Yes, please. I've never understood why someone wants to do pub/sub and only base the reliability on TCP/WebSockets. The concept of "fire and hope everyone who wants the message still has a connection open" always seemed fragile to me.
This has been the reason I've recommended implementing long-polling instead of WebSockets for real-time applications. And every time I see a real-time solution which only uses WebSockets I try to steer away from it. Once you have a reliable data model (which includes log position, retrieving old messages etc.) it's just as simple to implement long-polling as WebSockets. With WebSockets-only solution I can't help but think they base all the message delivery reliability on TCP.
This proposal looks like a clean (very Redis-like) solution, and I immediately see use cases for it.
Did you look into the EventSource API? It's basically automatic long-polling with a simple file format built into the browser. It supports pretty much every "real-time" pattern: regular polling, long polling, and streaming.
The browser handles the log position for you (via the id field and the Last-Event-ID header) and automatically reconnects when the server closes the stream or the connection is lost.
The proposed Redis API seems to fit extremely well to this model. My previous usage of EventSource with Redis worked by sending the entire state whenever someone (re-)connects and using PubSub afterwards. This works well for me, but likely doesn't scale very well.
Take a look at SSE (Server-Sent Events, aka EventSource). Unlike WebSockets, the it has the concept of an "event ID", which allows the protocol to automatically continue from the last event, no extra roundtrip needed to send a setup message to the server. Implemented by every major browser (except, of course, IE, where you can polyfill it pretty easily).
This is exactly what we built for Poll Everywhere with http://firehose.io/ to push updates in realtime reliably over crappy wifi connections. Since we're used by presenters at huge conferences with dicey networks we had to spend a lot of time thinking through all sorts of low quality connection scenarios.
It's exciting to see a proposal for an all-Redis implementation.
That's cool. In case you haven't seen this blog post [1], it provides an extended discussion of the uses of this abstraction.
One comment about the groups API is that while it is very convenient, it seems a bit fragile - if a consumer is nuked immediately after it gets a fresh batch from redis then this batch is lost forever.
Have you paid for Redis? If so, was it direct donation, Redis Labs services, or committing code?
Edit: Not asking to be a jerk or insist that s/he donate. I use lots of things heavily that I've never contributed to. I'm just curious how the Redis community works.
Agreed. I'm currently considering Kafka for a project but am hesitant because ideally we need a zero maintenance solution, and Kafka seems to always end up requiring some handholding.
I'm not familiar with the Redis dev process at all. How long might one expect it to take for a feature of this complexity to make it from RCP to production-ready?
This looks very interesting. I use Redis as a backend for an actor-system [1] as part of a larger functional framework for C# [2]. I use RPUSH to add messages to an actor's queue along with PUBLISH for saying 'new message' to the actor that should check its inbox, this is then followed by LINDEX 0 to peek at the item at the head of the queue, before calling LPOP when the message has been successfully processed. If I'm understanding this correctly, that process could be wrapped up in a single stream?
Yes. Instead of multiple operations, you would get this from a single operation. But instead of modifying the log on every "pop", the operation would merely update a "read position" on behalf of the client. The position semantics would allow you to jump backwards in the log if needed.
This has been the reason I've recommended implementing long-polling instead of WebSockets for real-time applications. And every time I see a real-time solution which only uses WebSockets I try to steer away from it. Once you have a reliable data model (which includes log position, retrieving old messages etc.) it's just as simple to implement long-polling as WebSockets. With WebSockets-only solution I can't help but think they base all the message delivery reliability on TCP.
This proposal looks like a clean (very Redis-like) solution, and I immediately see use cases for it.