[
https://issues.apache.org/jira/browse/IMPALA-15387?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksandr Efimov updated IMPALA-15387:
--------------------------------------
Description:
Since IMPALA-3841, reading two nested levels can return a scalar field of the
next top-level row, silently. It starts at the first page boundary of the
second or a later row group. We hit it on production data.
{code:sql}
select count(*), count(o.o_orderkey), sum(o.o_orderkey)
from customer_nested_multiblock_multipage c
left join c.c_orders o
left join o.o_lineitems l;
-- expected (pyarrow): 11908, 11808, 34964524398
-- actual: 11908, 9964, 29547222652
{code}
One level ({{left join c.c_orders o}}) is correct.
Cause: ComplexColumnReader::Reset() does not clear next_levels_consumed_. The
first NextLevels() of c_orders in a new row group skips o_lineitems, treats
c_orders as empty and reads levels again, so o_orderkey reads its levels twice.
At the next page boundary it drops one value and is one row ahead from then on.
Fix: clear next_levels_consumed_ in Reset().
was:
Since IMPALA-3841, reading two nested levels can return a scalar field of
the next top-level row, silently. It starts at the first page boundary of
the second or a later row group. We hit it on production data.
{code:sql}
select count(*), count(o.o_orderkey), sum(o.o_orderkey)
from customer_nested_multiblock_multipage c
left join c.c_orders o
left join o.o_lineitems l;
-- expected (pyarrow): 11908, 11808, 34964524398
-- actual: 11908, 9964, 29547222652
{code}
One level ({{left join c.c_orders o}}) is correct.
Cause: ComplexColumnReader::Reset() does not clear next_levels_consumed_.
The first NextLevels() of c_orders in a new row group skips o_lineitems,
treats c_orders as empty and reads levels again, so o_orderkey reads its
levels twice. At the next page boundary it drops one value and is one row
ahead from then on.
Fix: clear next_levels_consumed_ in Reset().
> Parquet: two nested levels return the next row's values in later row groups
> ---------------------------------------------------------------------------
>
> Key: IMPALA-15387
> URL: https://issues.apache.org/jira/browse/IMPALA-15387
> Project: IMPALA
> Issue Type: Bug
> Components: Backend
> Affects Versions: Impala 5.0.0
> Reporter: Aleksandr Efimov
> Assignee: Aleksandr Efimov
> Priority: Critical
> Labels: correctness
>
> Since IMPALA-3841, reading two nested levels can return a scalar field of the
> next top-level row, silently. It starts at the first page boundary of the
> second or a later row group. We hit it on production data.
> {code:sql}
> select count(*), count(o.o_orderkey), sum(o.o_orderkey)
> from customer_nested_multiblock_multipage c
> left join c.c_orders o
> left join o.o_lineitems l;
> -- expected (pyarrow): 11908, 11808, 34964524398
> -- actual: 11908, 9964, 29547222652
> {code}
> One level ({{left join c.c_orders o}}) is correct.
> Cause: ComplexColumnReader::Reset() does not clear next_levels_consumed_. The
> first NextLevels() of c_orders in a new row group skips o_lineitems, treats
> c_orders as empty and reads levels again, so o_orderkey reads its levels
> twice. At the next page boundary it drops one value and is one row ahead from
> then on.
> Fix: clear next_levels_consumed_ in Reset().
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
