nushell/crates/nu-protocol/src/value/lazy_record.rs
Reilly Wood 3b5172a8fa
LazyRecord (#7619)
This is an attempt to implement a new `Value::LazyRecord` variant for
performance reasons.

`LazyRecord` is like a regular `Record`, but it's possible to access
individual columns without evaluating other columns. I've implemented
`LazyRecord` for the special `$nu` variable; accessing `$nu` is
relatively slow because of all the information in `scope`, and [`$nu`
accounts for about 2/3 of Nu's startup time on
Linux](https://github.com/nushell/nushell/issues/6677#issuecomment-1364618122).

### Benchmarks

I ran some benchmarks on my desktop (Linux, 12900K) and the results are
very pleasing.

Nu's time to start up and run a command (`cargo build --release;
hyperfine 'target/release/nu -c "echo \"Hello, world!\""' --shell=none
--warmup 10`) goes from **8.8ms to 3.2ms, about 2.8x faster**.

Tests are also much faster! Running `cargo nextest` (with our very slow
`proptest` tests disabled) goes from **7.2s to 4.4s (1.6x faster)**,
because most tests involve launching a new instance of Nu.

### Design (updated)

I've added a new `LazyRecord` trait and added a `Value` variant wrapping
those trait objects, much like `CustomValue`. `LazyRecord`
implementations must implement these 2 functions:

```rust
// All column names
fn column_names(&self) -> Vec<&'static str>;

// Get 1 specific column value
fn get_column_value(&self, column: &str) -> Result<Value, ShellError>;
 ```

### Serializability

`Value` variants must implement `Serializable` and `Deserializable`, which poses some problems because I want to use unserializable things like `EngineState` in `LazyRecord`s. To work around this, I basically lie to the type system:

1. Add `#[typetag::serde(tag = "type")]` to `LazyRecord` to make it serializable
2. Any unserializable fields in `LazyRecord` implementations get marked with `#[serde(skip)]`
3. At the point where a `LazyRecord` normally would get serialized and sent to a plugin, I instead collect it into a regular `Value::Record` (which can be serialized)
2023-01-18 19:27:26 -08:00

35 lines
1.1 KiB
Rust

use crate::{ShellError, Span, Value};
use std::fmt;
// Trait definition for a lazy record (where columns are evaluated on-demand)
// typetag is needed to make this implement Serialize+Deserialize... even though we should never actually serialize a LazyRecord.
// To serialize a LazyRecord, collect it into a Value::Record with collect() first.
#[typetag::serde(tag = "type")]
pub trait LazyRecord: fmt::Debug + Send + Sync {
// All column names
fn column_names(&self) -> Vec<&'static str>;
// Get 1 specific column value
fn get_column_value(&self, column: &str) -> Result<Value, ShellError>;
fn span(&self) -> Span;
// Convert the lazy record into a regular Value::Record by collecting all its columns
fn collect(&self) -> Result<Value, ShellError> {
let mut cols = vec![];
let mut vals = vec![];
for column in self.column_names() {
cols.push(column.into());
let val = self.get_column_value(column)?;
vals.push(val);
}
Ok(Value::Record {
cols,
vals,
span: self.span(),
})
}
}