Backward scan of a table wasn't implemented yet in MVCC so this achieves
that. I added simple test for mixed btree and mvcc backward scan but I
should add more intense testing for this.
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> Implements backward scanning and last() in MVCC lazy cursor and adds
directional rowid iteration in the MVCC store, with new tests for mixed
MVCC+B-Tree backward scans.
>
> - **MVCC Cursor (`core/mvcc/cursor.rs`)**:
> - Implement `prev()` and `last()` with mixed MVCC/B-Tree
coordination using `IterationDirection`.
> - Add `PrevState` and extend state machine to handle backward
iteration.
> - Update `get_new_position_from_mvcc_and_btree(...)` to choose
rowids based on direction.
> - Integrate B-Tree cursor calls (`last`, `prev`) and adjust
`rewind`/rowid selection; tweak next-rowid when at `End`.
> - **MVCC Store (`core/mvcc/database/mod.rs`)**:
> - Add `get_prev_row_id_for_table(...)` and generalized
`get_row_id_for_table_in_direction(...)` supporting forward/backward
scans.
> - Add tracing and minor refactors around next/prev rowid retrieval.
> - **Tests (`core/mvcc/database/tests.rs`)**:
> - Add test for backward scan combining B-Tree and MVCC and an
ignored test covering delete during backward scan.
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
430bd457e6. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
Closes#3980
SQLite doesn't rewrite INSERT lists or WHEN clause, it instead
lets the trigger go "stale" and will cause runtime errors. This
may not be great behavior, but it's compatible...
- Disallow DROP COLUMN on columns referenced in triggers
- Propagate RENAME COLUMN to trigger SQL definitions
DROP COLUMN is not allowed when the column is mentioned in a trigger
on the table the column is dropped from, eg:
```
turso> CREATE TABLE t(x,y);
turso> CREATE TRIGGER foo BEFORE INSERT ON t BEGIN INSERT INTO t VALUES (NEW.x); END;
turso> ALTER TABLE t DROP COLUMN x;
× Parse error: cannot drop column "x": it is referenced in trigger foo
```
However, it is allowed if the trigger is on another table:
```
turso> CREATE TABLE t(x,y);
turso> CREATE TABLE u(x,y);
turso> CREATE TRIGGER bar BEFORE INSERT ON t BEGIN INSERT INTO u(y) VALUES (NEW.x); END;
turso> ALTER TABLE u DROP COLUMN y;
turso> INSERT INTO t VALUES (1,1);
× Parse error: table u has no column named y
```
Nearly all of the code here is vibecoded. I first asked Cursor Composer to create
an initial implementation. Then, I asked it to try to discover edge cases using the
`turso` and `sqlite3` CLIs, and write tests+fixes for the edge cases found.
The code is a bit slop, but this isn't a particularly performance-critical use case
and it should solve most of the issues with RENAME and DROP COLUMN.
## Trigger Support
This PR adds support for triggers:
- `CREATE TRIGGER`
- `DROP TRIGGER`
Supported
- `BEFORE/AFTER INSERT`
- `BEFORE/AFTER DELETE`
- `BEFORE/AFTER UPDATE [OF <col1,col2,col3>]`
Not supported:
- `INSTEAD OF`
- `TEMPORARY`
### Implementation details
- Triggers are executed within a new `Insn::Program` instruction. The
spec of the insn differs a bit from SQlite: we store a `Statement`
inside that instruction that we can `reset()` for every invocation.
- Like Sqlite, trigger programs take `NEW` and `OLD` rows as program
parameters.
Whenever there are triggers that would fire as the result of a DML
statement:
- `DELETE` writes the rows being deleted into a `RowSet` first.
- `UPDATE` and `INSERT` write the rows being updated into an ephemeral
table first.
### Other shit
Also added `EXPLAIN` support - the bytecode plans for trigger
subprograms are appended after the main program.
### AI disclosure
Used Cursor quite a bit for generating boilerplate code for this - you
can blame all the bad code on the AI of course 🤡
### Follow-ups:
1. ALTER TABLE ops need to rewrite the sql in the CREATE TRIGGER
statement e.g. if a column is renamed. Columns cannot be dropped if
referenced in triggers.
2. Fix weird rowid -1 fallback:
https://github.com/tursodatabase/turso/pull/3979#issuecomment-3547999449Closes#3979
rowid should only try to use the current's position. So if we are not
pointing to a `Loaded` row, then it should return None
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> Change `rowid()` to return `None` unless cursor is on a `Loaded` row,
removing the implicit seek from `BeforeFirst`.
>
> - **Core MVCC Cursor (`core/mvcc/cursor.rs`)**:
> - Adjust `rowid()` behavior: remove implicit first-row seek when
`BeforeFirst`; return `None` unless position is `Loaded`.
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
8848775a71. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#3977
RE: #3970
That Column::new having 14 boolean arguments was not great.
Also this removes the unneeded `parent_cols: Vec<String>` from
`ResolvedFkRef`
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#3973
