Rate Limiting system design | TOKEN BUCKET, Leaky Bucket, Sliding Logs

Rate limiting protects your APIs from overuse by limiting how often each user can call the API.
In this video following algorithms are discussed
Token Bucket
Leaky Bucket
Sliding Logs
Sliding window counters
Race Conditions in distributed systems
Donate/Patreon: / techdummies

[@khalidelgazzar]

04:16 Token bucket 10:40 Leaky bucket 12:50 Fixed window counter 16:15 Sliding logs 20:36 Sliding Window counter 25:21 Distributed system setup (Sticky sessions | locks)

[@CrusaderGeneral]

My implementation takes advantage of Redis expiration.. When a call comes in, I create a record and the increment the value. Consequent calls will increment the value until the quota is reached. If the quota is not reached by the time the record expires, consequential request will cause a creation of new record and restart the counter.. This way I dont need to check and compare dates at any point. Code is very simple. Albeit, I am not maintaining a perpetual quota, I am only preventing abuse, which is really the main gist of request throttling

[@varshard0]

This is the way I implemented for my org also. Simple and served its purpose well.

[@shelendrasharma9680]

how would you manage the concurrancy here in redis.

[@bazzalseed]

@@shelendrasharma9680 redis is single thread.

[@dhruveshkhandelwal8104]

this is indirectly fixed window counter

[@sid007ashish]

This is fixed window counter only.

[@praveenakarapu]

Narendra, very informative video, keep it up. About locking in case of distributed token bucket you can use following technique Optimistic locking or conditional put - many no sql databases support conditional put. This is how it works * Read current value, say 9 * You do a conditional put with value 10 only if current value is 9. * When 2 concurrent requests try to update the value to 10, only one of them will succeed and other will fail as current value for that request will be 10.

[@vcfirefox]

i was reading alex xu, i did not get good idea about sliding window and sliding window counter. now after i watched your explanation it is crystal clear and with pros and cons. thank you for doing this!!

[@logeshkumar8333]

This channel is just hidden Gem!

[@nikhilchopra9247]

Good Stuff Naren! Even famous profs are not able to explain this kind of stuff so clearly.

[@TechDummiesNarendraL]

Thanks

[@prabudasv]

Narendra, your video are great resources for learning system design. Your explanation of concepts is crystal clear. Big thumbs up for you

[@abasikhan100]

Great explanation. The pattern you followed is very good i.e. when you mention a problem with some approach, you also provide the solution for that instead of just identifying problems.

[@rabbanishahid]

Best explanation, almost searched everywhere for my scenario, but found this tutorial very very helpful, once again thanks man.

[@terigopula]

you have my respect Narendra.. great work! :)

[@RandomShowerThoughts]

I think you're easily the best youtuber for system design content

[@princenarayana]

Sliding window can be optimized by setting the size of the queue to Max Requests allowed and try to remove the old entries only if max size is reached by comparing timestamp

[@sbylk99]

Great tutorial. Tricky part comes at 25:12:)

[@valeriiryzhuk4126]

One additional case, were sliding logs should be used: limit a bitrate of video/audio/internet signal. In such case you need to store a packet size with a timestamp

[@JoshKemmerer]

I love your voice brother. It makes it exciting to listen to what you have to say about this very interesting design topic.

[@Awaarige]

Bro, You saved my months. Love from Pakistan

[@SanjayKumar-di5db]

You can solve this with the help of increment or decrement method on redis which works atomically on any key so there is no chance for data inconsistencies and no need to put any lock 😊

[@himanshu111284]

2 services firing increment concurrently will still face the same problem, so i think it will not work without locking. Read + Write has to be an atomic transaction.

[@SanjayKumar-di5db]

@@himanshu111284 in redis increment and decrement methods on id are atomic so no need for lock

[@rajsekharmahapatro]

@@SanjayKumar-di5db First time i am learning something new by going through KZbin comments bro. Thanks for it man.

[@xuanwang7400]

"compare and set" kind of logic works perfectly without explicit locking in simple operation case. But in complex situation, the app server may need a few requests. e.g. read the data first, the do some processing, then write back. and then two servers can do the same thing with same data at same time, thus race condition.

[@mohammadfarseenmanekhan4820]

very underrated youtube channel for system design

[@screen189]

Hi Narendra - You are doing a good job in your knowledge transfer. I suggest you cover these topics as well - a) Job Scheduler b) Internals of Zoo Keeper c) Dist.Sys concepts like 2PC, 3PC, Paxos d) DB Internals.

[@TechDummiesNarendraL]

Added to TODO, Thanks

[@screen189]

Thanks for your response. Looking forward for her videos!!@@TechDummiesNarendraL

[@1qwertyuiop1000]

I love your cap.. Looks like a trademark for you.. Thanks for all your videos..

[@r3jk8]

This video was a clear and concise explanation of these topics! Great job! You have a new subscriber.

[@ishansoni8494]

Great work Narendra..! I am currently planning to switch jobs and your videos on system design are amazing...!!

[@lolnikal6851]

20:36 Sliding Window counter The rate limit is 10R/M While in explanation , he considered 10R/S so please don't get confuse and think he is wrong

[@VirgiliuIonescu]

For the last example with concurrency. How about optimistic locking on the counter. Number of req has a version. If you try to update from 2 different RL, one of them will have the NoReq version smaller than the current one and will fail. The RL can retry or drop

[@ajaypuri1837]

Narendra L! You doing good job! I watched your couple of videos. Keep it up!

[@PankajKumar-mv8pd]

One of best explanation, thanks man :)

[@ShivamSingh-jw8ey]

04:15 Rate Limting Algorithms 25:11 Race Conditions in distributed systems

[@rajeshd7389]

Narendra L !! This is just superb ... keep going.

[@vinodcs80]

very comprehensive video. Great work. subscribed

[@shrimpo6416]

Perfect! I wish I can give you 1,000,000 likes!

[@molugueshwar1]

Hi Narendra, In token bucket scenario above, I would like to add one point that in order to reset the requests count after one minute to 5 again, we have to store the time(start time) of the first request so that we can check the difference of one minute to reset the count

[@nikhilneela]

Yes, I agree. If you simply reset the tokens to 5 when the minute changes, it would allow more than 5 requests/minute. Storing the start time and always comparing it with the current request time and if the delta is equal to or more than a minute, only then we can reset the tokens. @Eshwar, is this what you meant ?

[@molugueshwar1]

@@nikhilneela yes Nikhil. That's right

[@poojachauhan1509]

Great work, Searching for System design like leetcode or Hackerank...

[@dragonmohammad]

Distributed Systems, a necessary evil.. very nicely explained Narendra !!

[@adityagoel123able]

Awesome Narendra..

[@rationalthinker3223]

Outstanding Explanation

[@rbsrafa]

Great video, congrats!!

[@akhashramamurthy8774]

Thank you Narendra. The incredible content archive that you are building is invaluable. Thank you.

[@anand2009ish]

Excellent..hats off

[@Ghost_1823]

Your content is good. But please try to change your voice modulation. It really helps for long videos.

[@aeb242]

Great lesson! Thank you!

[@saurabhchako89]

Great video. Well explained.

[@amitchaudhary6199]

Great work Narendra👍👍

[@resetengineering]

Why are you using two caches? Your sync issues are solved by keeping one single cache. Then, coming to race conditions, redis automatically acquires a lock on the transaction since it is atomic and therefore, the other request(second) should get an updated value. For SPOF on one cache, we can keep a master slave nodes for redis

[@mszjuliak]

What's the difference between token bucket and fixed window? they seem so similar

[@mumbaibusa]

The key and value stores are different for the two. In the case of the fixed counter, the key is defined by the Userid+minute whereas for token bucket the key is userid. For value the FC is just number of reqs, for token you track the time and the number of requests so the checking algorithm has more to do.

[@preety202]

Burst problem at boundary seem to exist in token bucket as well right?

[@romangavrilovich8453]

@@preety202 yes

[@grantl3032]

seems they are about the same to be functionally, maybe a bit diff implement wise?

[@paraschawla3757]

token bucket is the number of tokens in a bucket, there is refill() happening in bucket after nth min/sec. Number of tokens represent number of request that can be served. with every new request, it keeps going down...but tokens keep increasing based on ratelimit as well. Fixed window counter is having User+TimeStamp as key and count as value for particular window and then start again.

[@thejaswiniuttarkar620]

the threshold is calculated per second, for example AWS API gateway 5000 req/sec .. we can just declare an Array Queue or Array stack and start pushing elements in to it and keep flushing it every second ... + or - 10/20 request would not matter .. if the stack/Queue fills up it would throw an error and that error could be propagated to the user !!

[@saip7137]

You have a new subscriber. Thanks for making this video.

[@keatmin]

Thanks for the great tutorial, but I have a question as how would a rate limit service obtain lock of a record in separate db affect another rate limiter service obtain the count from different db within a node？

[@IC-kf4mz]

Token Bucket and Fixed Window counter, what's the difference?

[@uditagrawal6603]

Yes this explanation for token bucket doesn't seem correct as in token bucket tokens are added at a particular rate in a particular window time , also there might be chances of going over rate limit in certain scenarios.

[@PABJEEGamer]

With token bucket algorithm we have control over cost of each operation(we can associate how many tokens an operation costs), where as in fixed window we dont, since we increase the counter by 1 each time

[@abcdef-fo1tf]

@@uditagrawal6603 why can't we have a set and compare operation on the counter, or just a restriction that it can't go over a certain amount, and have requests try to increment number by 1 and reject them if it can't?

[@prakashkaruppusamy3817]

Good one bro !

[@prasukjain8488]

Why he is looking like varun singla sir from Gate smashers , btw nice lecture

[@mostaza1464]

Great video! Thank you!

[@themynamesb]

Great video.. Thanks for the knowledge.

[@rangak7502]

Awesome work sir.. 👍🏼

[@vishalkohli3953]

What a guy!! bless you bro

[@nishathussain3672]

I love your videos. Thank you for making such detailed videos which explain the concepts so clearly. :)

[@dev-skills]

Redis provides INCR and DECR commands which are atomic operations for increment and decrement of its Integer Data Type. Will this not take care of distributed access without any lock ?

[@victoryang7734]

I think his assumption is redis is seperate

[@Priyam_Gupta]

Yes this will be taking care as they are atomic.

[@abcdef-fo1tf]

@@victoryang7734 what does separate redis mean. Is distributed redis not a shared cache?

[@bhaskargurram94]

Thanks for the nice explanation. One question - What is the difference between fixed window counter and token bucket? Are they not doing the same?

[@paraschawla3757]

token bucket is the number of tokens in a bucket, there is refill() happening in bucket after nth min/sec. Number of tokens represent number of request that can be served. with every new request, it keeps going down...but tokens keep increasing based on ratelimit as well. Fixed window counter is having User+TimeStamp as key and count as value for particular window and then start again. Essence of both alogos are very different.

[@curiousbhartiya8410]

@@paraschawla3757 But the underlying problem of both algorithms is the same is what the original comment meant. That they both might end up serving twice the amount of the desired RPM.

[@PABJEEGamer]

With token bucket algorithm we have control over cost of each operation(we can associate how many tokens an operation costs), where as in fixed window we dont, since we increase the counter by 1 each time

[@cantwaittowatch]

Well explained Narendra

[@indrajitbanerjee5131]

This is not efficient and optimized, cause it has to do linear O(N) time processing for each requests. The way it actually solves rate limiting is: 1. Create a container/ list of max N size, when you have to serve N requests/ min let say. 2. When a request comes: 2.1. If the container size is lesser than N, then add the timestamp. 2.2. If no, then do a binary search on the list with (TS - 1 min), this will return the index of the timestamp which got served at the beginning of the last minute. Get the index diff from that position and that is the number of requests you already served. 2.2.1. If that is more than or equal to N -> wait in the message queue with a signal or wait time. 2.2.2. If no, then add the TS entry at the list. 3. Keep a sanity check on each list size, that it should always contain the timestamps of last N requests. Keep on deleting the old requests. This way the response time reduces to O(logN) and also the latency is resolved.

[@anonym705]

Excellent videos, just lacking good sound system.

[@divyeshgaur]

thank you for sharing the video. neatly explained.

[@hemangakrishnaborah4987]

I found this video very useful. One thing that can be improved is the way it is presented. At times the material seems unorganized. For example, there are flashes on the screen because the speaker forgot to mention it verbally. Adding a few notes before making the video may help the presenter have a good flow.

[@153deep]

Consider this scenario for token bucket: We can only serve 5 request/5 min. One request (10.05), Two request(10.06), Two request(10.07) we have served all the 5 requests so at 10.07 we will have 0. Now when we get new request at 10.11 it should be the valid request because request at 10.05 & 10.06 should be removed but as per token bucket it won't be served because 10.07 is set to 0 & will be reset at 10.12

[@vaidyanathanpk9221]

Not really. Read about the token bucket algorithm. Before serving the operation at 10.12, it'll try to figure out the time elapsed so far ( 10.12 - 10:07 ) Then it'll figure out the number of tokens to add for this time elapsed ( For 5 minutes, we need to add 5 tokens ) So before doing the serving calculation, these addition of tokens will be done and then when you do the calculation, you should be able to serve these requests. The key point is maintaining something called as lastUpdateTime in the bucket.

[@karthikrangaraju9421]

The inconsistency problem is basically a common DB problem called "lost update" due to two threads reading committed data concurrently and performing writes without any locks. Solution is to introduce locking to enforce ordering. Or enforce ordering by sticky session at a much higher level

[@dbenbasa]

For token bucket - why do we need to update the timestamp (and not only the counter) when we are within the same minute, e.g. from 11:01:10 to 11:01:15? Why not just upata the timestamp when refilling the bucket, i.e. when we switched to a different minute, e.g.: from 11:01:10 to 11:02:07?

[@alivesurvive471]

You only set the timestamp on the first connection within the period or if you using something like memcached you can set the instance with a ttl value.

[@manasranjan4]

Good bro. Awesome

[@RandomShowerThoughts]

31:00 and you can't even lock across the nodes. If you are sharding then maybe, but as soon as you introduce replication, I don't think it'll just work like that

[@michael4799]

For the situation of distributed race limit, even though one user send two requests at the same time in one server, it dosen't mean that the actual two processing threads will deal them serially, so the inconsistency problem seems still exist. I think to address this problem we can make the read and update operation as atomic with redis+Lua.

[@prajwal9610]

Redis does this by having a lock which is already suggested in the video

[@rekhakalasare4910]

​@@prajwal9610 yea but in case of local memory suppose single user two request going to 2 regions and regions local cache first read from db and then update in cache and db. Then also there is inconsistency as both req operating parellely

[@081sidd]

at 10:37 video time, you mentioned that race condition may occur because of multiple requests coming from the different or same server. As you said, we are using Redis for this solution. Redis commands are atomic in itself and while executing atomic commands there is no scope of any data races. Did I get something wrong here?

[@gxbambu]

same question here!

[@musheerahmed5815]

Two request from the same user coming at the same time. Both get the same data one after the other. Both increment the count one after the other. The count ends up incremented only once.

[@mukeshbansiwal]

@@musheerahmed5815 Use optimistic locking by adding version column to avoid lost update

[@vivek9876]

Because here two operation are required. 1) Get the current counter value 2) And If its less than threshold then increment the counter. For example current counter value is 9 and threshold is 10 and if two request comes at the same time and both request see current value as 9 and so both request allowed but in real case one of the request must fail. You either has to take Lock implementation on Redis or have to write atomic operation using WATCH/MULTI or write LUA script for your usecase.

[@faizanfareed9076]

Using redis lock or lua scripts increases latency to user request.

[@InfinteMotivation]

you are the best

[@andriidanylov9453]

Thank You

[@gulati9]

At 19:18 How can we serve 11 requests , when the limit is set to 10?

[@nitinkulkarni7942]

Exactly. I dont think it will happen

[@arun5741]

As usual naren rocks !!!

[@andresantos-yx3bh]

amazing video

[@JitendraSarswat]

There is one con to all your videos. If you skip 10 sec of this video, you are doomed :-P Exceptional work, Narendra.

[@ashutoshbang5836]

Great video, keep up the good work :)

[@itsNaveen9]

You have already served 8 instead of 5 at 28:34 , your intention is right, but Cache 1 = U1:3 and Cache 2 = U1:2, should be the case, instead of u1:4 in both.

[@santoshdl]

thanks Narendra

[@madhusogam5823]

very nice tutorial .. great work :)

[@rahulsharma5030]

@31:00 you have confused me here, if we use locks, region 1 will have lock in region 1 redis only. Still regions 2 call can read old data from region 2 redis and allow more requests. R1 should take lock of all regions DB theoretically if u say locking is one way to solve consistency?

[@anuraggupta6890]

Narendra from where do you get such a great understanding of system

[@springtest540]

Sir please make video on elevator design and google doc design as well.

[@sanjanind]

I also want for these two.

[@TechDummiesNarendraL]

Sure I will work on it.

[@biboswanroy6699]

Amazing content

[@karthikeyaacharya5700]

why not use cache expiry to set rate limit? If the rate limit is set at 10 rpm, For a user, maintain a key in redis, set the cache expiry to 1 minute. Fetch the user key from redis for every API request, If the key is present, check if the count has exceeded. If yes, block the current request. If the count is under the rate limit, update the count for user. The cache will expire after a minute. Is there any problem with this approach?

[@vigneshbaskaran631]

Imo, redis key value pair is the only viable solution(first one). Others are examples for over engineering if we implement.

[@amanshivhare5592]

So Ideally, Token Bucket can have more request in particular time. Like if 5 request were made on 11:55:00 and the very next minute 11:56:00 5 more request are made, so total 10 request can be made in a minute. (or size of a bucket)? Right?

[@IC-kf4mz]

Yes. If it's implemented as explained you are right.

[@DebasisUntouchable]

thanks for this video

[@akshaytelang4532]

can't we use Zookeeper for synchronization to manage requests along multiple regions

[@singhalvikash]

Nice explanation. Could you please make a video for Google ad sense analytics collection system ?

[@helishah6719]

For the Local Memory solution that you provided, how is it different from the solution that you explained just before (where the rate limiter is connected directly to the Redis)?

[@ravisoni9610]

great explanation (y)

[@utkarshchugh8686]

In sliding window logs, how are we able to serve 11 (requests) in last minute, if we're checking the rate in real time. Ideally it shouldn't allow for more than 10.

[@javacoder1986]

Thanks for great video, very informative, however last several minutes of video is not very clear and crisp like other part of the video.

[@cbest3678]

isnt the token bucket and fixed window has the same problem of boundary request problem... ? since even in token bucket you can request more token in end of the first request window and request more token to the second of the window.?

[@shreysom2060]

Can't we use a sorted Redis set to avoid the concurrency issue?

[@krishankantsharma3655]

Sir, for amazon any particular series of questions you want to suggest.

[@DenisG631]

A good one. Thanks

[@dataguy7013]

@Naren, even with local memory, you can have inconsistency. It just is a bit faster. Do I have that right?

[@Priyam_Gupta]

yes it won't work. if we are even talking about updating it all the time its better to rely on redis cluster to do the copy then our application server.

[@sethuramanramaiah1132]

Don't the fixed window counter also run into concurrency issue like the first scenario ?

[@ersinerdem7285]

so, we are building a web application for example. Where do we put this rate limiter? As an aspect in java, as a cross cutting concern? Or as a server like load balancer in front of the application server?

[@vishnusannidi7830]

Generally, rate limiters, authorization are offloaded to Load balancer/API GW.