DISTRIBUTE BY and CLUSTER BY didn't resolve all the issues I've seen
with very large data sets. I mean I'm loading a couple terabytes in a
dataset and running into some rather interesting problems. I noticed
however loading a couple months or two at a time (and making sure they
were from the same time period) seem to resolve the problems I kept
hitting over and over again.
I have to keep reminding myself that hive / hadoop isn't a database and
not to treat it as such. :-)
On 08/14/2011 10:15 AM, bejoy...@yahoo.com wrote:
Ya I very much agree with you on those lines. Using the basic stuff
would literally run into memory issues with large datasets. I had some
of those resolved by using the DISTRIBUTE BY clause and so. In short a
little work around over your hive queries could help you out in some
cases.
Regards
Bejoy K S
------------------------------------------------------------------------
*From: * hadoopman <hadoop...@gmail.com>
*Date: *Sun, 14 Aug 2011 08:57:12 -0600
*To: *<user@hive.apache.org>
*ReplyTo: * user@hive.apache.org
*Subject: *Re: how to load data to partitioned table
Something else I've noticed is when loading LOTS of historical data,
if you can try to say load a month of data at a time, try to just load
THAT month of data and only that month. I've been able to load
several years of data (depending on the data) at a single load however
there have been times when loading a large dataset that I would run
into memory issues during the reduce phase (usually during
shuffle/sort). Things from out of memory to stack overflow messages
(I've compiled a list of the more fun ones).
Then I noticed that only loading data from say a single month loaded
quickly and without the memory headaches during the reduce.
Something to keep in mind and it works great!
On 08/12/2011 07:58 AM, bejoy...@yahoo.com wrote:
Hi Daniel
Just having a look at your requirement , to load data into a
partition based hive table from any input file the most hassle free
approach would be.
1. Load the data into a non partitioned table that shares similar
structure as the target table.
2. Populate the target table with the data from non partitioned one
using hive dynamic partition
approach.
With Dynamic partitions you don't need to manually identify the data
partitions and distribute data accordingly.
A similar implementation is described in the blog post
www.kickstarthadoop.blogspot.com/2011/06/how-to-speed-up-your-hive-queries-in.html
Hope it helps
Regards
Bejoy K S
------------------------------------------------------------------------
*From: * Vikas Srivastava <vikas.srivast...@one97.net>
*Date: *Fri, 12 Aug 2011 17:31:28 +0530
*To: *<user@hive.apache.org>
*ReplyTo: * user@hive.apache.org
*Subject: *Re: how to load data to partitioned table
Hey ,
Simpley you have run query like this
FROM sales_temp INSERT OVERWRITE TABLE sales partition(period_key)
SELECT *
Regards
Vikas Srivastava
2011/8/12 Daniel,Wu <hadoop...@163.com <mailto:hadoop...@163.com>>
suppose the table is partitioned by period_key, and the csv
file also has a column named as period_key. The csv file contains
multiple days of data, how can we load it in the the table?
I think of an workaround by first load the data into a
non-partition table, and then insert the data from non-partition
table to the partition table.
hive> INSERT OVERWRITE TABLE sales SELECT * FROM sales_temp;
FAILED: Error in semantic analysis: need to specify partition
columns because the destination table is partitioned.
However it doesn't work also. please help.
--
With Regards
Vikas Srivastava
DWH & Analytics Team
Mob:+91 9560885900
One97 | Let's get talking !
