The problem here is, it depends on WHY you are rate limiting. If it's just to balance so you don't overload your backends, then absolutely this is a great way to do it.
If however you are trying to limit client(s) because the service is an authentication gateway for instance, then you want to limit user/pass requests to X number then concurrency limiting isn't a good way to do that.
So you may need both, depending on your use-cases, so it's not a one-size fits all solution.
I heard about a case where someone built a phone app backend without rate limiting. Someone found this hole and successfully attacked 70,000 accounts by running password dictionary cracks against the authentication API.
Modern distributed systems are simply too fast and users are too dumb with picking passwords to allow unlimited password attempts.
Well you can rate limit via a lot of things, definitely by IP is a good idea, but for ipv6, you typically want to limit by blocks(since every ipv6 user currently usually gets a full block of IP's), but you also probably want to also limit by username, i.e. if you keep trying to login as user 'root', you only get 3 attempts/minute or something.
After X attempts via the same IP, you can block/ban that IP via a FW rule/etc for say 5 mins. or 30m, or whatever.
It all depends on your security posture, you could go crazy and actually lock the account after 3 failures, but then you hand bad people a free DDOS... so you have to be careful about doing that.. But you could soft-lock and require a correct password and an email address verification after X failed attempts(i.e. correct login, plus they have to click a link in an email).
Anyways, see what I'm saying here? Authentication access is something you really want to get right, and it's a complicated topic.
But this is not something you deal with a rate-limiter. By your logic, then a sufficiently large botnet would dwarf any live traffic whatsoever, implying you’d probably want to filter them on network level instead.
> A better approach is to use a “set-then-get” mindset, relying on atomic operators that implement locks in a very performant fashion, allowing you to quickly increment and check counter values without letting the atomic operations get in the way.
In a highly distributed system you’d probably want to avoid a centralised data store altogether for fast moving data like rate limits. CRDTs and bucket weighting might be a more effective strategy.
The article states that tracking per-node could cause a problem with race conditions but that assumes it’s the counter that’s the problem. If the node cluster is aware of the other nodes and the relative load of the cluster, you can use this value to weight the isolated rate limiter and the only data that needs to be shared can be broadcast between the nodes using a pub/sub mechanism.
If some variance is permitted (+/- a margin either side of the limit) then having the nodes synchronise their “token weight” based on the size of the cluster means that the nodes can then manage the rate limit in-memory without ever needing to track in a data store.
It does trade-off accuracy, but for accuracy you can then revert to the set-then-get centralised counter, the trade-off being performance because of increased round trip time to the day store.
In most rate limit scenarios, at least from what we’ve seen, extreme accuracy isn’t usually that important vs. being able to scale amd rate limit without having to also scale a data layer to handle the counters.
Disregarding the article completely, I'll share my opinion on Kong because we use quite a bit at my workplace. We use an old version (0.9.x as of right now). The things I shared below might not be true anymore in new versions but I can't tell and are regarding the plugin system.
IMO, the idea of taking things like authentication, rate limiting, etc in a proxy is a wonderful idea (https://2tjosk2rxzc21medji3nfn1g-wpengine.netdna-ssl.com/wp-...). So in theory, the approach Kong takes is wonderful, but I think the implementation is not so much.
Kong is layer over nginx and uses lua as a scripting language to add all sorts of stuff. Quickly, you reach the limit of the plugins capability so you think, well, I will write my own plugins. Well, to me it was a very unpleasing experience. I found that the thing was hard to test and hard to maintain. Maybe my criticism is more lua than Kong but since Kong relies heavily on lua, there is not much I can do.
There is also magic happening. You declare a schema.lua files to configure the store to hold your data. Then automatically you've got a DAO interface available with a bunch of method on it to work with the store. You don't know what methods are available or what arguments should be passed into these functions.
Anyway, this is my take after spending quite a few hours working on home made plugins in lua for Kong.
In the end, I'm glad Kong is open source and it's a great piece of software. It helped us reduce our applications complexity but make sure you don't start to shift to much logic into it because the plugin system can be hard to work with.
Marco here, Kong's CTO. We hear you, we are aware that plugin development is not as easy as we wish, and we are planning to rollout a few improvements next year to make it mainstream, including:
- A built-in Lua plugin API that abstracts away the underlying complexity of NGINX and OpenResty. Among other things the plugin API will make it easy to interact with the lifecycle of request/response objects.
- An official testing framework for plugins (unit and integration tests).
- A new DAO to get rid of some magic and make the overall system more robust and extensible (with an in-memory DAO implementation for example).
- Support for remote gRPC plugins to support plugin development in any gRPC supported languages.
- And finally supporting plugin installation/versioning within Kong itself using the HTTP Admin API to install/uninstall plugins cluster-wide on any Kong node (as opposed to installing plugins on the file system).
NGINX and Lua (on top of LuaJIT http://luajit.org/) were chosen for their reliability and performance, the next step for Kong is to focus more on making the overall Kong ecosystem bigger therefore simpler plugin development is a very high priority item in our roadmap.
Not OP, but we are using Tyk (tyk.io) at my workplace. It's definitely more user friendly than Kong, but we've had other issues with it. Having not used Kong extensively but knowing what it is, I'll go out on a limb and say Kong is likely more performant.
Actually it should be the opposite. Definitely do your own benchmark, but take a look at the latest numbers from Tyk and compare it with the latest numbers from Kong to get a general idea.
This analysis was done by BBVA comparing Tyk and Kong performance https://www.bbva.com/en/api-gateways-kong-vs-tyk/ - disclosure: I work for Kong, but Kong (the company) had no involvement with the benchmarking in that article.
Those BBVA benchmarks look very very wrong. BBVA achieved 3822 requests per second with a 16Gb server running Tyk.
I used a 16Gb Digital Ocean box, and a separate box to run benchmarks from.
Tyk CE with keyless api.
docker run --rm williamyeh/wrk -t4 -c100 -d1m --latency --timeout 2s http://10.131.45.89:8080/tykbench/get
Running 1m test @ http://10.131.45.89:8080/tykbench/get
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 5.79ms 6.97ms 206.76ms 87.28%
Req/Sec 6.35k 0.93k 10.89k 74.33%
Latency Distribution
50% 3.19ms
75% 6.28ms
90% 15.93ms
99% 31.14ms
1516573 requests in 1.00m, 185.13MB read
Requests/sec: 25244.99
Transfer/sec: 3.08MB
As per above results, I got 25,244 requests per second. 90% of requests were served with sub 15ms latency.
Another thing - you can freely scale with Tyk CE as much as you want at no-cost. I currently run 3 gateways in prod and it doesn't cost a penny. The BBVA article states that you need to pay Tyk to scale - simply not true:
>However, if we want to deploy a significantly larger number of microservices, or simply ensure high-tolerance to failure, we need to scale up the number of gateways. In the case of Kong, all it takes is adding a new node and connecting it to the database. With Tyk we can do the same, although it will require us to pay for the service. This could be a determining factor when choosing what API Gateway to use.
I went through process picking API gateway about 3 months ago. Tyk is not without it's warts, but has far more out-of-box features baked in (inc distributed rate limiting) than Kong. And if custom plugins are required, I can choose between lua, python, js or anything that supports gRPC. With Kong - stuck with Lua.... no thank you.
I guess my point here is that if Tyk not performant enough for you, just add another gateway. And if I need custom functionality - I don't need to learn Lua.
Hey Bob - you'll be pleased to hear Tyk took that feedback on-board. There is now very comprehensive documentation and user-guides. And true to the Open Source approach, anyone can contribute to them, the community have been great in working to improve them.
As you know, there are many plugins to do authentication in Kong. We started with jwt, then a coworker decided that we needed a more flexible approach so he basically forked the jwt plugin to add stuff for our needs. It quickly became confusing and hard to maintain.
When we tried to introduce a new feature (tokens similar to what Github offers with personal token, that is a revocable token with a given set of permissions), we had a rough time.
In the end, the decision to fork the plugin was maybe not the good one and the decision to bring the token feature into this plugin were maybe not the right one. But still, working in the plugin code was unpleasant in my opinion.
By the way, keep up the good work, it's a solid piece of software!
This may be how Kong does it but it's not really 'high performance'. The right way to do rate limiting is to limit by IP using a counting Bloom Filter or Cuckoo filter along with random samples. When you hit a false positive then you have a second normal rate limiter to 'mop up' IPs that are over the first limiter.
This doesn't give you a hard exact limit but gets the job done storing far less state. You also need to bucket by IP sub-ranges in IPV6 to stop people crap flooding you with tons of unique IP's
I have been using Kong since 2016 on various APIs and I have been impressed by the continuous increase of performance after every release. I wish upgrades were easier but I know they are working on it (the "migrations" between each major version), but overall it's a fast and pluggable layer for any API inside or outside the firewall, especially if you are running containers.
On this note for those of you who haven't noticed it yet, they have released an Alpine version of their Docker image, but it's still not the default one. I would actually recommend using it to further reduce the size of Kong containers: https://konghq.com/blog/kong-alpine-docker/
> A better approach is to use a “set-then-get” mindset, relying on atomic operators that implement locks in a very performant fashion, allowing you to quickly increment and check counter values without letting the atomic operations get in the way.
Can you elaborate on this? Why is it more performant? And what are the trade-offs vs. get-then-set?
The performance degradation of doing strict rate limit relying on Kong Cluster data store (Postgres or Cassandra) is amplified by the lake of database connection pool management. Each incoming http request induce the creation of a new database connection, this process is very expensive especially with Cassandra when authentication is enabled.
I use Kong EE for a client, I have read Kong EE documentation carefully and made a lot of tests.
The current implementation (0.29) have this behavior (that not meet our needs).
1) All incoming requests are take into account, including those which have been rejected (with 429 error). If the consumer exceed his limit during many consecutive time windows, all the requests of the consecutive time windows will be rejected (with 429 error).
2) For a windows size of 1 second, the computed weight of the previous windows is always 100%. If the limit was reach during the previous second, all requests made in the current windows will be rejected (with 429 errors).
One thing that RateLim.it uses to its advantage is calculating the “nextPossiblePass” for each limit. This allows clients to cache forever that something is over the limit until time X and not have to make another request.
In the bursty case this let’s clients effectively short circuit and protect the system.
The basic premise is not do do req/s limiting but rather concurrency limiting which results in req/s limiting by itself.
Concurrency limiting is rather simple and doesn't require a lot of code complexity.