The Fastest Redis Ever

This sort of makes me sad. Redis has strayed from what its original goal/purpose was.

I’ve been using it since it was in beta. Simple, clear, fast.

The company I’m working for now keeps trying to add more and more functionality using Redis, that doesn’t belong in Redis, and then complains about Redis scaling issues.

This is what I see everywhere. Something is a success and then everybody starts using it wrong. Like Elastic search as database, people use it for searching and then start using it as primary database. Mostly pushed by management BTW not always the software engineer.

You'd be surprised how many engineers make these kinda decisions.

What do you think doesn't belong in Redis? I've always viewed Redis as basically "generic datastructures in a database" — as opposed to say, Memcached, which is a very simple in-memory-only key/value store (that has always been much faster than Redis). It's hard for me to point to specific features and say: that doesn't belong in Redis! Because Redis has generally felt (to me) like a grab bag of data structures + algorithms, that are meant to be fairly low-latency but not maximally so, where your dataset has to fit in RAM (but is regularly flushed to disk so you avoid cold start issues).

That does not match my experience. Engineers learn a new tool, that tool is successful in solving a problem. Whether it is recency bias, incorrect pattern matching, or simply laziness, the tool is used again but with reduced success. Repeat that process a few more times (sometimes in different organizations) and now the tool is way outside the domain, ill-fit to the task at hand, and a huge pain.

That often happens with engineers who pushed that tool getting promoted a few times and building their career on said tool, which is where I have seen this being pushed down, but I think it is important that in most cases are still engineers

> The company I’m working for now keeps trying to add more and more functionality using Redis, that doesn’t belong in Redis, and then complains about Redis scaling issues.

This doesn't sound like a Redis issue, you're just not using the right tool for the job.

Generic data structures in memory, grab bag of structures and algorithms... sounds more like a programming language or library than an external tool. C++ STL for example would fit these descriptions perfectly.

Doing everything is a recipe for bloat. In a database, in a distributed cache, in a programming language, in anything.

If your application can't survive the Redis server being wiped without issues, you're using Redis wrong.

Don't think the argument is "everything", just the things that can be done within the protocol. There's really not much bloat being added considering the "limitations": https://redis.io/docs/latest/develop/reference/protocol-spec

I think it wouldn't be unfair to compare it to Golang, which has in my opinion a quite unbloated stdlib which allows you do almost anything without external libraries!

Rarely seen Redis viewed as a database, even if that has been their push for the last few years.

I may be biased, but I think this announcement is actually a very good sign for Redis, since it shows that the focus is back to the community edition, that is, the source tree you can just download from GitHub (and I believe this is an effect of the license change: it is possible for the company to work on the public tree without competitors to cut&paste the code in SAAS services).

There are few things that are interesting for me about this discussion related to complexity and use cases outside the scope.

1. You can still download Redis and type "make" and it builds without dependencies whatsoever like in the past, and that's it.

2. Then you run it and use just the subset of Redis that you like. The additional features are not imposed to the user, nor they impact the rest of the user experience. This is, actually, a lot like how it used to be when I handled the project: I added Pub/Sub, Lua scripting, geo indexing, streams, all stuff that, at first, people felt like they were out of scope, and yet many shown to be among the best features. Now it is perfectly clear that Pub/Sub belonged to Redis very well, for instance.

3. This release has improvements to the foundations, in terms of latency, for example. This means that even if you just use your small subset, you can benefit from the continued developments. Sometimes systems take the bad path of becoming less efficient over time.

So I understand the sentiment but I still see Redis remaining quite lean, at least for the version 8 that I just downloaded and I am inspecting right now.

Why not just use Memcached, then? Memcached is much better as an ephemeral cache than Redis — Redis is single-threaded. The point of Redis is all of its extra features: if you're limiting yourself to Memcached-style usage, IMO you're using Redis wrong and should just use Memcached.

Yup, agree. Or as I like to call Redis, your "db building kit"

Of course if what you need is a traditional DB then go with a traditional DB

But it offers those data structures and other stuff that fewer competitors have (or has it in a more quirky way)

Valkey is not single threaded

Also the datatypes of redis are practical for caching more complex stuff; they are not for using it as a database though

This.

Sure, there's persistence but it always seemed like an afterthought. It's also unavailable in most hosted Redis services or very expensive when it's available.

There's also HA and clustering, which makes data loss less likely but that might not be good enough.

For the people wondering who would ever use Redis this way, check out Sidekiq! https://sidekiq.org/ "Ephemeral" jobs can be a big trade-off that many Rails teams aren't really aware of until it's too late. Reading the Sidekiq docs doesn't mention this, last time I checked, so I can't really blame people when they go for the "standard"/"best" job system and they are surprised when it gets super expensive to host it.

If your application is happy with an empty Redis, then why run Redis in the first place?

What you say is good in theory, but doesn’t hold in practice.

We use memcached instead of Redis. Cache different layers in different instances so one going down hurts but doesn’t kill. Or at least it didn’t when I was there. They’ve been trying to squeeze cluster sizes and I guarantee that’s no longer sufficient and multiple open circuit breakers happen if more than one cache goes tits up.

Cache and Sessions

Both running in-memory speed up an application, but you can survive both being nuked (minus potentially logging everyone out).

Redis supports multiple forms of replication for HA

I worked somewhere where a team, that can only be described as a clown team, decided to use Elastic as the “database” for the entire login/auth microservice, that other teams depended on.

It was so slow and terrible.

I always just think of Redis as a HashMap As A Service that only supports string keys.

It's nice if the stuff stays there, because my application will be faster. If it goes down I need a few seconds to re-populate it and we're back.

There are Redis-protocol compatible databases like Aerospike and Kvrocks that are useful if you want a KV store that isn't always in-memory.

Redis Enterprise has started to lean into being able to do this too.

The key is in the name: "Redis-tribution".

If you're not redistributing, then you're using it wrong. Only once redistribution has successfully occurred (i.e. you can reboot the redis process and recover), is the goal of redis fulfilled.

No. Cache protects your other services from peak traffic. Which often leads to wrong sizing of those services to reap efficiency gains. Autoscaling can’t necessarily keep up with that sort of problem.

Remember how I mentioned circuit breakers?

The only time we had trouble with memcached was when we set the max memory a little too high and it restarted due to lack of memory. Which of course likes to happen during high traffic.

Not fixing those would have resulted in a metastable situation.

> that has always been much faster than Redis

Do you have some reliable recent benchmarks comparing the two?

You should use Memcached if you're only using Redis as an ephemeral hashmap. It's much faster.

Totally agree. It's definitely not the right tool for what they're doing, but some of the engineers don't seem to know better, or understand, the point of being able to run scripts on Redis.

Lots of Lua scripting and calculations being done on Redis that has nothing to do with the data that's local to Redis. It's infuriating.

Pub/Sub is a huge use case for me