[
https://issues.apache.org/jira/browse/HIVE-30246?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Denys Kuzmenko updated HIVE-30246:
----------------------------------
Description:
*Problem*
With {{hive.merge.tezfiles=true}}, a DELETE on an Iceberg table gets a merge
stage (HIVE-28267) for the small files it wrote. The merge task fails in two
cases:
# V3 merge-on-read: the DELETE writes deletion vectors in Puffin files, and
{{IcebergMergeRecordReader}} reads only Avro, ORC and Parquet.
# Copy-on-write, any format version: the DELETE writes data files, but the
merge stage's file sink gets the DELETE operation without the copy-on-write
flag ({{GenMapRedUtils}}, the {{isCustomDelete}}
branch), so it reads the data rows with the delete schema.
V2 merge-on-read DELETE merges fine. UPDATE and MERGE are not affected: they
never get a merge stage.
V3 merge-on-read:
{noformat}
TezRuntimeException: Vertex failed, vertexName=File Merge, ...
Caused by: java.lang.UnsupportedOperationException: Cannot read PUFFIN file:
file:/.../warehouse/v3_source/data/00000-0-...
at
org.apache.iceberg.mr.mapreduce.IcebergMergeRecordReader.openGeneric(IcebergMergeRecordReader.java:108)
{noformat}
Copy-on-write:
{noformat}
TezRuntimeException: Vertex failed, vertexName=File Merge, ...
Caused by: java.lang.ArrayIndexOutOfBoundsException: Index 1 out of bounds
for length 1
at org.apache.iceberg.data.GenericRecord.get(GenericRecord.java:129)
at
org.apache.iceberg.mr.hive.serde.objectinspector.IcebergRecordObjectInspector.lambda$getStructFieldsDataAsList$0(IcebergRecordObjectInspector.java:89)
{noformat}
*Reproducer* (TestIcebergLlapLocalCliDriver)
{code:sql}
--! qt:dataset:src
set hive.merge.mapredfiles=true;
set hive.merge.mapfiles=true;
set hive.merge.tezfiles=true;
set hive.optimize.sort.dynamic.partition.threshold=-1;
set mapred.reduce.tasks=5;
set hive.blobstore.supported.schemes=hdfs,file;
-- V3 merge-on-read: tblproperties('format-version'='3')
-- copy-on-write: tblproperties('format-version'='2',
'write.delete.mode'='copy-on-write')
create table t(key string) stored by iceberg stored as parquet
tblproperties('format-version'='3');
insert overwrite table t select key from src;
insert into table t select key from src;
insert into table t select key from src;
insert into table t select key from src;
delete from t where key in (select key from src where key < '200');
{code}
*Expected*
The DELETE succeeds.
*Proposed fix*
- V3 merge-on-read: skip DVs ({{ContentFileUtil.isDV}}) when
{{HiveIcebergOutputCommitter.getOutputFiles}} lists the merge input. A DV
already holds every deleted position of its data file, so it has nothing
to gain from the merge. For a V3 DELETE the list is then empty, and
{{ConditionalResolverMergeFiles}} runs only the move task.
- Copy-on-write: give a copy-on-write DELETE no merge stage, as UPDATE and
MERGE already have none ({{SemanticAnalyzer}} sets {{canBeMerged}} to false for
them).
was:
*Problem*
With {{hive.merge.tezfiles=true}}, a DELETE on a format-version 3 Iceberg
table fails in the File Merge vertex. The merge stage for delete files
(HIVE-28267) takes every delete file the DELETE wrote as merge
input. On V3 these are deletion vectors in Puffin files, and
{{IcebergMergeRecordReader}} reads only Avro, ORC and Parquet.
{noformat}
TezRuntimeException: Vertex failed, vertexName=File Merge, ...
Caused by: java.lang.UnsupportedOperationException: Cannot read PUFFIN file:
file:/.../warehouse/v3_source/data/00000-0-...
at
org.apache.iceberg.mr.mapreduce.IcebergMergeRecordReader.openGeneric(IcebergMergeRecordReader.java:108)
at
org.apache.iceberg.mr.mapreduce.IcebergMergeRecordReader.nextTask(IcebergMergeRecordReader.java:59)
at
org.apache.iceberg.mr.mapreduce.IcebergMergeRecordReader.initialize(IcebergMergeRecordReader.java:55)
...
at
org.apache.iceberg.mr.hive.HiveIcebergInputFormat.getRecordReader(HiveIcebergInputFormat.java:232)
at
org.apache.hadoop.hive.ql.io.CombineHiveRecordReader.<init>(CombineHiveRecordReader.java:104)
{noformat}
Only DELETE is affected: UPDATE and MERGE never get a merge stage.
*Reproducer* (TestIcebergLlapLocalCliDriver; iceberg_merge_delete_files.q
with format-version 3)
{code:sql}
--! qt:dataset:src
set hive.merge.mapredfiles=true;
set hive.merge.mapfiles=true;
set hive.merge.tezfiles=true;
set hive.optimize.sort.dynamic.partition.threshold=-1;
set mapred.reduce.tasks=5;
set hive.blobstore.supported.schemes=hdfs,file;
create table v3_source(key string) stored by iceberg stored as parquet
tblproperties('format-version'='3');
insert overwrite table v3_source select key from src;
insert into table v3_source select key from src;
insert into table v3_source select key from src;
insert into table v3_source select key from src;
delete from v3_source where key in (select key from src where key < '200');
{code}
*Expected*
The DELETE succeeds. A DV already holds every deleted position of its data
file, so DVs have nothing to gain from the row-based merge.
*Proposed fix*
Skip DVs ({{ContentFileUtil.isDV}}) when
{{HiveIcebergOutputCommitter.getOutputFiles}} lists the merge input. For a V3
DELETE the list is then empty, and {{ConditionalResolverMergeFiles}} runs only
the move
task. V2 position-delete files and data files are merged as before.
> Iceberg: DELETE with hive.merge.tezfiles=true fails in the File Merge vertex
> on V3 and copy-on-write tables
> ------------------------------------------------------------------------------------------------------------
>
> Key: HIVE-30246
> URL: https://issues.apache.org/jira/browse/HIVE-30246
> Project: Hive
> Issue Type: Bug
> Components: Iceberg integration
> Reporter: Denys Kuzmenko
> Priority: Major
>
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
