In the past, RepeatNode did not handle empty expressions.
It used DCHECK to check if the expression was non-empty.
In non-debug mode, this caused _child_block to remain unprocessed, resulting in a deadlock.
Now, if the expression is empty, the output block directly outputs _child_block
## Proposed changes
1. check data type whether can applied should not throw exception when real data type is subclass of signature data type
2. merge `SlotBinder` and `FunctionBinder` to `ExpressionAnalyzer` to skip rewrite the whole expression tree multiple times.
3. `ExpressionAnalyzer.buildCustomSlotBinderAnalyzer()` provide more refined code to bind slot by different parts and different priority
4. the origin slot binder has O(n^2) complexity, this pr use `Scope.nameToSlot` to support O(n) bind
5. modify some `Collection.stream()` to `ImmutableXxx.builder()` to remove some method call which are difficult to inline by jvm in the hot path, e.g. `Expression.<init>` and `AbstractTreeNode.<init>`
6. modify some `ImmutableXxx.copyOf(xxx)` to `Utils.fastToImmutableList(xxx)` to skip addition copy of the array
7. set init size to `Immutable.builder()` to skip some useless resize
8. lazy compute and cache some heavy operations, like `Scope.nameToSlot` and `CaseWhen.computeDataTypesForCoercion()`
(cherry picked from commit 83c2f5a95827136aac4f0a78c5e841e9a099858c)
The `PARTITION BY` syntax used by external catalogs has been added.
You can specify a column directly, or a partition function as a partition condition.
Like:
`PARTITION BY LIST(col1, col2, func(param), func(param1, param2), func(param1, param2, param3))`
NOTICE:
This PR change the grammar of `AUTO PARTITION`
From
```
AUTO PARTITION BY RANGE date_trunc(`TIME_STAMP`, 'month')
```
To
```
AUTO PARTITION BY RANGE (date_trunc(`TIME_STAMP`, 'month'))
```
* Problem:
Inconsistent behavior occurs when executing partial column update `UPDATE` statements and `INSERT` statements on merge-on-write tables with the Nereids optimizer enabled. The number of columns passed to BE differs; `UPDATE` operations incorrectly pass all columns, while `INSERT` operations correctly pass only the updated columns.
Reason:
The Nereids optimizer does not handle partial column update `UPDATE` statements properly. The processing logic for `UPDATE` statements rewrites them as equivalent `INSERT` statements, which are then processed according to the logic of `INSERT` statements. For example, assuming a MoW table structure with columns k1, k2, v1, v2, the correct rewrite should be:
* `UPDATE` table t1 set v1 = v1 + 1 where k1 = 1 and k2 = 2
* =>
* `INSERT` into table (v1) select v1 + 1 from table t1 where k1 = 1 and k2 = 2
However, the actual rewriting process does not consider the logic for partial column updates, leading to all columns being included in the `INSERT` statement, i.e., the result is:
* `INSERT` into table (k1, k2, v1, v2) select k1, k2, v1 + 1, v2 from table t1 where k1 = 1 and k2 = 2
This results in `UPDATE` operations incorrectly passing all columns to BE.
Solution:
Having analyzed the cause, the solution is straightforward: when rewriting partial column update `UPDATE` statements to `INSERT` statements, only retain the updated columns and all key columns (as partial column updates must include all key columns). Additionally, this PR includes error injection cases to verify the number of columns passed to BE is correct.
* 2
* 3
* 4
* 5
add tpc-h shape cases using leading hint
except:
single table without join q1 q6
not support feature include tables after subquery unnested q2 q16 q18 q20 q21 q22
