Currently we have a Record, which is a dumb vector of cloned values.
This is incredibly bad for performance as we do not want to clone
objects unless needed. Therefore, let's start by introducing this type
so that any record that has already been serialized will be returned
from btree in the format of a simple payload with reference to payload.
Closes#1176
- Use `CheckpointResult::default()` instead of `CheckpointResult::new()`
- Correct WAL frame header salt and checksum handling
- Ensure frame ID is 1-based and adjust frame offset calculation
- Add `Default` implementation for `CheckpointResult`
- Use random values for WAL header salts
Currently we have a Record, which is a dumb vector of cloned values.
This is incredibly bad for performance as we do not want to clone
objects unless needed. Therefore, let's start by introducing this type
so that any record that has already been serialized will be returned
from btree in the format of a simple payload with reference to payload.
There are two bugs in #1085.
1. `find_free_cell` accesses non-existent free blocks and returns their
size to `allocate_space`. This is out of range access error. The fix is
to add a loop termination condition that stops it when we hit the end of
the blocks
2. This bug is caused by `find_free_cell` some how swallowing the blocks
with size `4 bytes`. So `compute_free_space` consistently undercounts by
`4 bytes`. I've refactored that part of the code to make sure 4 sized
block are not deleted.
I've also added a unit test which proves these fixes work and also added
a function called `debug_print_freelist` which prints all free blocks of
a page.
For now I've silenced the `overflow_page` tests.
Fixes#1085Closes#1111
When I added frame reading support I thought, okay, who cares about the page id of this page it we read it from a frame because we don't need it to compute the offset to read from the file in this case. Fuck me, because it was needed in case we read `page 1` from WAL because it has a differnt `offset`.
Beep boop.
What happened you ask? I removed the dumb balancing algorithm I
implemented in favor of SQLite's implementation based on B*Tree[1] where
a page is 2/3 full instead of 1/2. It also tries to balance a page by
taking a maximum 3 pages and distributing cells evenly between them.
I've made some changes that are somewhat related:
* Moved most operations on pages out of BTreeCursor because those
operations are based on a page, not on a cursor, and it makes it easier
to test.
* Fixed `write_u16` and `read_u16` cases that didn't need a implicit
offset calculation. Added: `write_u16_no_offset` and
`read_u16_no_offset` to counter this.
* Added some tests with fuzz testing too.
* Fixed some important actions like: `compute_free_space`,
`defragment_page` and `drop_cell`.
[1] https://dl.acm.org/doi/10.1145/356770.356776Closes#968
The SerialType::try_from() was pretty high up in CPU profiles so I asked
my dear friend Claude to optimize it away by using the serial type
integer value directly instead of constructing a fancy enumeration.
I am on a bit of a mission to revisit a lot of the ref counting, this
was an easy first win.
It seems to be a linear path of function calls or hashmaps which can own
the completions directly, no cloning needed.
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#912
This PR aims to simplify text creation and to reduce allocation overhead
of creating a new OwnedValue::Text. Instead of creating Rc<String> every
time you need to create a text, you just pass the string slice and the
Rc<String> is created at the end. This change, at least on my machine,
has removed a lot of variability in the benchmarking performance, while
maintaining roughly the same performance.
Closes#961
We really need to make the WAL lock less expensive, but switching to
`parking_lot` is anyway something we should do.
Before:
```
Execute `SELECT 1`/Limbo
time: [56.230 ns 56.463 ns 56.688 ns]
```
After:
```
Execute `SELECT 1`/Limbo
time: [52.003 ns 52.132 ns 52.287 ns]
```
