Optimizing SQL Self-Diagnosis

Performance issues may occur when you query data or run the INSERT, DELETE, UPDATE, or CREATE TABLE AS statement. In this case, you can query the warning column in the GS_WLM_SESSION_STATISTICS, GS_WLM_SESSION_HISTORY, and GS_WLM_SESSION_QUERY_INFO_ALL views to obtain reference for performance optimization.

Alarms that can trigger SQL self diagnosis depend on the settings of resource_track_level. If resource_track_level is set to query, alarms about the failures in collecting column statistics and pushing down SQL statements will trigger the diagnosis. If resource_track_level is set to operator, all alarms will trigger the diagnosis.

Whether a SQL plan will be diagnosed depends on the settings of resource_track_cost. A SQL plan will be diagnosed only if its execution cost is greater than resource_track_cost. You can use the EXPLAIN keyword to check the plan execution cost.

The SQL self-diagnosis function is affected by the enable_analyze_check parameter. Ensure that the function is enabled before using it.

If a large number of statements are executed, certain data may fail to be collected due to memory control. In this case, you can increase the value of instr_unique_sql_count.

Alarms

Currently, the following performance alarms will be reported:

Some column statistics are not collected.

An alarm will be reported if some column statistics are not collected. (The current feature is a lab feature. Contact Huawei technical support before using it.) For details about the optimization, see Updating Statistics and Optimizing Statistics.

An alarm will also be reported if column statistics are not collected for queries on OBS foreign tables. (The current feature is a lab feature. Contact Huawei technical support before using it.) The performance of the ANALYZE statement executed for OBS foreign tables is poor and seems inefficient for a simple query. Run this statement as needed.

Example alarms:

No statistics about a table are not collected.

Statistic Not Collect:
    schema_test.t1

The statistics about a single column are not collected.

Statistic Not Collect:
    schema_test.t2(c1,c2)

The statistics about multiple columns are not collected.

Statistic Not Collect:
    schema_test.t3((c1,c2))

The statistics about a single column and multiple columns are not collected.

Statistic Not Collect:
    schema_test.t4(c1,c2)    schema_test.t4((c1,c2))

SQL statements are not pushed down.
The cause details are displayed in the alarms. For details about the optimization, see Optimizing Statement Pushdown.
- If the pushdown failure is caused by functions, the function names are displayed in the alarm.
- If the pushdown failure is caused by syntax, the alarm indicates that the syntax does not support pushdown. For example, the syntax containing the With Recursive, Distinct On, or Row expression, and syntax whose return value is of record type do not support pushdown.

Example alarms:

SQL is not plan-shipping, reason : "With Recursive" can not be shipped"
SQL is not plan-shipping, reason : "Function now() can not be shipped"
SQL is not plan-shipping, reason : "Function string_agg() can not be shipped"

In a hash join, the larger table is used as the inner table.

An alarm will be reported if the number of rows in the inner table reaches or exceeds 10 times of that in the outer table, more than 100 thousand of inner-table rows are processed on each DN in average, and the join statement has spilled to disks. You can view the query_plan column in GS_WLM_SESSION_HISTORY to check whether the hash join is used. For details about the optimization, see Hint-based Tuning.

Example alarms:

PlanNode[7] Large Table is INNER in HashJoin "Vector Hash Aggregate"

nestloop is used in a large-table equivalent join.

An alarm will be reported if nestloop is used in an equivalent join where more than 100 thousand of the larger-table rows are processed on each DN in average. You can view the query_plan column of GS_WLM_SESSION_HISTORY to check whether nestloop is used. For details about the optimization, see GS_WLM_SESSION_HISTORY.

Example alarms:

PlanNode[5] Large Table with Equal-Condition use Nestloop"Nested Loop"

A large table is broadcasted.

An alarm will be reported if more than 100 thousand of rows are broadcast on each DN in average. For details about the optimization, see Hint-based Tuning.

Example alarms:

PlanNode[5] Large Table in Broadcast "Streaming(type: BROADCAST dop: 1/2)"

Data skew occurs.

An alarm will be reported if the number of rows processed on any DN exceeds 100 thousand, and the number of rows processed on a DN reaches or exceeds 10 times of that processed on another DN. For the optimization, see Case: Selecting an Appropriate Distribution Key and Optimizing Data Skew.

Example alarms:

PlanNode[6] DataSkew:"Seq Scan", min_dn_tuples:0, max_dn_tuples:524288

Estimation is inaccurate.

An alarm will be reported if the maximum number or the estimated maximum number of rows processed on a DN is over 10 thousand, and the larger number reaches or exceeds 10 times of the smaller one. For details about the optimization, see Hint-based Tuning.

Example alarms:

PlanNode[5] Inaccurate Estimation-Rows: "Hash Join" A-Rows:0, E-Rows:52488

Restrictions

An alarm contains a maximum of 2048 characters. If the length of an alarm exceeds this value (for example, a large number of long table names and column names are displayed in the alarm when their statistics are not collected), a warning instead of an alarm will be reported.
```
WARNING, "Planner issue report is truncated, the rest of planner issues will be skipped"
```
If a query statement contains the Limit operator, alarms of operators lower than Limit will not be reported.
For alarms about data skew and inaccurate estimation, only alarms on the lower-layer nodes in a plan tree will be reported. This is because the same alarms on the upper-level nodes may be triggered by problems on the lower-layer nodes. For example, if data skew occurs on the Scan node, data skew may also occur in operators (for example, Hashagg) at the upper layer.