emliunix / dbslow Goto Github PK

Just write one in half an hour and get a first glance over the whole system.

There're some low hanging fruits right now:

• Merge Sort Join
• Aggregate/Window

Merge Sort Join

It shall be represented as a simple join in plan tree. Depending on the condition, it can be executed with nested loop or merge sort (only when there're only ANDed sequence of column equalities).

So the required stages are:

• Merge
• NestedLoop
• Sorted

Aggregate / Window

Currently the type checking/casting relies on op context (repository of function/op information). But it's not a simple function call.

So one solution might be to pre-split the expression before type checking. With the agg/win functions in separate registry.

In simple expression, it's natural to turn to null for insensible operations, eg. (1 + null => null).

By design, null is a special value included in all types. And in primitive operations, usually the result will be null if any operand is null except IS NULL.

However, there're more special handling required to make SQL useful.

In aggregation, we don't want to NULLify the whole state on first met of a null value. So maybe, for sum/cumsum, we can treat nulls as 0s.

Still not sure if any other special cases in aggregation functions.

Also, in sorting, should null be min or max.