[HIVE-20304] When hive.optimize.skewjoin and hive.auto.convert.join are both set to true, and the execution engine is mr, same stage may launch twice due to the wrong generated plan

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 1.2.1, 2.3.3
• Fix Version/s: 4.0.0-alpha-2
• Component/s: CLI
• Assignee: Hui Huang
• Reporter: Hui Huang
• Created: 03/Aug/18 09:21
• Updated: 16/Nov/22 13:50
• Resolved: 16/Nov/18 15:22

Description

`When hive.optimize.skewjoin and hive.auto.convert.join are both set to true, and the execution engine is set to mr, same stage of a query may launch twice due to the wrong generated plan. If hive.exec.parallel is also true, the same stage will launch at the same time and the job will failed due to the first completed stage clear the map.xml/reduce.xml file stored in the hdfs.

use following sql to reproduce the issue:

CREATE TABLE `tbl1`(
  `fence` string);

CREATE TABLE `tbl2`(
  `order_id` string,
  `phone` string,
  `search_id` string
)
PARTITIONED BY (
  `dt` string);


CREATE TABLE `tbl3`(
  `order_id` string,
  `platform` string)
PARTITIONED BY (
  `dt` string);


CREATE TABLE `tbl4`(
  `groupname` string,
  `phone` string)
PARTITIONED BY (
  `dt` string);


CREATE TABLE `tbl5`(
  `search_id` string,
  `fence` string)
PARTITIONED BY (
  `dt` string);

SET hive.exec.parallel = TRUE;

SET hive.auto.convert.join = TRUE;

SET hive.optimize.skewjoin = TRUE;


SELECT dt,
               platform,
               groupname,
               count(1) as cnt
        FROM
        (SELECT dt,
                platform,
                groupname
         FROM
         (SELECT fence
          FROM tbl1)ta
           JOIN
           (SELECT a0.dt,
                   a1.platform,
                   a2.groupname,
                   a3.fence
            FROM
            (SELECT dt,
                    order_id,
                    phone,
                    search_id
             FROM tbl2
             WHERE dt =20180703 )a0
              JOIN
              (SELECT order_id,
                      platform,
                      dt
               FROM tbl3
               WHERE dt =20180703 )a1 ON a0.order_id = a1.order_id
              INNER JOIN
              (SELECT groupname,
                      phone,
                      dt
               FROM tbl4
               WHERE dt =20180703 )a2 ON a0.phone = a2.phone
              LEFT JOIN
              (SELECT search_id,
                      fence,
                      dt
               FROM tbl5
               WHERE dt =20180703)a3 ON a0.search_id = a3.search_id)t0 ON ta.fence = t0.fence)t11
        GROUP BY dt,
                 platform,
                 groupname;

DROP TABLE tbl1;
DROP TABLE tbl2;
DROP TABLE tbl3;
DROP TABLE tbl4;
DROP TABLE tbl5;

We will get some error message like this:

Examining task ID: task_1531284442065_3637_m_000000 (and more) from job job_1531284442065_3637

Task with the most failures(4):

---

Task ID:
task_1531284442065_3637_m_000000

URL:
http://0.0.0.0:8088/taskdetails.jsp?jobid=job_1531284442065_3637&tipid=task_1531284442065_3637_m_000000

---

Diagnostic Messages for this Task:
File does not exist: hdfs://test/tmp/hive-hadoop/hadoop/fe5efa94-abb1-420f-b6ba-ec782e7b79ad/hive_2018-08-03_17-00-17_707_592882314975289971-5/-mr-10045/757eb1f7-7a37-4a7e-abc0-4a3b8b06510c/reduce.xml
java.io.FileNotFoundException: File does not exist: hdfs://test/tmp/hive-hadoop/hadoop/fe5efa94-abb1-420f-b6ba-ec782e7b79ad/hive_2018-08-03_17-00-17_707_592882314975289971-5/-mr-10045/757eb1f7-7a37-4a7e-abc0-4a3b8b06510c/reduce.xml

Looking into the plan by executing explain, I found that the Stage-4 and Stage-5 can reached from multi root tasks.

Explain
STAGE DEPENDENCIES:
  Stage-21 is a root stage , consists of Stage-34, Stage-5
  Stage-34 has a backup stage: Stage-5
  Stage-20 depends on stages: Stage-34
  Stage-17 depends on stages: Stage-5, Stage-18, Stage-20 , consists of Stage-32, Stage-33, Stage-1
  Stage-32 has a backup stage: Stage-1
  Stage-15 depends on stages: Stage-32
  Stage-10 depends on stages: Stage-1, Stage-15, Stage-16 , consists of Stage-31, Stage-2
  Stage-31
  Stage-9 depends on stages: Stage-31
  Stage-2 depends on stages: Stage-9
  Stage-33 has a backup stage: Stage-1
  Stage-16 depends on stages: Stage-33
  Stage-1
  Stage-5
  Stage-27 is a root stage , consists of Stage-37, Stage-38, Stage-4
  Stage-37 has a backup stage: Stage-4
  Stage-25 depends on stages: Stage-37
  Stage-12 depends on stages: Stage-4, Stage-22, Stage-23, Stage-25, Stage-26 , consists of Stage-36, Stage-5
  Stage-36
  Stage-11 depends on stages: Stage-36
  Stage-19 depends on stages: Stage-11 , consists of Stage-35, Stage-5
  Stage-35 has a backup stage: Stage-5
  Stage-18 depends on stages: Stage-35
  Stage-38 has a backup stage: Stage-4
  Stage-26 depends on stages: Stage-38
  Stage-4
  Stage-30 is a root stage , consists of Stage-42, Stage-43, Stage-3
  Stage-42 has a backup stage: Stage-3
  Stage-28 depends on stages: Stage-42
  Stage-14 depends on stages: Stage-3, Stage-28, Stage-29 , consists of Stage-41, Stage-4
  Stage-41
  Stage-13 depends on stages: Stage-41
  Stage-24 depends on stages: Stage-13 , consists of Stage-39, Stage-40, Stage-4
  Stage-39 has a backup stage: Stage-4
  Stage-22 depends on stages: Stage-39
  Stage-40 has a backup stage: Stage-4
  Stage-23 depends on stages: Stage-40
  Stage-43 has a backup stage: Stage-3
  Stage-29 depends on stages: Stage-43
  Stage-3
  Stage-0 depends on stages: Stage-2

After skewjoin optimization, the processed node is added into the listTasks of ConditionalTask and the parentTask of the processed node is removed and during the commonJoin optimization of listTasks of ConditionalTask, the new generated condTask will be added into root task list due to parentTask is null.

workaround: do not set hive.optimize.skewjoin and hive.auto.convert.join to true at the same time.

Issue Links

• duplicates: HIVE-14557 Nullpointer When both SkewJoin and Mapjoin Enabled