▲ 229 points
back
112 comments
> Rate limiting is one of the most common Redis use cases. Traditionally, users implemented rate limiters using server-side Lua scripts combined with client logic. In Redis 8.8, we introduce a window counter rate limiter (by @raffertyyu, together with the Redis team).
I had a look for this and it turns out it's slightly mis-described there - it's not a window counter, it's a "GCRA (Generic Cell Rate Algorithm)" - a leaky bucket algorithm. Code here: https://github.com/redis/redis/blob/unstable/src/gcra.c
The code comments say it was heavily influenced by https://github.com/brandur/redis-cell by Brandur Leach.
It's a neat algorithm (I just learned about it today) - it only needs to store a single integer for each rate-limited key, which is the "Theoretical Arrival Time" when the bucket would next be empty.
Also, the “cell” in Generic Cell Rate Algorithm is an ATM cell. GCRA is 1990s telecom, the scheduling algorithm ATM switches used to check that 53-byte cells were arriving on the wire at the agreed rate.
Asynchronous Transfer Mode (I had to look it up too.)
I've never liked token buckets (vs sliding window counters) due to the extra work CPU cycles required to "fill" them. It seems like doing an atomic incr on a key based on a `time % 1 minute` or something would be more efficient and then let that key TTL expire X duration later. This results in zero work for rate limiters not in use and only a single push change -> resulting count for ones in use. Nothing but setting the TTL extra is required.
Thanks for the links, I'll checkout the Generic Cell Rate Algorithm!
While I love Redis as a versatile tool for external data structures, it's still lacking in two areas IMHO:
One, it would be cool to be able to embed it, similar to sqlite, directly into applications.
Two, the HA story is so much more complicated than it should be. I totally acknowledge that concurrency and distributed computing is hard, but it should not require reading heaps of documentation and understanding two entirely separate multi-node approaches only to figure out there are lots of subtle strings attached that make it impractical for many applications.
What would be the point of embedding Redis into an application? What's the advantage of using Redis over using the builtin (or third party) data structures of the language the application is developed in?
I'm asking as a non-webdev who never quite got what Redis actually does, but would love to learn.
To me the thing I like about Redis is that it gives you a storage engine very suitable for caches; it handles TTLs and memory pressure, as well as built-in serialization with the ability to get better performance by allowing for some data loss. At the same time, many users will be deploying small programs to individual machines. If you could just have Redis be embedded this would make it very operationally simple: no additional daemons and a single file to backup if you want to.
It would also be useful because of the ability to switch modalities. When running a multi node service, you can use Redis to share data between nodes and use Redis pubsub as a communication bus. If you wanted to support a simple single node configuration too, then it wouldn't need to be a special case, it could just go through the same mechanism but with an embedded Redis instance.
It's pretty similar to SQLite: being able to embed more or less a complete storage engine into your app can be very convenient and powerful.
Probably because Redis gives you a very well-defined/understood set of rich data structures with built-in behavior like TTL, atomic operations, eviction, and persistence. These things are otherwise usually scattered across native types, helper classes, or entirely separate libraries.
A few nice things about doing this in no particular order:
Embedding would make local dev/CI integration testing convenient.
Embedding replicated Redis with each application instance would give you HA benefits while infra-management complexity.
Embedded redis (even via local RPC) is still going to be faster than a lot of languages or frameworks’ built-in data structures. Large array operations in, say, Python are gonna slower than RPCing to Redis (assuming that the data structures are built gradually and not built all at once); to beat Redis you’d have to use numpy or something—-which is definitely preferable, but is extra work if your app already uses Redis for other things.
Just like choosing SQLite over e.g. LMDB or RocksDB, embedded Redis would be a nice future proofing option for small apps during the prototype phase; less would have to be changed to move Redis out of the app than if a different cache or persistence service were chosen.
I have 15 processes handling API requests over HTTP. I want rate limiting, so users don't DoS my API. I need shared state across these processes to store information about rate limits for each client key. I store those in redis.
Or queues, caching, pub/sub... Redis can do a lot, it's really (and i mean really) fast, very easy to get up and running and the protocol is pretty simple.
Locality and latency.
Network hops are not free! Those milliseconds are an eternity compared to local function calls.
The optimal architecture is something like what Service Fabric or Orleans can do with their distributed dictionary types: reads are generally in-process and take only nanoseconds (but writes require a synchronous replica copy to a remote host.)
Obviously this requires load balancers to steer traffic consistently, but that’s a common feature… outside of the public clouds where they forgot latency exists.
Why would you embed SQLite?
It’s the same use case with a different api.
A typical (meaningful) example might be communication between threads or actors in a single process, or idempotent tests.
As with SQLite, an external xxx that does this for you is certainly better, etc. but it’s convenient sometimes, to have an application that doesn’t go “now before you run this install Postgres…”.
It’s seldom useful for a web app where you control everything.
In practice, mostly scaling sessions and ephemeral data (caching) across multiple intances of a microservice on multiple machines. Seperating the kv store and the application allows upgrading each application while retaining availability and avoiding loss of session data.
I mostly use redis for pub/sub communication between services. If the app wasn't a collection of knative functions, and instead a monolith, it would be cool to also use redis for event based communication.
For simple cases, it is probably a total overkill to even consider it, but for something heavier, embedding the database gives you a chance to trivially migrate later to a separate database server.
> One, it would be cool to be able to embed it, similar to sqlite, directly into applications.
I've found myself wanting this on several occasions too. I.e. wanting all my rust backend processes (k8s pods) to have some minimal shared state, without having to spin up a Redis cluster. I've talked to Claude about it a couple of times, and it descends into something like, "you gotta use Raft or CRDTs, and pick 2 out of 3 from CAP". Which honestly seems pretty fair, and indicates to me that I'm dreaming for something magical.
Nonetheless, it is nice to hear someone else asking for this. If this is indeed feasible (even if simple/limited), then I'd be interested to try it.
I don't know if that'll make you feel any better but yeah, you're indeed asking for the impossible! You need consensus between your nodes that store state _somehow_, either these nodes are Redis and it does that for you, or these nodes are your pods and you need to do consensus yourself (zookeeper might help, but you're definitely in "complicated stuff" territory).
Spinning up an in-memory (no persistence) Redis cluster in your k8s should be easy enough, hopefully?
Genuinely interested why we need HA in redis, just not read round robin from multiple non-HA instances?
Redis (and memcache) are memory caches and should be treated like that, not like highly consistent distributed session store.
> Redis (and memcache) are memory caches and should be treated like that
If you haven't come across Kvrocks yet, it may be worth a look: https://github.com/apache/kvrocks https://kvrocks.apache.org/ . It's a database with a Redis-compatible wire protocol, but the database is stored on disk. This means your working set is not limited by RAM and can be a few orders of magnitude larger! On modern SSDs this is still very fast. I think it improves the durability story as well. But the big win is the orders of magnitude larger database space.
As I've been improving my side project https://totalrealreturns.com/ recently I've ended up using both Redis and Kvrocks together. Redis is great for small global state that needs to be super fast. Kvrocks is great for larger bulk data storage (large precomputed datasets), but also supports a lot of the Redis data structures as well as Lua scripts.
Redis is used for plenty of things, not just memory caches.
For example if you use it for session storage, you can't have your application read from a random instance that may or may not contain the session.
For the project I've been working on for more than 15 years, we make extensive use of the pub/sub functionality for distributing live data. Pub/sub scales well across the cluster. Publish to one, and it goes out to subscribers on any of the nodes that they've connected to.
Will millions of users, high availability is critical for this functionality.
Redis doesn't necessarily have to be used as a cache. Streams, for example, make it a great message queue; but a single-node message queue is a single point of failure and thus not viable for many setups.
Redka can be used embed, in Go: https://github.com/nalgeon/redka
> it's still lacking in two areas
This is entirely different than what Redis is and tries to solve.
Sqlite is embedded. It's not a distributed SQL. Redis is a distributed data structure store and concurrency primitive. These are worlds apart.
> HA story is so much more complicated than it should be
It is precisely as complicated as it needs to be. You don't want data loss.
If you're in the business of high available fault tolerance, you read the manual and learn how to Redis.
> Two, the HA story is so much more complicated than it should be.
Really? I am curious, how would simplify it? It’s a very well defined problem and all the “solutions” are very complicated and with many strings attached. I have managed one or two systems that came in different modalities and you had to pick your poison and had to make sure the other engineering teams understand the trade offs. Some were more successful than others, but “easy” never crossed my mind.
Redis is single threaded and doesn’t concern itself with these things, directly, exactly because Antitez understood the trade offs and made all the right choices.
How would you improve the HA story without sacrificing ease of use and performance on a single thread?
Where did everyone end up on the Redis/Valkey split? Is there still a reason to use Redis after the license kerfuffle?
We switched to Valkey two years ago. I haven't really looked back. I think both projects have done a lot of nice stuff since the split but it's not really impacting anything I use. The feature set was fine five years ago and I don't think we're using anything in Valkey that wouldn't work in Redis. There are probably a lot of projects that never switched over because they had no real need.
But most of the cloud providers now offer Valkey because of the license changes. Of course, cloud providers not offering Redis was the intention of the license change from the Redis point of view. So mission accomplished for Redis.
But the flip side of course is that if you want to deploy on standard infrastructure rather than self hosting Redis, Valkey is now the easy, low risk path that probably should be the default for most companies that target AWS, Azure, GCP, etc. Same with Elasticsearch vs. Opensearch and a few other products where the community forked because of license changes.
Mentioning Elasticsearch because I know people in both communities and I'm deeply familiar with the stack. A few years on, Opensearch has taken a lot of the momentum from Elasticsearch.
For those who may not know, you can cut your costs in AWS by going with Valkey over Redis for about 33% savings.
https://aws.amazon.com/blogs/database/reduce-your-amazon-ela...
I've switched to Valkey and I'm not really looking back. I'm much more comfortable with those people maintaining the software.
Valkey, because our cloud provider is hosting it and that's obviously what they prefer.
I feel like we're using about 1% of its features at this point - really just as a fast K/V store - so it would be easy to switch if needed, but I can't see a case where we would.
We use almost exclusively Valkey now, mostly because we host on AWS and Render, which both use Valkey. It's faster, cheaper and compatible. I'd consider Garnet too but I believe it doesn't support LUA(or didn't at the time we needed it).
We switched to Valkey after the Redis license kerfuffle happened, discovered we were saving money on our AWS bill, and have no motivation to go back to Redis.
So we’ve stayed with Valkey.
Valkey continues to be the real redis, released under a permissive open source license like redis originally was. The fork of redis known as redis because those that decided to fork it had the trademark at first wasn't under an OSI-certified open source license and now is under a non-permissive OSI-certified open source license.
I'm taking liberties with the concept of forking.
We're a self hosted shop, we went with Valkey. Valkey also has support for RDMA, which we already is running in our infrastructure.
Went with 100% ValKey, if you are solely on AWS it is a no-brainer
Most people seems to have switched to Valkey, and it's backed by the Linux foundation.
We went with DragonFlyDB
There's also a separate blog post that goes into the details of why existing data structures Redis already supported, which could provide array-like behavior, weren't good enough:
https://redis.io/blog/diving-deep-into-rediss-new-array-data...
@antirez wrote about the development of that data structure last month, which includes how he used LLMs to do it (which was before ds4 for the co-comment mentioning it ;). The PR he linked goes into the motivation.
Are we still using Redis? License change, no more Kube operators.
window counter rate limiter!
This is awesome!
And arrays look great too. Lots to play with.
given his ds4 project, likely collaborated with DeepSeek for this release:
Possibly, but the array type code was implemented using GPT/Claude models before DS4 was a thing. I really recommend this write up on how he used LLMs which I think is a more sane/safe way to code with them vs the YOLOing even I'm subject to unfortunately...
The experimental SSD streaming feature (author's demo @ https://x.com/antirez/status/2062536214675067322 - recently merged into the main branch) is great news for that project, allowing for SOTA inference (DeepSeek V4 Flash and Pro!) on RAM-limited machines. Now we need work on large-ish scale batching in order to recover tok/s under the SSD streaming scenario. It's not helpful when running normally (at least not on Apple Silicon) since thermal/power throttling is the constraint in that case, but SSD streaming is a whole other consideration.
And here we see the reason for the sudden AI enthusiasm of Redis authors: array data structures are used in AI. This was clear weeks ago.
The website looks like openclaw's website.