[
https://issues.apache.org/jira/browse/DRILL-5480?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16016659#comment-16016659
]
Jinfeng Ni edited comment on DRILL-5480 at 5/18/17 11:37 PM:
-------------------------------------------------------------
We may argue that in the above reproduce sample data, hbase table has a schema
change, since different regions have different columns, hence schema change
exception seems to be reasonable, but the incorrect query result is still a bug.
Here is the case where there is no schema change across different hbase
regions; all 3 regions have 'row_key', and 'orders.id'. However, query failed
with schema change conflicts intermittently.
{code}
create 'customer', 'orders', {SPLITS => ['3', '6']}
put 'customer', '1', 'orders:id', '100'
put 'customer', '5', 'orders:id', '200'
put 'customer', '7', 'orders:id', '300'
alter session set `planner.enable_broadcast_join`=false;
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
{code}
In the following, the first run hit schema conflicts in UnorderReceiver, and
second run was successful.
{code}
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
Error: SYSTEM ERROR: IllegalStateException: Failure while reading vector.
Expected vector class of org.apache.drill.exec.vector.NullableIntVector but was
holding vector class org.apache.drill.exec.vector.complex.MapVector, field=
orders(MAP:REQUIRED) [id(VARBINARY:OPTIONAL) [$bits$(UINT1:REQUIRED),
id(VARBINARY:OPTIONAL) [$offsets$(UINT4:REQUIRED)]]]
Fragment 2:0
[Error Id: 4a249fc4-0323-4c25-82bf-2313e05a3d82 on drlvm-245.qa.lab:31010]
(state=,code=0)
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
| row_key | orders | o_orderkey | o_custkey | o_orderstatus |
o_totalprice | o_orderdate | o_orderpriority | o_clerk |
o_shippriority | o_comment |
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
| [B@12c7d5cf | {"id":"MTAw"} | 100 | 1471 | O |
198978.27 | 1998-02-28 | 4-NOT SPECIFIED | Clerk#000000577 | 0
| heodolites detect slyly alongside of the ent |
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
{code}
The cause of the schema change: with filter pushdown, the third region from
HBase API returns nothing. This forces Drill to inject nullable-int column,
which causes the schema conflict in downstream operator (UnorderedReceiver).
{code}
04-03 Scan(groupscan=[HBaseGroupScan
[HBaseScanSpec=HBaseScanSpec [tableName=customer, startRow=, stopRow=,
filter=SingleColumnValueFilter (orders, id, LESS_OR_EQUAL, 200)],
columns=[`*`]]]) : rowType = RecordType(ANY row_key, (VARCHAR(65535), ANY) MAP
orders): rowcount = 1572864.0, cumulative cost = {1572864.0 rows, 3145728.0
cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 33866
{code}
was (Author: jni):
We may argue that in the above reproduce sample data, hbase table has a schema
change, since different regions have different columns, hence schema change
exception seems to be reasonable, but the incorrect query result is still a bug.
Here is the case where there is no schema change across different hbase
regions; all 3 regions have 'row_key', and 'orders.id'. However, query failed
with schema change conflicts intermittently.
{code}
create 'customer', 'orders', {SPLITS => ['3', '6']}
put 'customer', '1', 'orders:id', '100'
put 'customer', '5', 'orders:id', '200'
put 'customer', '7', 'orders:id', '300'
alter session set `planner.enable_broadcast_join`=false;
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
{code}
In the following, the first run hit schema conflicts in UnorderReceiver, and
second run was successful.
{code}
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
Error: SYSTEM ERROR: IllegalStateException: Failure while reading vector.
Expected vector class of org.apache.drill.exec.vector.NullableIntVector but was
holding vector class org.apache.drill.exec.vector.complex.MapVector, field=
orders(MAP:REQUIRED) [id(VARBINARY:OPTIONAL) [$bits$(UINT1:REQUIRED),
id(VARBINARY:OPTIONAL) [$offsets$(UINT4:REQUIRED)]]]
Fragment 2:0
[Error Id: 4a249fc4-0323-4c25-82bf-2313e05a3d82 on drlvm-245.qa.lab:31010]
(state=,code=0)
select * from hbase.customer c, cp.`tpch/orders.parquet` o where
cast(c.orders.id as bigint) = o.o_orderkey and c.orders.id <= '200';
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
| row_key | orders | o_orderkey | o_custkey | o_orderstatus |
o_totalprice | o_orderdate | o_orderpriority | o_clerk |
o_shippriority | o_comment |
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
| [B@12c7d5cf | {"id":"MTAw"} | 100 | 1471 | O |
198978.27 | 1998-02-28 | 4-NOT SPECIFIED | Clerk#000000577 | 0
| heodolites detect slyly alongside of the ent |
+--------------+----------------+-------------+------------+----------------+---------------+--------------+------------------+------------------+-----------------+-----------------------------------------------+
{code}
The cause of the schema change: with filter pushdown, the second region from
HBase API returns nothing. This forces Drill to inject nullable-int column,
which causes the schema conflict in downstream operator (UnorderedReceiver).
{code}
04-03 Scan(groupscan=[HBaseGroupScan
[HBaseScanSpec=HBaseScanSpec [tableName=customer, startRow=, stopRow=,
filter=SingleColumnValueFilter (orders, id, LESS_OR_EQUAL, 200)],
columns=[`*`]]]) : rowType = RecordType(ANY row_key, (VARCHAR(65535), ANY) MAP
orders): rowcount = 1572864.0, cumulative cost = {1572864.0 rows, 3145728.0
cpu, 0.0 io, 0.0 network, 0.0 memory}, id = 33866
{code}
> Empty batch returning from HBase may cause SchemChangeException or incorrect
> query result
> -----------------------------------------------------------------------------------------
>
> Key: DRILL-5480
> URL: https://issues.apache.org/jira/browse/DRILL-5480
> Project: Apache Drill
> Issue Type: Bug
> Reporter: Jinfeng Ni
> Assignee: Jinfeng Ni
>
> The following repo was provided by [~haozhu].
> 1. Create a Hbase table with 4 regions
> {code}
> create 'myhbase', 'cf1','cf2', {SPLITS => ['a', 'b', 'c']}
> put 'myhbase','a','cf1:col1','somedata'
> put 'myhbase','b','cf1:col2','somedata'
> put 'myhbase','c','cf2:col1','somedata'
> {code}
> One region has cf1.col1. One region has column family 'cf1', but does not
> have 'col1' under 'cf1'. One region has only column family 'cf2'. And last
> region is complete empty.
> 2. Prepare a csv file.
> {code}
> select * from dfs.tmp.`joinhbase.csv`;
> +-------------------+
> | columns |
> +-------------------+
> | ["1","somedata"] |
> | ["2","somedata"] |
> | ["3","somedata"] |
> {code}
> Now run the following query on drill 1.11.0-SNAPSHOT:
> {code}
> select cast(H.row_key as varchar(10)) as keyCol, CONVERT_FROM(H.cf1.col1,
> 'UTF8') as col1
> from
> hbase.myhbase H JOIN dfs.tmp.`joinhbase.csv` C
> ON CONVERT_FROM(H.cf1.col1, 'UTF8')= C.columns[1]
> ;
> {code}
> The correct query result show be:
> {code}
> +---------+-----------+
> | keyCol | col1 |
> +---------+-----------+
> | a | somedata |
> | a | somedata |
> | a | somedata |
> +---------+-----------+
> {code}
> Turn off broadcast join, then we will see SchemaChangeException, or incorrect
> result randomly. By 'randomly', it means in the same session, the same query
> would hit SchemaChangeException in one run, while gets incorrect result in a
> second run.
> {code}
> alter session set `planner.enable_broadcast_join`=false;
> {code}
> {code}
> select cast(H.row_key as varchar(10)) as keyCol, CONVERT_FROM(H.cf1.col1,
> 'UTF8') as col1
> . . . . . . . . . . . . . . . . . .> from
> . . . . . . . . . . . . . . . . . .> hbase.myhbase H JOIN
> dfs.tmp.`joinhbase.csv` C
> . . . . . . . . . . . . . . . . . .> ON CONVERT_FROM(H.cf1.col1, 'UTF8')=
> C.columns[1]
> . . . . . . . . . . . . . . . . . .> ;
> Error: SYSTEM ERROR: SchemaChangeException: Hash join does not support schema
> changes
> {code}
> {code}
> +---------+-------+
> | keyCol | col1 |
> +---------+-------+
> +---------+-------+
> No rows selected (0.302 seconds)
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
