HDFS-NAMENODE is your HDFS server, default is your namespace (assuming you have none), test is your table name, and f is the column family name. You can find out the exact path to your HFiles by using the HDFS browse command recursively:
hdfs dfs -ls /hbase/data

Answer 2

[Updated] We worked with Cloudera and found the issue was due to the Hbase region servers getting overlapped. Cloudera fixed it for us. I dont have the full details how they did it.