1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
//! Deep learning models are usually implemented to make efficient use of a GPU by batching inputs together
//! in "mini-batches". However, applications serving these models often receive requests one-by-one.
//! So using a conventional single or multi-threaded server approach will under-utilize the GPU and lead to latency that increases
//! linearly with the volume of requests.
//!
//! `batched-fn` is a drop-in solution for deep learning webservers that queues individual requests and provides them as a batch
//! to your model. It can be added to any application with minimal refactoring simply by inserting the [`batched_fn`](crate::batched_fn)
//! macro into the function that runs requests through the model.
//!
//! ## Features
//!
//! - 🚀 Easy to use: drop the `batched_fn!` macro into existing code.
//! - 🔥 Lightweight and fast: queue system implemented on top of the blazingly fast [flume crate](https://github.com/zesterer/flume).
//! - 🙌 Easy to tune: simply adjust [`max_delay`](crate::batched_fn#config) and [`max_batch_size`](crate::batched_fn#config).
//! - 🛑 [Back pressure](https://medium.com/@jayphelps/backpressure-explained-the-flow-of-data-through-software-2350b3e77ce7) mechanism included:
//! just set [`channel_cap`](crate::batched_fn#config) and handle
//! [`Error::Full`](crate::Error#variant.Full) by returning a 503 from your webserver.
//!
//! ## Examples
//!
//! Suppose you have a model API that look like this:
//!
//! ```rust
//! // `Batch` could be anything that implements the `batched_fn::Batch` trait.
//! type Batch<T> = Vec<T>;
//!
//! #[derive(Debug)]
//! struct Input {
//! // ...
//! }
//!
//! #[derive(Debug)]
//! struct Output {
//! // ...
//! }
//!
//! struct Model {
//! // ...
//! }
//!
//! impl Model {
//! fn predict(&self, batch: Batch<Input>) -> Batch<Output> {
//! // ...
//! # batch.iter().map(|_| Output {}).collect()
//! }
//!
//! fn load() -> Self {
//! // ...
//! # Self {}
//! }
//! }
//! ```
//!
//! Without `batched-fn` a webserver route would need to call `Model::predict` on each
//! individual input, resulting in a bottleneck from under-utilizing the GPU:
//!
//! ```rust
//! use once_cell::sync::Lazy;
//! # use batched_fn::{batched_fn, Batch as BatchTrait};
//! # type Batch<T> = Vec<T>;
//! # #[derive(Debug)]
//! # struct Input {}
//! # #[derive(Debug)]
//! # struct Output {}
//! # struct Model {}
//! # impl Model {
//! # fn predict(&self, batch: Batch<Input>) -> Batch<Output> {
//! # batch.iter().map(|_| Output {}).collect()
//! # }
//! # fn load() -> Self { Self {} }
//! # }
//! static MODEL: Lazy<Model> = Lazy::new(Model::load);
//!
//! fn predict_for_http_request(input: Input) -> Output {
//! let mut batched_input = Batch::with_capacity(1);
//! batched_input.push(input);
//! MODEL.predict(batched_input).pop().unwrap()
//! }
//! ```
//!
//! But by dropping the [`batched_fn`](crate::batched_fn) macro into your code you automatically get batched
//! inference behind the scenes without changing the one-to-one relationship between inputs and
//! outputs:
//!
//! ```rust
//! # use batched_fn::{batched_fn, Batch as BatchTrait};
//! # type Batch<T> = Vec<T>;
//! # #[derive(Debug)]
//! # struct Input {}
//! # #[derive(Debug)]
//! # struct Output {}
//! # struct Model {}
//! # impl Model {
//! # fn predict(&self, batch: Batch<Input>) -> Batch<Output> {
//! # batch.iter().map(|_| Output {}).collect()
//! # }
//! # fn load() -> Self { Self {} }
//! # }
//! async fn predict_for_http_request(input: Input) -> Output {
//! let batch_predict = batched_fn! {
//! handler = |batch: Batch<Input>, model: &Model| -> Batch<Output> {
//! model.predict(batch)
//! };
//! config = {
//! max_batch_size: 16,
//! max_delay: 50,
//! };
//! context = {
//! model: Model::load(),
//! };
//! };
//! batch_predict(input).await.unwrap()
//! }
//! ```
//!
//! ❗️ *Note that the `predict_for_http_request` function now has to be `async`.*
//!
//! Here we set the [`max_batch_size`](crate::batched_fn#config) to 16 and [`max_delay`](crate::batched_fn#config)
//! to 50 milliseconds. This means the batched function will wait at most 50 milliseconds after receiving a single
//! input to fill a batch of 16. If 15 more inputs are not received within 50 milliseconds
//! then the partial batch will be ran as-is.
//!
//! ## Tuning max batch size and max delay
//!
//! The optimal batch size and delay will depend on the specifics of your use case, such as how big of a batch you can fit in memory
//! (typically on the order of 8, 16, 32, or 64 for a deep learning model) and how long of a delay you can afford.
//! In general you want to set `max_batch_size` as high as you can, assuming the total processing time for `N` examples is minimized
//! with a batch size of `N`, and keep `max_delay` small relative to the time it takes for your
//! handler function to process a batch.
//!
//! ## Implementation details
//!
//! When the `batched_fn` macro is invoked it spawns a new thread where the
//! [`handler`](crate::batched_fn#handler) will
//! be ran. Within that thread, every object specified in the [`context`](crate::batched_fn#context)
//! is initialized and then passed by reference to the handler each time it is run.
//!
//! The object returned by the macro is just a closure that sends a single input and a callback
//! through an asyncronous channel to the handler thread. When the handler finishes
//! running a batch it invokes the callback corresponding to each input with the corresponding output,
//! which triggers the closure to wake up and return the output.
extern crate flume;
extern crate once_cell;
#[doc(hidden)]
pub use flume::{bounded, unbounded, Sender};
#[doc(hidden)]
pub use once_cell::sync::Lazy;
/// The `Batch` trait is essentially an abstraction of `Vec<T>`. The input and output of a batch
/// [`handler`](crate::batched_fn#handler) must implement `Batch`.
///
/// It represents an owned collection of ordered items of a single type.
pub trait Batch: IntoIterator<Item = <Self as Batch>::Item> {
type Item;
fn with_capacity(n: usize) -> Self;
fn len(&self) -> usize;
fn push(&mut self, item: <Self as Batch>::Item);
fn is_empty(&self) -> bool {
self.len() == 0
}
}
impl<T> Batch for Vec<T> {
type Item = T;
fn with_capacity(n: usize) -> Vec<T> {
Vec::<T>::with_capacity(n)
}
fn len(&self) -> usize {
self.len()
}
fn push(&mut self, item: T) {
self.push(item);
}
}
#[doc(hidden)]
pub struct Config {
pub max_batch_size: usize,
pub max_delay: u128,
pub channel_cap: Option<usize>,
// Used to avoid clippy linting errors within the macro-generated code
// when updating the fields of this struct.
pub _phantom: std::marker::PhantomData<bool>,
}
impl Default for Config {
fn default() -> Self {
Self {
max_batch_size: 8,
max_delay: 50,
channel_cap: None,
_phantom: std::marker::PhantomData,
}
}
}
/// Error types that can occur while calling a batched function.
#[derive(Debug, Copy, Clone)]
pub enum Error {
/// Channel is full.
///
/// This can happen if you've set `channel_cap`, and should usually be handled
/// by returning a 503 error code from your server to signal that the server is too
/// busy at the moment to handle any more requests.
Full,
/// Channel has been disconnected, most likely due to the handler thread crashing.
Disconnected,
}
/// Created by the [`batched_fn`](crate::batched_fn) macro.
///
/// A `BatchedFn` is a wrapper around a [`handler`](crate::batched_fn#handler)
/// that provides the interface for evaluating a single input as part of a batch of other inputs.
pub struct BatchedFn<T, R>
where
T: 'static + Send + Sync + std::fmt::Debug,
R: 'static + Send + Sync + std::fmt::Debug,
{
tx: Sender<(T, Sender<R>)>,
}
impl<T, R> BatchedFn<T, R>
where
T: 'static + Send + Sync + std::fmt::Debug,
R: 'static + Send + Sync + std::fmt::Debug,
{
pub fn new(tx: Sender<(T, Sender<R>)>) -> Self {
Self { tx }
}
/// Evaluate a single input as part of a batch of other inputs.
pub async fn evaluate_in_batch(&self, input: T) -> Result<R, Error> {
// Can use `unbounded` channel because we already get backpressure from
// the channel that `self.tx` sends to.
let (result_tx, result_rx) = unbounded::<R>();
self.tx.try_send((input, result_tx)).map_err(|e| match e {
flume::TrySendError::Full(_) => Error::Full,
flume::TrySendError::Disconnected(_) => Error::Disconnected,
})?;
result_rx
.recv_async()
.await
.map_err(|_| Error::Disconnected)
}
}
#[doc(hidden)]
#[macro_export]
macro_rules! __batched_fn_internal {
(
handler = |$batch:ident: $batch_input_type:ty $(, $ctx_arg:ident: &$ctx_arg_ty:ty )*| -> $batch_output_type:ty $fn_body:block ;
config = {
$( $cfg:ident: $cfg_init:expr ),* $(,)?
};
context = {
$( $ctx:ident: $ctx_init:expr ),* $(,)?
} $(;)?
) => {{
static BATCHED_FN: $crate::Lazy<
$crate::BatchedFn<
<$batch_input_type as $crate::Batch>::Item,
<$batch_output_type as $crate::Batch>::Item,
>,
> = $crate::Lazy::new(|| {
let config = $crate::Config {
$( $cfg: $cfg_init, )*
..Default::default()
};
let (tx, mut rx) = match config.channel_cap {
None => {
$crate::unbounded::<(
<$batch_input_type as $crate::Batch>::Item,
$crate::Sender<<$batch_output_type as $crate::Batch>::Item>,
)>()
}
Some(cap) => {
$crate::bounded::<(
<$batch_input_type as $crate::Batch>::Item,
$crate::Sender<<$batch_output_type as $crate::Batch>::Item>,
)>(cap)
}
};
std::thread::spawn(move || {
// Create handler closure.
let handler = |$batch: $batch_input_type $(, $ctx_arg: &$ctx_arg_ty )*| -> $batch_output_type {
$fn_body
};
// Set config vars.
let max_batch_size: usize = config.max_batch_size;
let max_delay: u128 = config.max_delay;
// Initialize handler context.
struct _Context {
$( $ctx_arg: $ctx_arg_ty, )*
}
let context = _Context {
$( $ctx: $ctx_init, )*
};
// Wait for an input.
while let Ok((input, result_tx)) = rx.recv() {
let mut batch_input =
<$batch_input_type as $crate::Batch>::with_capacity(max_batch_size);
let mut batch_txs = Vec::with_capacity(max_batch_size);
batch_input.push(input);
batch_txs.push(result_tx);
let mut vacancy = max_batch_size - 1;
let mut time_left = max_delay as u64;
let start = std::time::Instant::now();
// While there is still room in the batch we'll wait at most `max_delay`
// milliseconds to try to fill it.
while vacancy > 0 && time_left > 0 {
if let Ok((next_input, next_result_tx)) =
rx.recv_timeout(std::time::Duration::from_millis(time_left))
{
batch_input.push(next_input);
batch_txs.push(next_result_tx);
vacancy -= 1;
let elapsed = start.elapsed().as_millis();
time_left = if elapsed > max_delay {
0
} else {
(max_delay - elapsed) as u64
};
} else {
break;
}
}
let batch_output = handler(batch_input $(, &context.$ctx_arg )*);
for (output, mut result_tx) in batch_output.into_iter().zip(batch_txs) {
result_tx.send(output).ok();
}
}
});
$crate::BatchedFn::new(tx)
});
|input| BATCHED_FN.evaluate_in_batch(input)
}};
}
/// Macro for creating a batched function.
///
/// This macro has 3 parameters: [`handler`](#handler), [`config`](#config), and
/// [`context`](#context). It returns an async function that wraps
/// [`BatchedFn::evaluate_in_batch`](struct.BatchedFn.html#method.evaluate_in_batch).
///
/// # Parameters
///
/// ### `handler`
///
/// The handler must be in the form of a closure declaration that takes a batch
/// and any number of references to objects in the context as input and
/// returns a different type of batch.
///
/// ### `config`
///
/// Within the config you can specify the `max_batch_size`, `max_delay`, and `channel_cap`.
///
/// The batched function will wait at most `max_delay` milliseconds after receiving a single
/// input to fill a batch of size `max_batch_size`. If enough inputs to fill a full batch
/// are not received within `max_delay` milliseconds then the partial batch will be ran as-is.
///
/// The `channel_cap` option allows you to apply back pressure if too many inputs are waiting for
/// the handler thread to accept another batch. By default `channel_cap` is `None`, but if
/// set to `Some(usize)` then
/// [`BatchedFn::evaluate_in_batch`](struct.BatchedFn.html#method.evaluate_in_batch) will
/// return [`Error::Full`](crate::Error#variant.Full) if the channel between the calling thread and the handler thread is at this
/// capacity. You probably want to set this to some multiple of `max_batch_size`.
///
/// ## `context`
///
/// Any additional reference that the handler takes as input must be defined within
/// the context.
///
/// # Examples
///
/// ```rust
/// # #[macro_use] extern crate batched_fn;
/// use batched_fn::{batched_fn, Error};
///
/// async fn double(x: i32) -> Result<i32, Error> {
/// let batched_double = batched_fn! {
/// handler = |batch: Vec<i32>| -> Vec<i32> {
/// batch.into_iter().map(|x| x*2).collect()
/// };
/// config = {
/// max_batch_size: 4,
/// max_delay: 50,
/// channel_cap: Some(20),
/// };
/// context = {};
/// };
///
/// batched_double(x).await
/// }
/// ```
///
/// You can also provide an arbitrary number of additional arguments to the handler by reference.
/// All of the objects have to be initialized in the [`context`](#context):
///
/// ```rust
/// # #[macro_use] extern crate batched_fn;
/// # use batched_fn::{batched_fn, Error};
/// async fn multiply(x: i32) -> Result<i32, Error> {
/// let batched_multiply = batched_fn! {
/// handler = |batch: Vec<i32>, factor: &i32| -> Vec<i32> {
/// batch.into_iter().map(|x| *factor * x ).collect()
/// };
/// config = {
/// max_batch_size: 4,
/// max_delay: 50
/// };
/// context = {
/// factor: 3
/// };
/// };
///
/// batched_multiply(x).await
/// }
/// ```
#[macro_export]
macro_rules! batched_fn {
(
handler = |$batch:ident: $batch_input_type:ty $(, $ctx_arg:ident: &$ctx_arg_ty:ty )*| -> $batch_output_type:ty $fn_body:block ;
config = {
$( $cfg:ident: $cfg_init:expr ),* $(,)?
};
context = {
$( $ctx:ident: $ctx_init:expr ),* $(,)?
} $(;)?
) => {
$crate::__batched_fn_internal!(
handler = |$batch: $batch_input_type $(, $ctx_arg: &$ctx_arg_ty )*| -> $batch_output_type $fn_body ;
config = {
$( $cfg: $cfg_init, )*
};
context = {
$( $ctx: $ctx_init, )*
};
);
};
}