Skip to content

Latest commit

 

History

History
107 lines (81 loc) · 8.38 KB

README.md

File metadata and controls

107 lines (81 loc) · 8.38 KB

reactive-mutiny crate

reactive-mutiny GitHub Actions reactive-mutiny on crates.io reactive-mutiny on docs.rs

async & zero-cost-abstraction Event-Driven Reactive Library for Rust with advanced & optimized containers and Stream executors

Browse the Docs.

Rust's reactive-mutiny was designed to allow building efficient & elegant asynchronous event processing pipelines (using Streams -- a.k.a. "async Iterators"), easing flexible & decoupled microservice architectures (distributed or not), ready for production.

The core of this library is composed of a Uni and a Multi -- hence the name "Mutiny". Both process streams of events:

  • Uni allows a single listener OR multiple consumers for each produced payload -- also definable as allows a single event processing pipeline;
  • Multi allows multiple listeners AND multiple consumers for each produced payload, allowing several event processing pipelines -- or, in Kafka parlance, allowing several consumer groups
  • Multi may do what Uni does, but the later does it faster -- hence, justifying its existence: Uni doesn't use any reference counting for the payloads and uses a single queue/channel (where Multi requires as many as there are listeners).

Moreover, zero-cost-abstractions over metrics, logs & retrying are available -- getting optimized away if not used, as specified in const time initialization options and on functional, deeper API opt-ins.

Taste the library in this excerpt:

    use reactive_mutiny::prelude::*;

    fn logic_1(events_stream: impl Stream<Item=InputEventType>) -> impl Stream<Item=OutputEventType> {
        // your logic goes here using Rust's Stream / Iterator functions
    }

    fn main() {
        // build the event processing pipeline
        let events_handle = UniZeroCopy::<InputEventType, 1024, 1>::new()
        .spawn_non_futures_non_fallible_executor("Consumer of InputEventType and issiuer of OutputEventType",
                                                 |events_stream| {
                                                     logic_2(logic_1(events_stream))
                                                         .inspect(|outgoing_event| send(outgoing_event))
                                                 },
                                                 |_executor| async { /* on-close logic */ });

    }

    // see more details in examples/uni-microservice

Core components:

  1. A set of channels through which events are sent from producers to consumers -- all context-switch-free (AKA "lock-free") -- including zero-copy & mmap log based implementations;
  2. A set of generic Stream executors for all possible combinations of Future/non-Future & Fallible/non-Fallible event types, with the option of enforcing or not a Timeout on each event's resolution of their Future. The API was carefully designed to allow the compiler to fully optimize everything: most of times, all of the reactive code ends up in the executors and the whole Multi / Uni abstractions are zeroed out;
  3. Instrumentation & Metrics collectors for visibility of the performance and operation (as said earlier, as a zero-cost-abstraction);
  4. The main Multi and Uni objects, along with a set of prelude type aliases binding the channels and allocators together.
  5. Constant-pool based allocators, for superior performance and flexibility -- see the AtomicZeroCopy channel benchmarks;

NOTE: This crate is rather new (less than 1yo), but actively maintained and used in production: no known bugs exist (and MIRI says we're fine), speed is amazing, API has been throughoutly tested & reviewed and is stable, but improved docs & code cleanup / refactorings will still be (slowly) addressed to improve the cosmetics. Anyway, evolutions are always driven from community feedback

MIRI: Not all parts of this crate are testable with MIRI, as of 2023-06-14: "ready events from epoll_wait is not yet implemented"; "mmap syscalls" and some other functionalities are not available in MIRI -- but what is able to be tested, passes.

Performance

This crate was very loosely inspired by the SmallRye's Mutiny library for Java, from which some names were borrowed. Little had to be done to bring the same functionality to Rust, due to the native functional approach, powerful error handling, async support and wonderful & flexible Iterator/Stream APIs supported by the language, so the focus of this work went into bringing the events to their maximum processing speed & operability: special queues, topics, stacks, channels and Stream executors have been created, offering a superior performance over the Rust's native & community versions -- inspect the benches folder for details:

reactive-mutiny's channels latencies reactive-mutiny's channels throughput performance characteristics of the standard/community vs our provided raw senders of payloads from one thread to another

reactive-mutiny's allocators & type wrappers performance characteristics comparison of standard vs our provided type wrappers and allocators, used for zero-copy channels -- with raw memcopy and allocators baselines

Where to go next

Docs will still be improved. Meanwhile, the following sequence is suggested for new users of this crate:

  1. Look at the examples/;
  2. Study the type aliases in reactive-mutiny::prelude::advanced::* -- at this point, it is safe to trust that the docs will provide everything you'll need;
  3. For an advanced usage example, inspect the reactive-messaging crate -- in special, how easily & decoupled the reactive abstractions allow upgrading a processor that doesn't pass any answers back to one that does pass them back to the peers.

Comparisons

If you're familiar with SmallRye's Mutiny, here are some key differences:

  • Both our Uni and Multi here process streams of events. On the original library, a Uni is like a single "async future" and, since we don't need that in Rust, the names were repurposed: the other Multi is our Uni (may also work as our Multi when using "subscriptions") and the other Uni you may get by just using any Rust's async calls & handling any Result<>, for error treatment;
  • Each event fed into the pipeline will be executed, regardless if there is an answer at the end; also, there is no "subscription" (subscription semantics is achieved by adding pipelines to a Multi);
  • Executors & their settings are set when the pair producer/pipeline comes to be (when the Uni / Multi object is created): there is no .merge() nor .executeAt() to call in the pipeline;
  • No Multi/Uni pipeline conversion and the corresponding plethora of functions -- they are simply not needed;
  • No timeouts are set in the pipeline -- they are a matter for the executor, which will simply cancel events (that are Futures) that take longer than the configured executor's maximum (SmallRye's Uni timeouts are attainable using Tokio's "futures" timeouts, just like one would do for any async function call);
  • Incredibly faster: Rust's compiler makes your pipelines (and most of this library) behave as a zero-cost abstraction (when compiled in Release mode). The used-and-abused Const Generics play a great role in such optimizations -- at the expense of our rather complex type definitions in reactive_mutiny::prelude::advanced.
  • To fully get the original Mutiny's behavior, you'll have to use:
    • Rust's reactive-mutiny (for reactive async event-processing);
    • Tokio (to get responses from Futures and to specify timeouts in async calls, async sleeps... saving a ton of APIs for this crate);
    • Streams (here we don't mix Multi & Stream & Iterator functionalities -- which, in practice, leads to inefficient abuses of the original Java library's abstractions -- for using a new instance of their Multi where a Stream or Iterator could be used is a common bad parctice / anti-pattern);