I have finally gotten around to testing this functionality, and it would
doesn't work. The ALTER table change columns command just changes the
metadata for the table, not for the partition. Follows is exactly what I did
to test this, and the (broken) result:
hive> create table multischema_test (
id int,
crdt string,
name string,
age int
)
partitioned by (dtp string)
row format delimited
fields terminated by '\t'
lines terminated by '\n'
stored as textfile;
OK
Time taken: 0.345 seconds
hive> alter table multischema_test add partition (dtp=20110101) location
'/user/hive/warehouse/test/multischema_test/20110101';
OK
Time taken: 0.662 seconds
hive> alter table multischema_test replace columns (id int, name string,
gender string, age int, crdt string) ;
OK
Time taken: 0.119 seconds
hive> alter table multischema_test add partition (dtp=20110102) location
'/user/hive/warehouse/test/multischema_test/20110102';
OK
Time taken: 0.186 seconds
hive> select * from multischema_test ;
OK
1 2010-07-01 Jeff 32 NULL 20110101
2 2010-07-01 Lisa 33 NULL 20110101
3 2010-07-01 Bob 22 NULL 20110101
4 2010-07-01 Fred 27 NULL 20110101
100 Gregory Male 45 2010-08-01 20110102
101 Horus Male 14 2010-08-01 20110102
102 Verdann Male 33 2010-08-01 20110102
103 Gennefer Female 32 2010-08-01 20110102
Time taken: 0.348 seconds
hive> select name,gender from multischema_test ;
Total MapReduce jobs = 1
Launching Job 1 out of 1
Number of reduce tasks is set to 0 since there's no reduce operator
Starting Job = job_201108291505_28322, Tracking URL =
http://laxhadoop1-001:50030/jobdetails.jsp?jobid=job_201108291505_28322Kill Command = /usr/lib/hadoop/bin/hadoop job
-Dmapred.job.tracker=laxhadoop1-001:54311 -kill job_201108291505_28322
2011-10-06 11:02:27,099 Stage-1 map = 0%, reduce = 0%
2011-10-06 11:02:31,129 Stage-1 map = 100%, reduce = 0%
2011-10-06 11:02:33,144 Stage-1 map = 100%, reduce = 100%
Ended Job = job_201108291505_28322
OK
Gregory Male
Horus Male
Verdann Male
Gennefer Female
2010-07-01 Jeff
2010-07-01 Lisa
2010-07-01 Bob
2010-07-01 Fred
Time taken: 8.977 seconds
hive>
The two text files that make up the data in the table are like this:
[hdfs@laxhadoop1-012 ~/Tim] :) cat multischema-datafile-20110101 | hadoop fs
-put - /user/hive/warehouse/test/multischema_test/20110101/datafile
[hdfs@laxhadoop1-012 ~/Tim] :) cat multischema-datafile-20110102 | hadoop fs
-put - /user/hive/warehouse/test/multischema_test/20110102/datafile
[hdfs@laxhadoop1-012 ~/Tim] :) cat multischema-datafile-20110101
1 2010-07-01 Jeff 32
2 2010-07-01 Lisa 33
3 2010-07-01 Bob 22
4 2010-07-01 Fred 27
[hdfs@laxhadoop1-012 ~/Tim] :) cat multischema-datafile-20110102
100 Gregory Male 45 2010-08-01
101 Horus Male 14 2010-08-01
102 Verdann Male 33 2010-08-01
103 Gennefer Female 32 2010-08-01
Did I do something wrong?
On Mon, Aug 29, 2011 at 10:46 PM, Ashutosh Chauhan wrote:Hi Tim,
I figured that both reading the code and manual. I don't think
its explicitly documented anywhere, so it will be great if you document
this. This page looks right place where this place of information can live.
Thanks for the help in making Hive better.
Ashutosh
On Mon, Aug 29, 2011 at 15:26, Time Less wrote:Hello, Ashutosh,
I did nothing like that... :)
It seems the problem here is I didn't RTFM. Perchance, could you say where
you figured this out? I am going from the Hive DDL page on confluence[1],
and although it mentions partitions and it mentions the "replace columns"
you've mentioned here, it doesn't mention them together that I see. I would
like to document this for future generations. Is that the proper page where
I'd document this?
I would probably explicitly create a section titled "Different Schemas per
Partition" and basically give the syntax you give (from quoted, assuming
after I test it, it works).
[1]
https://cwiki.apache.org/confluence/display/Hive/LanguageManual+DDL#LanguageManualDDL-AlterTable%2FPartitionStatementsOn Wed, Aug 24, 2011 at 6:14 PM, Ashutosh Chauhan wrote:Hey Tim,
Hive does support different schema's for different partitions. If your
data comes out garbled, that seems to be a bug then. In your case, is the
following sequence of steps resemble what you did:
a) create table tbl (id: int, name: string, level: int) partitioned by
date;
b) -- add partitions
c) alter table tbl replace columns (id: int, level: int, name_id: int)
d) -- add more partitions.
If you do select * from tbl, then this should work. You need not to
rewrite any of your data. Can you provide more info about what output you
were expecting and what you got. Are there any error logs?
Ashutosh
On Mon, Aug 22, 2011 at 14:34, Time Less wrote:I found a set of slides from Facebook online about Hive that claims you
can have a schema per partition in the table, this is exciting to us,
because we have a table like so:
id int
name string
level int
date string
And it's broken up into partitions by date. However, on a particular
date last year, the table dramatically changed its schema to:
id int
level int
date string
name_id int
So now if I do "select * from table" in hive, the data is completely
garbled for whichever portion of data doesn't fit the Hive schema. We are
considering re-writing the datafiles so they're the same before/after that
date, but if Hive supports having two entirely different schemas depending
on the partition, that'd be really convenient, since these datafiles are
hundreds of gigabytes in size (and we do sort of like the idea of knowing
how the datafile looked back then...).
This page:
https://cwiki.apache.org/confluence/display/Hive/LanguageManual+DDL#LanguageManualDDL-AlterTable%2FPartitionStatementsdoesn't seem to have an appropriate example, so I'm left wondering.
Has anyone done anything like this?
--
Tim Ellis
Data Architect, Riot Games
--
Tim
--
Tim
