Use knowledge of query plan to inform how much memory to initially
allocate for `ProgramBuilder` vectors
Some of them are exact, some are semi-random estimates
```sql
Prepare `SELECT 1`/Limbo/SELECT 1
time: [756.93 ns 758.11 ns 759.59 ns]
change: [-4.5974% -4.3153% -4.0393%] (p = 0.00 < 0.05)
Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
2 (2.00%) low severe
1 (1.00%) low mild
3 (3.00%) high mild
1 (1.00%) high severe
Prepare `SELECT * FROM users LIMIT 1`/Limbo/SELECT * FROM users LIMIT 1
time: [1.4739 µs 1.4769 µs 1.4800 µs]
change: [-7.9364% -7.7171% -7.4979%] (p = 0.00 < 0.05)
Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
1 (1.00%) high mild
Prepare `SELECT first_name, count(1) FROM users GROUP BY first_name HAVING count(1) > 1 ORDER BY cou...`
time: [3.7440 µs 3.7520 µs 3.7596 µs]
change: [-5.4627% -5.1578% -4.8445%] (p = 0.00 < 0.05)
Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
1 (1.00%) high severe
```
Closes#899
This PR adds support for `json_set`.
There are three helper functions added:
1. `json_path_from_owned_value`, this function turns an `OwnedValue`
into a `JsonPath`.
2. `find_or_create_target`, this function is similar to `find_target`
with the added bonus of creating the target if it doesn't exist. There
is a caveat with this function and that is that it will create
objects/arrays as it goes, meaning if you send `{}` into it and try
getting the path `$.some.nested.array[123].field`, it will return
`{"some":{"nested":array:[]}}` since creation of `some`, `nested` and
`array` will succeed, but accessing element `123` will fail.
3. `create_and_mutate_json_by_path`, this function is very similar to
`mutate_json_by_path` but calls `find_or_create_target` instead of
`find_target`
Related to #127Closes#878
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]
```
I was baffled previously, because any time that `free` was called on a
type from an extension, it would hang even when I knew it wasn't in use
any longer, and hadn't been double free'd.
After #737 was merged, I tried it again and noticed that it would no
longer hang... but only for extensions that were staticly linked.
Then I realized that we are using a global allocator, that likely wasn't
getting used in the shared library that is built separately that won't
inherit from our global allocator in core, causing some symbol mismatch
and the subsequent hanging on calls to `free`.
This PR adds the global allocator to extensions behind a feature flag in
the macro that will prevent it from being used in `wasm` and staticly
linked environments where it would conflict with limbos normal global
allocator. This allows us to properly free the memory from returning
extension functions over FFI.
This PR also changes the Extension type to a union field so we can store
int + float values inline without boxing them.
any additional tips or thoughts anyone else has on improving this would
be appreciated 👍Closes#803
This patch adds some libSQL vector extension functions such as
`vector()` and `vector_distance_cos()`, which can be used for exact
nearest neighbor search as follows:
```
limbo> SELECT embedding, vector_distance_cos(embedding, '[9, 9, 9]')
...> FROM movies ORDER BY vector_distance_cos(embedding, '[9, 9, 9]');
[4, 5, 6]|0.013072490692138672
[1, 2, 3]|0.07417994737625122
```
The current status of the PR is halfway. The new framing of simulation
runner where `setup_simulation` is separated from `run_simulation`
allows for injecting custom plans easily. The PR is currently missing
the functionality to update the `SimulatorEnv` ad hoc from the plan, as
the environment tables were typically created during the planning phase.
The next steps will be to implement a function `fn
mk_env(InteractionPlan, SimulatorEnv) -> SimulatorEnv`, add `--load`
flag to the CLI for loading a serialized plan file, making a
corresponding environment and running the simulation.
We can optionally combine this with a `--save` option, in which we keep
a seed-vault as part of limbo simulator, corresponding each seed with
its generated plan and save the time to regenerate existing seeds by
just loading them into memory. I am curious to hear thoughts on this?
Would the maintainers be open to adding such a seed-vault? Do you think
the saved time would be worth the complexity of the approach?
Reviewed-by: Pere Diaz Bou <pere-altea@homail.com>
Closes#720
- add `--watch` flag
- start saving seeds in persistent storage
- make a separate version of execution functions that use `vector of interaction` instead of `InteractionPlan`
#708
This PR adds basic support for the following API for defining
Aggregates, and changes the existing API for scalars.
```rust
register_extension! {
scalars: { Double },
aggregates: { MedianState },
}
#[derive(ScalarDerive)]
struct Double;
impl Scalar for Double {
fn name(&self) -> &'static str {
"double"
}
fn call(&self, args: &[Value]) -> Value {
if let Some(arg) = args.first() {
match arg.value_type() {
ValueType::Float => {
let val = arg.to_float().unwrap();
Value::from_float(val * 2.0)
}
ValueType::Integer => {
let val = arg.to_integer().unwrap();
Value::from_integer(val * 2)
}
_ => {
println!("arg: {:?}", arg);
Value::null()
}
}
} else {
Value::null()
}
}
}
#[derive(AggregateDerive)]
struct MedianState;
impl AggFunc for MedianState {
type State = Vec<f64>;
fn name(&self) -> &'static str {
"median"
}
fn args(&self) -> i32 { 1 }
fn step(state: &mut Self::State, args: &[Value]) {
if let Some(val) = args.first().and_then(Value::to_float) {
state.push(val);
}
}
fn finalize(state: Self::State) -> Value {
if state.is_empty() {
return Value::null();
}
let mut sorted = state;
sorted.sort_by(|a, b| a.partial_cmp(b).unwrap());
let len = sorted.len();
if len % 2 == 1 {
Value::from_float(sorted[len / 2])
} else {
let mid1 = sorted[len / 2 - 1];
let mid2 = sorted[len / 2];
Value::from_float((mid1 + mid2) / 2.0)
}
}
}
```
I know it's a bit more verbose than the previous version, but I think in
the long run this will be more robust, and it matches the aggregate API
of implementing a trait on a struct that you derive the relevant trait
on.
Also this allows for better registration of functions, I think passing
in the struct identifiers just feels much better than the `"func_name"
=> function_ptr`
Closes#721
- makes interaction plans serializable
- fixes the shadowing bug where non-created tables were assumed to be created in the shadow tables map
- makes small changes to make clippy happy
- reorganizes simulation running flow to remove unnecessary plan regenerations while shrinking and double checking
