Tuple pointers in the generated row batches may not be initialized
if a tuple has byte size 0. There are some codes which compare these
uninitialized pointers against nullptr so having them uninitialized
may return wrong (and non-deterministic) results, e.g.:
impala::TupleIsNullPredicate::GetBooleanVal()
The following query produces non-deterministic results currently:
SELECT count(v.x) FROM functional.alltypestiny t3 LEFT OUTER JOIN (
SELECT true AS x FROM functional.alltypestiny t1 LEFT OUTER JOIN
functional.alltypestiny t2 ON (true)) v ON (v.x = t3.bool_col)
WHERE t3.bool_col = true;
The alltypestiny table has 8 records, 4 records of them has the true
boolean value for bool_col. Therefore, the above query is a fancy
way of saying "8 * 8 * 4", i.e. it should return 256.
The solution is that scanners initialize tuple pointers to a non-null
value when there are no materialized slots. This non-null value is
provided by the new static member Tuple::POISON.
I extended QueryTest/scanners.test with the above query. This test
executes the query against all table formats.
This change has the biggest performance impact on count(*) queries on
large kudu tables. For my quick benchmark I copied tpch_kudu.lineitem
and doubled its data. The resulting table has 12,002,430 rows.
Without this patch 'select count(*) from biglineitem' runs for ~0.12s.
With the patch applied, the overhead is around a dozens of ms. I measured
the query on my desktop PC using a relase build of Impala. On debug builds,
the execution time of the patched version is around 160% of the original
version.
Without this patch:
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
| Operator | #Hosts | Avg Time | Max Time | #Rows | Est. #Rows | Peak Mem | Est. Peak Mem | Detail |
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
| 03:AGGREGATE | 1 | 127.50us | 127.50us | 1 | 1 | 28.00 KB | 10.00 MB | FINALIZE |
| 02:EXCHANGE | 1 | 22.32ms | 22.32ms | 3 | 1 | 0 B | 0 B | UNPARTITIONED |
| 01:AGGREGATE | 3 | 1.78ms | 1.89ms | 3 | 1 | 16.00 KB | 10.00 MB | |
| 00:SCAN KUDU | 3 | 8.00ms | 8.28ms | 12.00M | -1 | 512.00 KB | 0 B | default.biglineitem |
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
With this patch:
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
| Operator | #Hosts | Avg Time | Max Time | #Rows | Est. #Rows | Peak Mem | Est. Peak Mem | Detail |
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
| 03:AGGREGATE | 1 | 129.01us | 129.01us | 1 | 1 | 28.00 KB | 10.00 MB | FINALIZE |
| 02:EXCHANGE | 1 | 33.00ms | 33.00ms | 3 | 1 | 0 B | 0 B | UNPARTITIONED |
| 01:AGGREGATE | 3 | 1.99ms | 2.13ms | 3 | 1 | 16.00 KB | 10.00 MB | |
| 00:SCAN KUDU | 3 | 13.13ms | 13.97ms | 12.00M | -1 | 512.00 KB | 0 B | default.biglineitem |
+--------------+--------+----------+----------+--------+------------+-----------+---------------+---------------------+
Change-Id: I298122aaaa7e62eb5971508e0698e189519755de
Reviewed-on: http://gerrit.cloudera.org:8080/9239
Reviewed-by: Tim Armstrong <tarmstrong@cloudera.com>
Tested-by: Impala Public Jenkins
0eaab69fff