Crates.io | syndicus |
lib.rs | syndicus |
version | 0.3.1 |
source | src |
created_at | 2024-09-14 03:37:14.155618 |
updated_at | 2024-10-14 02:40:04.524014 |
description | Publish/Subscribe with types for topics and subscriber (re)synchronisation |
homepage | |
repository | https://github.com/arnolddevos/syndicus |
max_upload_size | |
id | 1374437 |
size | 49,517 |
A rusty interpretation of publish/subscribe using types for topics and compaction to control backlog.
A Syndicate
exchanges messages between publishers and subscribers on a many to many basis.
It is an in-process, async data structure built on tokio watch
.
Types are used for topics. An application defines a unified type for communication,
typically an enum
. Call this type A
.
B
requires B: Into<A>
.B
requires A: TryInto<B>
.The derive-more crate can neatly produce these conversions. Here is a toy example:
// Individual message types
#[derive(Debug, Clone)]
struct Temperature(i64);
#[derive(Debug, Clone)]
struct Voltage(i64);
// Unified message type
#[derive(Debug, Clone, From, TryInto)]
enum Message {
T(Temperature),
V(Voltage),
}
// The syndicate.
let syndicate: Syndicate<Message> = Default::default();
Then publishing and subscribing tasks for Temperature
might look like this:
// a task that publishes temperatures
async fn temp_sensor(p: Publisher<Message, Temperature>) -> Result<()> {
// see examples/basics/main.rs
loop {
// ....
p.push(Temperature(t)).await;
}
}
// a task that monitors temperature
async fn temp_monitor(mut s: Subscription<Message, Temperature>) -> Result<()> {
// see examples/basics/main.rs
while let Some(Temperature(t)) = s.pull().await {
// ...
}
}
// run the tasks
spawn(temp_sensor(syndicate.publish()));
spawn(temp_monitor(syndicate.subscribe()));
A Syndicate
has no backlog limit meaning publishers are never blocked and
subscribers never get lagging errors. With certain assumptions, Syndicate
will also operate in bounded space. The price of this is compaction.
The Syndicate
will drop certain older messages.
The last linear_min
messages are always retained. Any older message may be
dropped if it has the same compaction_key
as a younger message.
The order of publication of messages is preserved in any case.
The assumption is that a subscriber only needs to see the latest message with each key to converge on a valid state.
For this example, compaction will ensure that at least the most recent message
of each type is retained. The compaction key is just the enum discriminant
:
impl Compactable for Message {
type Key = Discriminant<Self>;
fn compaction_key(&self) -> Self::Key {
discriminant(self)
}
}
The ability to extract a compaction key is expressed by a trait, Compactable
.
This effectively imposes a key-value structure on the data.
The space complexity of a Syndicate
is O(n) where n is the number of distinct
compaction keys among the published messages. This is comparable to the
space requirement of a key-value store.
Function scope
implements a form of structured concurrency intended to
work with Syndicate
. This is built on tokio JoinSet
. Tasks spawned within a scope
will be joined at the close of the scope.
Tasks are assumed to return Result<(), Error>
where Error
is an error type used
throughout the application. This keeps errors and messages separate.
For example:
// Run tasks until all finished or any one errors
scope(|local| {
local.spawn(temp_sensor(syndicate.publish()));
local.spawn(volt_sensor(syndicate.publish()));
local.spawn(temp_monitor(syndicate.subscribe()));
Ok(())
})
.await