[
https://issues.apache.org/jira/browse/DRILL-5830?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16187677#comment-16187677
]
ASF GitHub Bot commented on DRILL-5830:
---------------------------------------
Github user jinfengni commented on the issue:
https://github.com/apache/drill/pull/968
I'm not convinced that it's a good idea to back out the change to HBase
specific changes made in DRILL-5546.
You are right that project push-down in planner ideally should do its job
and push the list of columns. However, until planner could claim there is no
issues at all ( just like what happened prior to DRILL-5546), execution still
may face a columns list with "*". That's why in HBaseGroupScan we have
verifyColumnsAndConvertStar, just in case planner's project push-down did not
work in the way we want. If as you suggested, such conversion in
HBaseGroupScan is redundant, why would it cause regression (if planner's rule
works as expected)? or is it your intention to still keep "*" in
HBaseRecordReader? If that's your intention, I think we are going in the wrong
direction. HBase has a unified schema at table level.
I agree that the analysis of empty map {}, vs {a:varbinary}. It's
something we have to deal with. As a matter of fact, such scenarios does not
have to come from empty batch. It could happen with two regions with >0 rows.
For instance, regrion 1 has 10 rows, with cf1.c1 appears in only first 5
rows, while region 2 has 20 rows with cf1.c1 appears in every rows. For the
following query:
select CF1 FROM table where some_condition_on_row_key;
if "some_condition_on_row_key" is pushed to hbase and prunes the first 5
rows, region1 will return a batch with 5 rows, but with cf1 as an empty map,
while region2 will have map with cf1 as {c1:varbinary}.
In that sense, DRILL-5546 exposes such issues, and force us to have a
solution to handle empty map {} vs {a:varbinary}/
> Resolve regressions to MapR DB from DRILL-5546
> ----------------------------------------------
>
> Key: DRILL-5830
> URL: https://issues.apache.org/jira/browse/DRILL-5830
> Project: Apache Drill
> Issue Type: Bug
> Affects Versions: 1.12.0
> Reporter: Paul Rogers
> Assignee: Paul Rogers
> Fix For: 1.12.0
>
>
> DRILL-5546 added a number of fixes for empty batches. One part of the fix was
> for HBase. Key changes:
> * Add code to expand wildcards in the planner. (i.e. SELECT *)
> * Remove support for wildcards in the HBase record reader.
> As noted in DRILL-5775, this change had the effect of breaking support for
> MapR-DB binary (which is API compatible with HBase.) DRILL-5775 does this by
> expanding wildcards in the planner for MapR DB as was done for HBase in
> DRILL-5546.
> Unfortunately, this change introduced other regressions into the code as
> described by DRILL-5706.
> Investigation of those issues revealed that we should back out the original
> DRILL-5546 changes and go down a different route.
> As it turns out, HBase already had a project push-down rule that expanded
> wildcards. However, that rule didn't work correctly some of the time.
> DRILL-5546 fixed that bug, ensuring that wildcards are expanded (at least in
> the cases tested for this ticket.)
> The actual issue turned out to be a bug in the {{RecordBatchLoader}} class
> which did not consider map contents when detecting schema change. As a
> result, results like (row_key, cf\{}) were treated the same as (row_key,
> cf\{mycol}) and the actual data colums were discarded, but randomly depending
> on batch arrival order.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
