[
https://issues.apache.org/jira/browse/HIVE-6131?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13963641#comment-13963641
]
Pala M Muthaia commented on HIVE-6131:
--------------------------------------
I looked into the failures above and revisited HIVE-3833 with more context now:
1. LazyBinaryColumnarSerde requires partition level metadata to read existing
data, it needs exact metadata used when serializing the data. So cannot use
table level metadata which could have changed.
2. Other serdes/format, which support schema change, needs updated schema to
support newly appended data with new columns.
So, it seems we should pass the table metadata or partition metadata
selectively, depending on what the storage/serde supports. Is there a way to
identify the serdes/format that do not support newer schema, programmatically?
I don't see anything obvious. Alternative is to
a. Add such metadata to serde info and populate that for all serdes. This may
have been discussed briefly in HIVE-3833, and looks like this will be a large
change because it essentially modifies interface for a plugin.
b. Hardcode a white or blacklist of serdes and pass table/partition level
metadata accordingly.
[~ashutoshc], [~szehon], any thoughts on the above, particularly are there
other alternatives?
> New columns after table alter result in null values despite data
> ----------------------------------------------------------------
>
> Key: HIVE-6131
> URL: https://issues.apache.org/jira/browse/HIVE-6131
> Project: Hive
> Issue Type: Bug
> Affects Versions: 0.11.0, 0.12.0, 0.13.0
> Reporter: James Vaughan
> Priority: Minor
> Attachments: HIVE-6131.1.patch
>
>
> Hi folks,
> I found and verified a bug on our CDH 4.0.3 install of Hive when adding
> columns to tables with Partitions using 'REPLACE COLUMNS'. I dug through the
> Jira a little bit and didn't see anything for it so hopefully this isn't just
> noise on the radar.
> Basically, when you alter a table with partitions and then reupload data to
> that partition, it doesn't seem to recognize the extra data that actually
> exists in HDFS- as in, returns NULL values on the new column despite having
> the data and recognizing the new column in the metadata.
> Here's some steps to reproduce using a basic table:
> 1. Run this hive command: CREATE TABLE jvaughan_test (col1 string)
> partitioned by (day string);
> 2. Create a simple file on the system with a couple of entries, something
> like "hi" and "hi2" separated by newlines.
> 3. Run this hive command, pointing it at the file: LOAD DATA LOCAL INPATH
> '<FILEDIR>' OVERWRITE INTO TABLE jvaughan_test PARTITION (day = '2014-01-02');
> 4. Confirm the data with: SELECT * FROM jvaughan_test WHERE day =
> '2014-01-02';
> 5. Alter the column definitions: ALTER TABLE jvaughan_test REPLACE COLUMNS
> (col1 string, col2 string);
> 6. Edit your file and add a second column using the default separator
> (ctrl+v, then ctrl+a in Vim) and add two more entries, such as "hi3" on the
> first row and "hi4" on the second
> 7. Run step 3 again
> 8. Check the data again like in step 4
> For me, this is the results that get returned:
> hive> select * from jvaughan_test where day = '2014-01-01';
> OK
> hi NULL 2014-01-02
> hi2 NULL 2014-01-02
> This is despite the fact that there is data in the file stored by the
> partition in HDFS.
> Let me know if you need any other information. The only workaround for me
> currently is to drop partitions for any I'm replacing data in and THEN
> reupload the new data file.
> Thanks,
> -James
--
This message was sent by Atlassian JIRA
(v6.2#6252)
