[
https://issues.apache.org/jira/browse/HIVE-6010?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13846029#comment-13846029
]
Sergey Shelukhin commented on HIVE-6010:
----------------------------------------
I looked at CliDriver generation/code/flow... the plan is as such (this can
also be used for other stuff later if needed).
There will be new CliDriver template, called TestCompareCliDriver, with
separate set of .q files.
Unlike normal CliDriver, it will not use .out files; instead, there will be
multiple .qv (query version) initialization files; I haven't decided yet
whether these should be a set per query (q file), or a set applied to all
queries. The latter is simpler and solves the problem for vectorization, but
the former may make sense for other things, esp. if we need to compare more
things, Nqv x Nq combinations to run will quickly become ugly. Perhaps
per-query qv files can be added when needed.
The test, for each of its q files, will concatenate all the requisite qv files
in turn with the q file, run each of the resulting queries w/different output
files, and diff the outputs with each other. It will fail if they don't match.
So, for vectorization we can have some simple queries (arithmetics, functions,
etc.), with qv files being one-liners to enable and disable vectorization.
[~ehans] [~jnp] opinions?
> create a test that would ensure vectorization produces same results as
> non-vectorized execution
> -----------------------------------------------------------------------------------------------
>
> Key: HIVE-6010
> URL: https://issues.apache.org/jira/browse/HIVE-6010
> Project: Hive
> Issue Type: Test
> Components: Tests, Vectorization
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
>
> So as to ensure that vectorization is not forgotten when changes are made to
> things. Obviously it would not be viable to have a bulletproof test, but at
> least a subset of operations can be verified.
--
This message was sent by Atlassian JIRA
(v6.1.4#6159)
