I am working on a mobile app that will broadcast a push message to hundreds of thousands of devices at a time. When each user opens their app from the push message, the app will hit our API for data. The API resource will be identical for each user of this push.
Now let's assume that all 500,000 users open their app at the same time. API Gateway will get 500,000 identical calls.
Because all 500,000 nearly concurrent requests are asking for the same data, I want to cache it. But keep in mind that it takes about 2 seconds to compute the requested value.
What I want to happen
I want API Gateway to see that the data is not in the cache, let the first call through to my backend service while the other requests are held in queue, populate the cache from the first call, and then respond to the other 499,999 requests using the cached data.
What is (seems to be) happening
API Gateway, seeing that there is no cached value, is sending every one of the 500,000 requests to the backend service! So I will be recomputing the value with some complex db query way more times than resources will allow. This happens because the last call comes into API Gateway before the first call has populated the cache.
Is there any way I can get this behavior?
I know that based on my example that perhaps I could prime the cache by invoking the API call myself just before broadcasting the bulk push job, but the actual use-case is slightly more complicated than my simplified example. But rest assured, solving this simplified use-case will solve what I am trying to do.
If you anticipate that kind of burst concurrency, priming the cache yourself is certainly the best option. Have you also considered adding throttling to the stage/method to protect your backend from a large surge in traffic? Clients could be instructed to retry on throttles and they would eventually get a response.
I'll bring your feedback and proposed solution to the team and put it on our backlog.
Related
I have a client which has a pretty popular ticket selling service, to the point that the microservice based backend is struggling to keep up, I need to come up with a solution to optimize and loadbalance the system. The infrastructure works through a series of interconnected microservices.
When a user enter the sales channels (mobile or web app), the request is directed to an AWS API Gateway which is in charge of orchestrating the communication towards the microservice in charge of obtaining the requested resources.
These resources are provided from a third party API
This third party has physical servers in each venue in charge of synchronizing the information between the POS systems and the digital sales channels.
We have a REDIS instance in charge of caching these requests that we make to the third party API, we cache each endpoint with a TTL relative to the frequency of updating the information.
Here is some background info:
We get traffic mostly from 2 major countries
On a normal day, about 100 thousands users will use the service, with an 70%/30% traffic relation in between the two countries
On important days, each country has different opening hours (Country A starts sales at 10 am UTC, but country B starts at 5 pm UTC), on these days the traffic increases some n times
We have a main MiddleWare through which all requests made by clients are processed.
We have a REDIS cache database that stores GETs with different TTLs for each endpoint.
We have a MiddleWare that decides to make the request to the cache or to the third party's API, as the case may be.
And these are the complaints I have gotten that need to be deal with:
When a country receives a high amount of requests, the country with the least traffic gets negatively affected, the clients do not respond, or respond partially because the computation layer's limit was exceeded and so the users have a bad experience
Every time the above happens, the computation layer must be manually increased from the infrastructure.
Each request has different response times, stadiums respond in +/- 40 seconds and movie theaters in 3 seconds. These requests enter a queue and are answered in order of arrival.
The error handling is not clear. The errors are mixed up and you can't tell from which country the errors are coming from and how many errors there are
The responses from the third party API are not cached correctly in the cache layer since errors are stored for the time of the TTL
I was thinking of a couple of thinks that I could suggest:
Adding in instrumentation of the requests by using AWS X-Ray
Adding in a separate table for errors in the redis cache layer (old data has to be better than no data for the end user)
Adding in AWS elastic load balancing for the main middleware
But I'm not sure how realistic would be to implement these 3 things, I'm also not sure if they would even solve the problem, I personally don't really have experience with optimizing this type of backed. I would appreciate any suggestions, recommendations, links, documentation, etc. I'm really desperate for a solution to this problem
few thoughts:
When a country receives a high amount of requests, the country with the least traffic gets negatively affected, the clients do not respond, or respond partially because the computation layer's limit was exceeded and so the users have a bad experience
A common approach in aws is to regionalize stack - assuming you are using cdk/cloud formation creating regionalized stack should be a straightforward task.
But it is a question if this will solve the problem. Your system suffers from availability issues, regionalization will isolate this problem down to regions. So we should be able to do better (see below)
Every time the above happens, the computation layer must be manually increased from the infrastructure.
AWS has an option to automatically scale up and down based on traffic patterns. This is a neat feature, given you set limits to make sure you are not overcharged.
Each request has different response times, stadiums respond in +/- 40 seconds and movie theaters in 3 seconds. These requests enter a queue and are answered in order of arrival.
It seems that the large variance is because you have to contact the servers at venues. I recommend to decouple that activity. Basically calls to venues should be done async; there are several ways you could do that - queues and customer push/pull are the approaches (please, comment if more details are needed. but this is quite standard problem - lots of data in the internet)
The error handling is not clear. The errors are mixed up and you can't tell from which country the errors are coming from and how many errors there are
That's a code fix, when you do send data to cloudwatch (do you?). You could put country as a context to all request, via filter or something. And when error is logged that context is logged as well. You probably need venue id even more than country, as you can conclude country from venue id.
The responses from the third party API are not cached correctly in the cache layer since errors are stored for the time of the TTL
Don't store errors + add a circuit breaker pattern.
We have a webpage that queries an item from an API gateway which in turn calls a service that calls another service and so on.
Webpage --> API Gateway --> service#1 --> service#2 --> data store (RDMS, S3, Azure blob)
We want to make the operation resilient so we added a retry mechanism at every layer.
Webpage --retry--> API Gateway --retry--> service#1 --retry--> service#2 --retry--> data store.
This however could case a cascading failure because if the data store doesn't response on time, it will cause every layer to timeout and retry. In other words, if each layer has the same connection timeout and is configured to retry 3 times, then there will be a total of 81 retries to the data store (which is called a retry storm).
One way to fix this is to increase the timeout at each layer in order to give the layer below time to retry.
Webpage --5m timeout--> API Gateway --2m timeout--> service#1
This however is unacceptable because the timeout at the webpage will be too long.
How should I address this problem?
Should there only be one layer that retries? Which layer? And how can the layer know if the error is transient?
A couple possible solutions (and you can/should use both) would be to retry on different conditions and implement rate limiters/circuit breakers.
Retry On is a technique where you don't retry on every condition, but only specific conditions. This could be a specific error code or a specific header value. E.g. in your current situation, DO NOT retry on timeouts; only retry on server failures. In addition, you could have each layer retry on different conditions
Rate limiting would be to stick either a local or global rate limiter service inline to the connections. This would just help to short-circuit the thundering herd in the case that it starts up. E.g. rate limit the data layer to X req/s (insert real values here) and the gateway to Y req/s and then even if a service attempts lots of retries it won't pass too far down the chain. Similarly to this is circuit breaking, where each layer only permits X active connections to any downstream, so just another way to slow those retry storms.
I am wondering if there is a standard way to reject requests with the same body sent within a few seconds at the API gateway itself.
Forex: Reddit rejects if I try to post the same content within few seconds in a different group. Similarly, if I make a credit card payment for the second time, it automatically rejects it.
I am wondering if there is a way to have the same behavior in the AWS API gateway itself so that we are not handling it in lambda functions with dynamoDB and stuff.
Looking forward to efficient ways of doing it.
The API Gateway currently doesn't offer a feature like that, you'd have to implement this yourself.
If I was to implement this, I'd probably use an in-memory cache like ElastiCache for Redis or Memcached as the storage backend for deduplications.
For each incoming request I'd determine what makes it unique and create a hash from that.
Then I check if that hash value is in the cache already. If that's the case it 's a duplicate and I reject the request. If it isn't already in the cache, I'd add it with a time to live of n seconds (The time interval in which I wish to deduplicate).
I have hosted my node app in Cloud Run and all of my requests served within 300 - 600ms time. But one endpoint that gets data from a 3rd party service so that request takes 1.2s - 2.5s to complete the request.
My doubts regarding this are
Is 1.2s - 2.5s requests suitable for cloud run? Or is there any rule that the requests should be completed within xx ms?
Also see the screenshot, I got a message along with the request in logs "The request caused a new container instance to be started and may thus take longer and use more CPU than a typical request"
What caused a new container instance to be started?
Is there any alternative or work around to handle long requests?
Any advice / suggestions would be greatly appreciated.
Thanks in advance.
I don't think that will be an issue unless you're worried about the cost of the CPU/memory time, which honestly should only matter if you're getting 10k+ requests/day. So, probably doesn't matter and cloud run can handle that just fine (my own app does requests longer than that with no problem)
It's possible that your service was "scaled to zero" meaning that there were no containers left running to serve requests. In that case, it would be necessary to start up a new instance and wait for whatever initializing/startup costs are associated with that process. It's also possible that it was auto-scaled due to all other instances being at their request limits. Make sure that your setting for max concurrent requests per instance is set greater than one - Node/Express can handle multiple requests at once. Plus, you'll only get charged for the total time spend, not per request:
In situations where you get very long (30 seconds, minutes+) operations, it may be a good idea to switch to some different data transfer method. You could use polling, where the client makes a request every 5 seconds and checks if the response is ready. You could also switch to some kind of push-based system like WebSockets, but Cloud Run doesn't have support for that.
TL;DR longer requests (~10-30 seconds) should be fine unless you're worried about the cost of the increased compute time they may occur at scale.
AWS Lambda functions are supposed to respond quickly to events. I would like to create a function that fires off a quick request to a slow API, and then terminates without waiting for a response. Later, when a response comes back, I would like a different Lambda function to handle the response. I know this sounds kind of crazy, when you think about what AWS would have to do to hang on to an open connection from one Lambda function and then send the response to another, but this seems to be very much in the spirit of how Lambda was designed to be used.
Ideas:
Send messages to an SQS queue that represent a request to be made. Have some kind of message/HTTP proxy type service on an EC2 / EB cluster listen to the queue and actually make the HTTP requests. It would put response objects on another queue, tagged to identify the associated request, if necessary. This feels like a lot of complexity for something that would be trivial for a traditional service.
Just live with it. Lambda functions are allowed to run for 60 seconds, and these API calls that I make don't generally take longer than 10 seconds. Not sure how costly it would to have LFs spend 95% of their running time waiting on a response, but "waiting" isn't what LFs are for.
Don't use Lambda for anything that interacts with 3rd party APIs that aren't lightning fast :( That is what most of my projects do these days, though.
It depends how many calls will this lambda execute monthly, and how many memory are you allocating for those lambda. The new timeout for lambda is 5 minutes, which should (hopefully :p) be more than enough for an API to respond. I think you should let lambda deal with all of it to not over complicate the workflow. Lambda pricing is generally really cheap.
E.g: a lambda executed 1 million times with 128 MB allocated during 10 seconds would cost approximatively 20$ - this without considering the potential free tier.