Sometimes there are cycles in extracted graph

[Matts966]

• We can prevent cycles by making dependency only between table references and create table statements if a table is not created by multiple statements and serial execution can create required tables.
• However, we are making dependency of INSERT on CREATE, UPDATE on INSERT and other DMLs on UPDATE for it is useful in preprocessing (before v0.5).
• If it is needed, we should implement an option to make dependency only between table references and create table statements simply.

[Matts966]

We can remove cycles by introducing some total order to the cycle like the dead lock prevention in DB.
We are now planning to implement an option to use file names order for the use and another option to remove all the edges in the cycle.

[Matts966]

Automatically removing cycles can ignore problems of SQL files.
It is simpler to make nodes depend only on create table statements, and if parallel execution is undesirable, write a script file to execute sequentially.

When the above simple rule implemented, we can ignore the possibility of cycles (in anyway, the type checking fails with cycle even now)

1. create table with unique name
2. only use existing resources

The cycle free property of this algorithm is proved by simple induction even for functions and variables.

When modifying existing scripts, we should be cautious about cycles, but dag command teaches us when it happens.

[Matts966]

Also, we can always remove cycles by refactoring SQL files to separate depended resources as a new file because all the resources can be generated by another SQL.

This can be viewed as an example of DIP (dependency inversion principle).

[Matts966]

From v0.6.0, we only resolve dependencies between table references and create table statements.

[Matts966]

Making dependency of INSERT on CREATE, UPDATE on INSERT, and other DMLs on UPDATE for the preprocessing context is a little bit complex, and the order of dependency is not always reasonable. Also, coding rules to prevent cycles will be more complex than the current resolution algorithm.

We can express arbitrary dependencies as written in #5 by using dummy tables.