#queue #persistent #disk #data-structures

yaque

Yaque is yet another disk-backed persistent queue for Rust

18 releases

0.6.6 Nov 4, 2023
0.6.4 Aug 21, 2022
0.6.3 Feb 4, 2022
0.6.2 May 15, 2021
0.4.2 Jun 25, 2020

#56 in Concurrency

Download history 269/week @ 2024-07-20 238/week @ 2024-07-27 145/week @ 2024-08-03 582/week @ 2024-08-10 878/week @ 2024-08-17 637/week @ 2024-08-24 450/week @ 2024-08-31 1117/week @ 2024-09-07 416/week @ 2024-09-14 603/week @ 2024-09-21 1085/week @ 2024-09-28 833/week @ 2024-10-05 533/week @ 2024-10-12 612/week @ 2024-10-19 1119/week @ 2024-10-26 460/week @ 2024-11-02

2,838 downloads per month
Used in atm0s-media-server

Apache-2.0

150KB
2.5K SLoC

Yaque: Yet Another QUEue

Yaque is yet another disk-backed persistent queue (and mutex) for Rust. It implements an SPSC channel using your OS' filesystem. Its main advantages over a simple VecDeque<T> are that

  • You are not constrained by your RAM size, just by your disk size. This means you can store gigabytes of data without getting OOM killed.
  • Your data is safe even if you program panics. All the queue state is written to the disk when the queue is dropped.
  • Your data can persist, that is, can exist through multiple executions of your program. Think of it as a very rudimentary kind of database.
  • You can pass data between two processes.

Yaque is asynchronous and built directly on top of mio and notify. It is therefore completely agnostic to the runtime you are using for you application. It will work smoothly with tokio, with async-std or any other executor of your choice.

Sample usage

To create a new queue, just use the channel function, passing a directory path on which to mount the queue. If the directory does not exist on creation, it (and possibly all its parent directories) will be created.

use yaque::channel;

futures::executor::block_on(async {
    let (mut sender, mut receiver) = channel("data/my-queue").unwrap();
})

You can also use Sender::open and Receiver::open to open only one half of the channel, if you need to.

The usage is similar to the MPSC channel in the standard library, except that the receiving method, Receiver::recv is asynchronous. Writing to the queue with the sender is basically lock-free and atomic.

use yaque::{channel, queue::try_clear};

futures::executor::block_on(async {
    // Open using the `channel` function or directly with the constructors.
    let (mut sender, mut receiver) = channel("data/my-queue").unwrap();
    
    // Send stuff with the sender...
    sender.send(b"some data").await.unwrap();

    // ... and receive it in the other side.
    let data = receiver.recv().await.unwrap();

    assert_eq!(&*data, b"some data");

    // Call this to make the changes to the queue permanent.
    // Not calling it will revert the state of the queue.
    data.commit();
});

// After everything is said and done, you may delete the queue.
// Use `clear` for awaiting for the queue to be released.
try_clear("data/my-queue").unwrap();

The returned value data is a kind of guard that implements Deref and DerefMut on the underlying type.

queue::RecvGuard and transactional behavior

One important thing to notice is that reads from the queue are transactional. The Receiver::recv returns a queue::RecvGuard that acts as a dead man switch. If dropped, it will revert the dequeue operation, unless queue::RecvGuard::commit is explicitly called. This ensures that the operation reverts on panics and early returns from errors (such as when using the ? notation). However, it is necessary to perform one more filesystem operation while rolling back. During drop, this is done on a "best effort" basis: if an error occurs, it is logged and ignored. This is done because errors cannot propagate outside a drop and panics in drops risk the program being aborted. If you have any cleanup behavior for an error from rolling back, you may call queue::RecvGuard::rollback which will return the underlying error.

Batches

You can use the yaque queue to send and receive batches of data , too. The guarantees are the same as with single reads and writes, except that you may save on OS overhead when you send items, since only one disk operation is made. See Sender::send_batch, Receiver::recv_batch and Receiver::recv_until for more information on receiver batches.

Tired of .awaiting? Timeouts are supported

If you need your application to not stall when nothing is being put on the queue, you can use Receiver::recv_timeout and Receiver::recv_batch_timeout to receive data, awaiting up to a completion of a provided future, such as a delay or a channel. Here is an example:

use yaque::channel;
use std::time::Duration;
use futures_timer::Delay;

futures::executor::block_on(async {
    let (mut sender, mut receiver) = channel("data/my-queue-2").unwrap();
    
    // receive some data up to a second
    let data = receiver
        .recv_timeout(Delay::new(Duration::from_secs(1)))
        .await
        .unwrap();

    // Nothing was sent, so no data...
    assert!(data.is_none());
    drop(data);
    
    // ... but if you do send something...
    sender.send(b"some data").await.unwrap();
 
    // ... now you receive something:
    let data = receiver
        .recv_timeout(Delay::new(Duration::from_secs(1)))
        .await
        .unwrap();

    assert_eq!(&*data.unwrap(), b"some data");  
});

Ctrl+C and other unexpected events

First of all, "Don't panic©"! Writing to the queue is an atomic operation. Therefore, unless there is something really wrong with your OS, you should be fine in terms of data corruption most of the time.

In case of a panic (the program's, not the programmer's), the queue is guaranteed to save all the most up-to-date metadata for the receiver. For the reader it is even simpler: there is nothing to be saved in the first place. The only exception to this guarantee is if the saving operation fails due to an IO error. Remember that the program is not allowed to panic during a panic. Therefore in this case, yaque will not attempt to recover from an error.

The same thing cannot be said from OS signals. Signals from the OS are not handled automatically by this library. It is understood that the application programmer knows best how to handle them. If you chose to close queue on Ctrl+C or other signals, you are in luck! Saving both sides of the queue is async-signal-safe so you may set up a bare signal hook directly using, for example, signal_hook(https://docs.rs/signal-hook/), if you are the sort of person that enjoys unsafe code. If not, there are a ton of completely safe alternatives out there. Choose the one that suits you the best.

Unfortunately, there are also times when you get aborted or killed. These signals cannot be handled by any library whatsoever. When this happens, not everything is lost yet. We provied a whole module, recovery, to aid you in automatic queue recovery. Please check the module for the specific function names. From an architectural perspective, we offer two different approaches to queue recovery, which may be suitable to different use cases:

  1. Recover with replay (the standard): we can reconstruct a lower bound of the actual state of the queue during the crash, which consists of the maximum of the following two positions:
    • the bottom of the smallest segment still present in the directory.
    • the position indicated in the metadata file.

Since this is a lower bound, some elements may be replayed. If your processing is idempotent, this will not be an issue and you lose no data whatsoever.

  1. Recover with loss: we can also reconstruct an upper bound for the actual state of the queue: the bottom of the second smallest segment in the queue. In this case, the smallest segment is simply erased and the receiver caries on as if nothing has happened. If replays are intollerable, but some data loss is, this might be the right alternative for you. You can limit data loss by constraining the segment size, configuring this option on SenderBuilder.

If you really want to err on the safer side, you may use Receiver::save to periodically back the receiver state up. Just choose you favorite timer implementation and set a simple periodical task up every hundreds of milliseconds. However, be warned that this is only a mitigation of consistency problems, not a solution.

Known issues and next steps

  • This is a brand new project. Although I have tested it and it will certainly not implode your computer, don't trust your life on it yet. This code is running in production for non-critical applications.
  • Wastes too much kernel time when the queue is small enough and the sender sends many frequent small messages non-atomically. You can mitigate that by writing in batches to the queue.
  • There are probably unknown bugs hidden in some corner case. If you find one, please fill an issue on GitHub. Pull requests and contributions are also greatly appreciated.

Dependencies

~2–13MB
~90K SLoC