This commit adds suport for DROP INDEX.
Bytecode produced by this commit differs from SQLITE's bytecode, main
reason we don't do autovacuum or repacking of pages like SQLITE does.
Closes#1280Closes#1444
The following code reproduces the leak (memory usage increases over
time):
```rust
#[tokio::main]
async fn main() {
let db = Builder::new_local(":memory:").build().await.unwrap();
let conn = db.connect().unwrap();
conn.execute("SELECT load_extension('./target/debug/liblimbo_series');", ())
.await
.unwrap();
loop {
conn.execute("SELECT * FROM generate_series(1,10,2);", ())
.await
.unwrap();
}
}
```
After switching to the system allocator, the leak becomes detectable
with Valgrind:
```
32,000 bytes in 1,000 blocks are definitely lost in loss record 24 of 24
at 0x538580F: malloc (vg_replace_malloc.c:446)
by 0x62E15FA: alloc::alloc::alloc (alloc.rs:99)
by 0x62E172C: alloc::alloc::Global::alloc_impl (alloc.rs:192)
by 0x62E1530: allocate (alloc.rs:254)
by 0x62E1530: alloc::alloc::exchange_malloc (alloc.rs:349)
by 0x62E0271: new<limbo_series::GenerateSeriesCursor> (boxed.rs:257)
by 0x62E0271: open_GenerateSeriesVTab (lib.rs:19)
by 0x425D8FA: limbo_core::VirtualTable::open (lib.rs:732)
by 0x4285DDA: limbo_core::vdbe::execute::op_vopen (execute.rs:890)
by 0x42351E8: limbo_core::vdbe::Program::step (mod.rs:396)
by 0x425C638: limbo_core::Statement::step (lib.rs:610)
by 0x40DB238: limbo::Statement::execute::{{closure}} (lib.rs:181)
by 0x40D9EAF: limbo::Connection::execute::{{closure}} (lib.rs:109)
by 0x40D54A1: example::main::{{closure}} (example.rs:26)
```
Interestingly, when using mimalloc, neither Valgrind nor mimalloc’s
internal statistics report the leak.
Reviewed-by: Preston Thorpe (@PThorpe92)
Closes#1447
The following code reproduces the leak, with memory usage increasing
over time:
```
#[tokio::main]
async fn main() {
let db = Builder::new_local(":memory:").build().await.unwrap();
let conn = db.connect().unwrap();
conn.execute("SELECT load_extension('./target/debug/liblimbo_series');", ())
.await
.unwrap();
loop {
conn.execute("SELECT * FROM generate_series(1,10,2);", ())
.await
.unwrap();
}
}
```
After switching to the system allocator, the leak becomes detectable
with Valgrind:
```
32,000 bytes in 1,000 blocks are definitely lost in loss record 24 of 24
at 0x538580F: malloc (vg_replace_malloc.c:446)
by 0x62E15FA: alloc::alloc::alloc (alloc.rs:99)
by 0x62E172C: alloc::alloc::Global::alloc_impl (alloc.rs:192)
by 0x62E1530: allocate (alloc.rs:254)
by 0x62E1530: alloc::alloc::exchange_malloc (alloc.rs:349)
by 0x62E0271: new<limbo_series::GenerateSeriesCursor> (boxed.rs:257)
by 0x62E0271: open_GenerateSeriesVTab (lib.rs:19)
by 0x425D8FA: limbo_core::VirtualTable::open (lib.rs:732)
by 0x4285DDA: limbo_core::vdbe::execute::op_vopen (execute.rs:890)
by 0x42351E8: limbo_core::vdbe::Program::step (mod.rs:396)
by 0x425C638: limbo_core::Statement::step (lib.rs:610)
by 0x40DB238: limbo::Statement::execute::{{closure}} (lib.rs:181)
by 0x40D9EAF: limbo::Connection::execute::{{closure}} (lib.rs:109)
by 0x40D54A1: example::main::{{closure}} (example.rs:26)
```
Interestingly, when using mimalloc, neither Valgrind nor mimalloc’s
internal statistics report the leak.
This commit adds suport for DROP INDEX.
Bytecode produced by this commit differs from SQLITE's bytecode, main
reason we don't do autovacuum or repacking of pages like SQLITE does.
Closes#1413 . Basically, SQLite emits a check in a transaction to see
if it is attempting to write. If the db is in read only mode, it throws
an error, else the statement is executed. Mirroring how Rusqlite does
it, I modified the `OpenFlags` to use bitflags to better configure how
we open our VFS. This modification, will enable us to run tests against
the same database in parallel.
Closes#1433
we had an incorrect optimization in `eliminate_orderby_like_groupby()`
where it could remove e.g. the first term of the ORDER BY if it matched
the first GROUP BY term and the result set was naturally ordered by that
term. this is invalid. see e.g.:
```sql
main branch - BAD: removes the `ORDER BY id` term because the results are naturally ordered by id.
However, this results in sorting the entire thing by last name only!
limbo> select id, last_name, count(1) from users GROUP BY 1,2 order by id, last_name desc limit 3;
┌──────┬───────────┬───────────┐
│ id │ last_name │ count (1) │
├──────┼───────────┼───────────┤
│ 6235 │ Zuniga │ 1 │
├──────┼───────────┼───────────┤
│ 8043 │ Zuniga │ 1 │
├──────┼───────────┼───────────┤
│ 944 │ Zimmerman │ 1 │
└──────┴───────────┴───────────┘
after fix - GOOD:
limbo> select id, last_name, count(1) from users GROUP BY 1,2 order by id, last_name desc limit 3;
┌────┬───────────┬───────────┐
│ id │ last_name │ count (1) │
├────┼───────────┼───────────┤
│ 1 │ Foster │ 1 │
├────┼───────────┼───────────┤
│ 2 │ Salazar │ 1 │
├────┼───────────┼───────────┤
│ 3 │ Perry │ 1 │
└────┴───────────┴───────────┘
I also refactored sorters to always use the ast `SortOrder` instead of boolean vectors, and use the `compare_immutable()` utility we use inside btrees too.
Closes#1365
This PR adds `PRAGMA schema_version` to get the value of the schema-
version integer at offset 40 in the database header.
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#1427
Since `page_size` in `DatabaseHeader` can be 1 representing 65526 bytes,
it can't be used it directly. Additionally, we should use `u32` instead
of `u16` or `usize` in other contexts.
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#1411
This issue was introduced in #819. However, I believe the solution is
suboptimal because `pragma page_count` can never return 1, which is
inconsistent with SQLite.
<img width="442" alt="image" src="https://github.com/user-
attachments/assets/c772eae7-3e9f-4687-a94a-230deb0eb034" />
To align with SQLite's behavior, we should allocate the first page when
the first schema object is created, rather than immediately after
creating database. And it's always preferable to return an accurate page
count.
Reviewed-by: Pere Diaz Bou <pere-altea@homail.com>
Closes#1407
DeleteState had a bit too many unnecessary states so I removed them.
Usually we care about having a different state when I/O is triggered
requiring a state to be stored for later.
Furthermore, there was a bug with op_idx_delete where if balance is
triggered, op_idx_delete wouldn't be re-entrant. So a state machine was
added to prevent that from happening.
Previously `DELETE FROM ...` only emitted deletes for main table, but
this is incorrect as we want to remove entries from index tables as
well.
Closes#1383
## Problem:
- We have cases where we are evaluating expressions in a hot loop that
could only be evaluated once. For example: `CAST('2025-01-01' as
DATETIME)` -- the value of this never changes, so we should only run it
once.
- We have no robust way of doing this right now for entire _expressions_
-- the only existing facility we have is
`program.mark_last_insn_constant()`, which has no concept of how many
instructions translating a given _expression_ spends, and breaks very
easily for this reason.
## Main ideas of this PR:
- Add `expr.is_constant()` determining whether the expression is
compile-time constant. Tries to be conservative and not deem something
compile-time constant if there is no certainty.
- Whenever we think a compile-time constant expression is about to be
translated into bytecode in `translate_expr()`, start a so called
`constant span`, which means a range of instructions that are part of a
compile-time constant expression.
- At the end of translating the program, all `constant spans` are
hoisted outside of any table loops so they only get evaluated once.
- The target offsets of any jump instructions (e.g. `Goto`) are moved to
the correct place, taking into account all instructions whose offsets
were shifted due to moving the compile-time constant expressions around.
- An escape hatch wrapper `translate_expr_no_constant_opt()` is added
for cases where we should not hoist constants even if we otherwise
could. Right now the only example of this is cases where we are reusing
the same register(s) in multiple iterations of some kind of loop, e.g.
`VALUES(...)` or in the `coalesce()` function implementation.
## Performance effects
Here is an example of a modified/simplified TPC-H query where the
`CAST()` calls were previously run millions of times in a hot loop, but
now they are optimized out of the loop.
**BYTECODE PLAN BEFORE:**
```sql
limbo> explain select
l_orderkey,
3 as revenue,
o_orderdate,
o_shippriority
from
lineitem,
orders,
customer
where
c_mktsegment = 'FURNITURE'
and c_custkey = o_custkey
and l_orderkey = o_orderkey
and o_orderdate < cast('1995-03-29' as datetime)
and l_shipdate > cast('1995-03-29' as datetime);
addr opcode p1 p2 p3 p4 p5 comment
---- ----------------- ---- ---- ---- ------------- -- -------
0 Init 0 26 0 0 Start at 26
1 OpenRead 0 10 0 0 table=lineitem, root=10
2 OpenRead 1 9 0 0 table=orders, root=9
3 OpenRead 2 8 0 0 table=customer, root=8
4 Rewind 0 25 0 0 Rewind lineitem
5 Column 0 10 5 0 r[5]=lineitem.l_shipdate
6 String8 0 7 0 1995-03-29 0 r[7]='1995-03-29'
7 Function 0 7 6 cast 0 r[6]=func(r[7..8]) <-- CAST() executed millions of times
8 Le 5 6 24 0 if r[5]<=r[6] goto 24
9 Column 0 0 9 0 r[9]=lineitem.l_orderkey
10 SeekRowid 1 9 24 0 if (r[9]!=orders.rowid) goto 24
11 Column 1 4 10 0 r[10]=orders.o_orderdate
12 String8 0 12 0 1995-03-29 0 r[12]='1995-03-29'
13 Function 0 12 11 cast 0 r[11]=func(r[12..13])
14 Ge 10 11 24 0 if r[10]>=r[11] goto 24
15 Column 1 1 14 0 r[14]=orders.o_custkey
16 SeekRowid 2 14 24 0 if (r[14]!=customer.rowid) goto 24
17 Column 2 6 15 0 r[15]=customer.c_mktsegment
18 Ne 15 16 24 0 if r[15]!=r[16] goto 24
19 Column 0 0 1 0 r[1]=lineitem.l_orderkey
20 Integer 3 2 0 0 r[2]=3
21 Column 1 4 3 0 r[3]=orders.o_orderdate
22 Column 1 7 4 0 r[4]=orders.o_shippriority
23 ResultRow 1 4 0 0 output=r[1..4]
24 Next 0 5 0 0
25 Halt 0 0 0 0
26 Transaction 0 0 0 0 write=false
27 String8 0 8 0 DATETIME 0 r[8]='DATETIME'
28 String8 0 13 0 DATETIME 0 r[13]='DATETIME'
29 String8 0 16 0 FURNITURE 0 r[16]='FURNITURE'
30 Goto 0 1 0
```
**BYTECODE PLAN AFTER**:
```sql
limbo> explain select
l_orderkey,
3 as revenue,
o_orderdate,
o_shippriority
from
lineitem,
orders,
customer
where
c_mktsegment = 'FURNITURE'
and c_custkey = o_custkey
and l_orderkey = o_orderkey
and o_orderdate < cast('1995-03-29' as datetime)
and l_shipdate > cast('1995-03-29' as datetime);
addr opcode p1 p2 p3 p4 p5 comment
---- ----------------- ---- ---- ---- ------------- -- -------
0 Init 0 21 0 0 Start at 21
1 OpenRead 0 10 0 0 table=lineitem, root=10
2 OpenRead 1 9 0 0 table=orders, root=9
3 OpenRead 2 8 0 0 table=customer, root=8
4 Rewind 0 20 0 0 Rewind lineitem
5 Column 0 10 5 0 r[5]=lineitem.l_shipdate
6 Le 5 6 19 0 if r[5]<=r[6] goto 19
7 Column 0 0 9 0 r[9]=lineitem.l_orderkey
8 SeekRowid 1 9 19 0 if (r[9]!=orders.rowid) goto 19
9 Column 1 4 10 0 r[10]=orders.o_orderdate
10 Ge 10 11 19 0 if r[10]>=r[11] goto 19
11 Column 1 1 14 0 r[14]=orders.o_custkey
12 SeekRowid 2 14 19 0 if (r[14]!=customer.rowid) goto 19
13 Column 2 6 15 0 r[15]=customer.c_mktsegment
14 Ne 15 16 19 0 if r[15]!=r[16] goto 19
15 Column 0 0 1 0 r[1]=lineitem.l_orderkey
16 Column 1 4 3 0 r[3]=orders.o_orderdate
17 Column 1 7 4 0 r[4]=orders.o_shippriority
18 ResultRow 1 4 0 0 output=r[1..4]
19 Next 0 5 0 0
20 Halt 0 0 0 0
21 Transaction 0 0 0 0 write=false
22 String8 0 7 0 1995-03-29 0 r[7]='1995-03-29'
23 String8 0 8 0 DATETIME 0 r[8]='DATETIME'
24 Function 1 7 6 cast 0 r[6]=func(r[7..8]) <-- CAST() executed twice
25 String8 0 12 0 1995-03-29 0 r[12]='1995-03-29'
26 String8 0 13 0 DATETIME 0 r[13]='DATETIME'
27 Function 1 12 11 cast 0 r[11]=func(r[12..13])
28 String8 0 16 0 FURNITURE 0 r[16]='FURNITURE'
29 Integer 3 2 0 0 r[2]=3
30 Goto 0 1 0 0
```
**EXECUTION RUNTIME BEFORE:**
```sql
limbo> select
l_orderkey,
3 as revenue,
o_orderdate,
o_shippriority
from
lineitem,
orders,
customer
where
c_mktsegment = 'FURNITURE'
and c_custkey = o_custkey
and l_orderkey = o_orderkey
and o_orderdate < cast('1995-03-29' as datetime)
and l_shipdate > cast('1995-03-29' as datetime);
┌────────────┬─────────┬─────────────┬────────────────┐
│ l_orderkey │ revenue │ o_orderdate │ o_shippriority │
├────────────┼─────────┼─────────────┼────────────────┤
└────────────┴─────────┴─────────────┴────────────────┘
Command stats:
----------------------------
total: 3.633396667 s (this includes parsing/coloring of cli app)
```
**EXECUTION RUNTIME AFTER:**
```sql
limbo> select
l_orderkey,
3 as revenue,
o_orderdate,
o_shippriority
from
lineitem,
orders,
customer
where
c_mktsegment = 'FURNITURE'
and c_custkey = o_custkey
and l_orderkey = o_orderkey
and o_orderdate < cast('1995-03-29' as datetime)
and l_shipdate > cast('1995-03-29' as datetime);
┌────────────┬─────────┬─────────────┬────────────────┐
│ l_orderkey │ revenue │ o_orderdate │ o_shippriority │
├────────────┼─────────┼─────────────┼────────────────┤
└────────────┴─────────┴─────────────┴────────────────┘
Command stats:
----------------------------
total: 2.0923475 s (this includes parsing/coloring of cli app)
````
Reviewed-by: Pere Diaz Bou <pere-altea@homail.com>
Closes#1359
Closes#1384 . This PR implements Primary Key constraint for inserts. As
can be seen in the issue, if you created an Index with a Primary Key
constraint, it could trigger `Unique Constraint` error, but still insert
the record. Sqlite uses the opcode `NoConflict` to check if the record
already exists in the Btree. As we did not have this Opcode yet, I
implemented it. It is very similar to `NotFound` with the difference
that if any value in the Record is Null, it will immediately jump to the
offset. The added benefit of implementing this, is that now we fully
support Composite Primary Keys. Also, I think with the current
implementation, it will be trivial to implement the Unique opcode for
Insert. To support Updates, I need to understand more of the plan
optimizer to and find where we are Making the Record and opening the
autoindex.
For testing, I have written a test generator to generate many different
tables that can have a varying numbers of Primary Keys.
```sql
limbo> CREATE TABLE users (id INT, username TEXT, PRIMARY KEY (id, username));
limbo> INSERT INTO users VALUES (1, 'alice');
limbo> explain INSERT INTO users VALUES (1, 'alice');
addr opcode p1 p2 p3 p4 p5 comment
---- ----------------- ---- ---- ---- ------------- -- -------
0 Init 0 16 0 0 Start at 16
1 OpenWrite 0 2 0 0
2 Integer 1 2 0 0 r[2]=1
3 String8 0 3 0 alice 0 r[3]='alice'
4 OpenWrite 1 3 0 0
5 NewRowId 0 1 0 0
6 Copy 2 5 0 0 r[5]=r[2]
7 Copy 3 6 0 0 r[6]=r[3]
8 Copy 1 7 0 0 r[7]=r[1]
9 MakeRecord 5 3 8 0 r[8]=mkrec(r[5..7])
10 NoConflict 1 12 5 2 0 key=r[5]
11 Halt 1555 0 0 users.id, users.username 0
12 IdxInsert 1 8 5 0 key=r[8]
13 MakeRecord 2 2 4 0 r[4]=mkrec(r[2..3])
14 Insert 0 4 1 0
15 Halt 0 0 0 0
16 Transaction 0 1 0 0 write=true
17 Goto 0 1 0 0
limbo> INSERT INTO users VALUES (1, 'alice');
× Runtime error: UNIQUE constraint failed: users.id, users.username (19)
limbo> INSERT INTO users VALUES (1, 'bob');
limbo> INSERT INTO users VALUES (1, 'bob');
× Runtime error: UNIQUE constraint failed: users.id, users.username (19)
```
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#1393
1) Fix a bug where cli pretty mode would not print pragma results;
2) Add ability to read page_size using PRAGMA page_size;
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#1394
Apply affinities to a range of P2 registers starting with P1.
P4 is a string that is P2 characters long. The N-th character of the string indicates the column affinity that should be used for the N-th memory cell in the range.
If P4==0 then register P3 holds a blob constructed by MakeRecord. If P4>0 then register P3 is the first of P4 registers that form an unpacked record.
Cursor P1 is on an index btree. If the record identified by P3 and P4 is not the prefix of any entry in P1 then a jump is made to P2. If P1 does contain an entry whose prefix matches the P3/P4 record then control falls through to the next instruction and P1 is left pointing at the matching entry.
This operation leaves the cursor in a state where it cannot be advanced in either direction. In other words, the Next and Prev opcodes do not work after this operation.
