IGNITE-19615: Index is not used while performing SELECT over an indexed column.

[lowka]

Removed int type / interval type checks from TypeCoercion::needToCast(Type, Type) - it behaves exactly the same way as TypeUtils::needToCast (which was removed in https://issues.apache.org/jira/browse/IGNITE-19128).

• Partially reverts changes, that were made in https://issues.apache.org/jira/browse/IGNITE-19128 and moved type cast checks to TypeUtils.
• Adds test cases with indexes to ImplicitCastsTest.
• Adds saturation casts for numeric types in index bounds to prevent runtime exceptions (see comments).

[korlov42]

After this patch an index is used indeed. But resulting plan is rather invalid. For a query like SELECT val FROM my WHERE id = 100500, where id is of type TINYINT, I've got following plan:

IgniteExchange<...>
  IgniteIndexScan(<...>, searchBounds=[[ExactBounds [bound=CAST(100500):TINYINT NOT NULL]]], filters=[=(CAST($t0):INTEGER NOT NULL, 100500)]<...>

this results in RuntimeException, since 100500 is out of range of TINYINT

[lowka]

After this patch an index is used indeed. But resulting plan is rather invalid. For a query like SELECT val FROM my WHERE id = 100500, where id is of type TINYINT, I've got following plan:

IgniteExchange<...>
  IgniteIndexScan(<...>, searchBounds=[[ExactBounds [bound=CAST(100500):TINYINT NOT NULL]]], filters=[=(CAST($t0):INTEGER NOT NULL, 100500)]<...>

this results in RuntimeException, since 100500 is out of range of TINYINT

@korlov42
I going to look into it. I looked at the result of IgniteSqlValidator and it add INTEGER casts to both id and literal (this is expected). But somehow the bounds are still use TINYINT.

[lowka]

Optimiser starts from:

LogicalProject(VAL=[$1]), id = 116
LogicalFilter(condition=[=(CAST($0):INTEGER NOT NULL, 100500)]), id = 114
IgniteLogicalTableScan(table=[[PUBLIC, MY]]), id = 113

And produces:

IgniteExchange(distribution=[single]), id = 144
IgniteIndexScan(table=[[PUBLIC, MY]], index=[MY_PK], type=[HASH], searchBounds=[[ExactBounds [bound=CAST(100500):TINYINT NOT NULL]]], filters=[=(CAST($t0):INTEGER NOT NULL, 100500)], projects=[[$t1]], requiredColumns=[{0, 1}], collation=[[0 CLU]]), id = 140

[lowka]

Yep. Types are correct.

SqlValidator BINARY_COMPARISON: CAST(MY.ID AS INTEGER) = 100500 types: INTEGER INTEGER

So the problem should be in SearchBound assembly code.

[korlov42]

First problem with this patch is that now IgniteTypeCoercion has two similar method with different behaviour:

boolean needToCast(RelDataType fromType, RelDataType toType)
boolean needToCast(SqlValidatorScope scope, SqlNode node, RelDataType toType)

The former returns false for (TINYINT, INTEGER), while the latter returns true in the very similar case (node::TINYINT, INTEGER).

The second problem is that it deals with literals only, while leaving dynamic params and columns references out of scope for some reason

[lowka]

The second problem is that it deals with literals only, while leaving dynamic params and columns references out of scope for some reason

What makes you think so? Search bounds can be built from arbitrary expressions, if it was not the case then the optimizer wouldn't be able to select an index for SELECT * FROM t WHERE c1=?
But nonetheless I have reverted the changes to the state as were copied from ignite2.

[lowka]

I have reverted the changes to the state as were copied from ignite2.

That would not work. Left all as it is - just renamed a method to needToCastInIndex.

[lowka]

• Partially reverted changes, that were made in https://issues.apache.org/jira/browse/IGNITE-19128 and moved type cast checks to TypeUtils.
• Added test cases with indexes to ImplicitCastsTest.

[korlov42]

Suggested change

	* Checks whether one type can be casted to another if one of type is a custom data type.
	* This method expects at least one of its arguments to be a custom data type.
	* Checks whether one type can be casted to another if one of type is a custom data type.
	*
	* <p>This method expects at least one of its arguments to be a custom data type.

[korlov42]

why do we need both this assertion and one on line 584? Let's leave only one on line 584 with improved message

[korlov42]

this predicate doesn't seem right... Method customDataTypeNeedCast actually expects that at least one of the argument is instance of IgniteCustomType. If there is an invariant that there should not be any ANY type but IgniteCustomType, it's better to check it explicitly

[lowka]

@korlov42 yeap. I removed the first assertion since the second one is the only one that is needed.

[korlov42]

why do even need call in needCastInSearchBounds? This is the very same rule as in TypeCoercion

[lowka]

Fixed.