We are checking the feasibility of migrating one of our application to Amazon Web Services (AWS) . We decide to use AWS API Gateway to expose the services and AWS Lambda (java) for back end data processing. The lambda function has to fetch a large amount of data from our database.
Currently using Cassandra for data storage, which has been set up with in an EC2 instance and it has no public ip.
Can anyone suggest a way to access Cassandra(EC2) from AWS Lambda using the private Ip ( 10.0.x.x)?
Is it a right choice to use AWS Lambda for large scale applications?
Since your Cassandra instance is using private IP, you will need to configure your AWS lambda Network to use a VPC. It could be the VPC you are running Cassandra in, or a VPC you create for the purpose of your lambdas, and that you VPC-peer to your cassandra VPC. A few things to note from the documentation :
when your lambda runs in a VPC, it doesn't have internet access by default, you will need to configure a NAT for that.
There is an additional latency due to the configuration of the ENI (you only pay that penalty on cold start)
You need to make sure your lambda has the right permission to manage the ENI, you should use this role: AWSLambdaVPCAccessExecutionRole
Your plan to use API / AWS lambda has at least 3 potential issues which you need to consider carefully:
Cost. API gateway per request cost is higher than AWS lambda per request cost. Make sure you are familiar with the cost.
cold start. When AWS start an underlying container to execute your lambda, you pay a cold start latency (which get worse when using VPC due to the management of the ENI). If you execute your lambda concurrently, there will be multiple underlying containers. Each of them will have this cold start the first time. AWS tends to keep the underlying containers ready for a warm start, for a few minutes (users report 5 to 40 minutes). You might try to keep your container warm by pinging your aws lambda, obviously if you have multiple container in parallel, it is getting tricky.
Cassandra session. You will probably want to avoid creating and destroying your Cassandra session each time you invoke your lambda (costly). I haven't tried yet, but there are reports of keeping the session alive in a warm container, you might want to check this SO answer.
Having say all that, currently the biggest limitation in using AWS lambda is concurrent execution and cold start latency. For data processing, that's usually fine. For user-facing usage, the percentage of slow cold start might affect your user experience.
Related
I am in process of coming up with a multi-region high availability (active-active) architecture for my product. A simplified version of our stack is that we use Lambda to implement our micro services, which are exposed as APIs using API Gateway. These micro services integrate with downstream services or databases like DynamoDB, Aurora RDS. So, '
Route 53 >> Api gateway >> Lambda >> Downstream service/Database
'
I am trying to figure out what is the best mechanism to configure Route 53 such that it understands any of the services in the stack fails so that it routes the incoming requests to another region. Eg if Lambda service in region-1 fails, then it is easy because I would create Health Check records pointing to these Lambdas and once they are not reachable Route 53 will itself route to next requests to region-2.
However, if the downstream resource eg RDS that Lambda is dependent on fails, how will Route 53 know this so as to route the next requests to region-2?
Appreciate any pointers on this.
It depends a bit on your envision failover setup.
Let us assume you have two regions: region1,region2
Now you could have two failure scenarios:
Lambda fails in region1 => you failover to Lambda in region2
RDS fails in region1 => you failover to RDS in region2
In both cases you need to ask yourself: What I want to do. If for example, in case 1 you connect from Lambda in region2 to RDS in region1 then high region transfer costs may occur, so you may want to trigger in any case a fail-over of RDS to region2.
Note: Generally it is very advisable to not connect directly with Lambda to RDS, but use instead RDS proxy (to avoid hammering the database with requests, slowing it down etc.): https://docs.aws.amazon.com/AmazonRDS/latest/UserGuide/rds-proxy.html
Generally, with RDS these region failovers are much more complicated (can answer on that bit if needed). It is also not simply changing IP to another region, because usually you need to promote in the other region the database (cluster) as a designated node to allow write operations.
For the databases (DynamoDB, Aurora) you mentioned there is though a solution: Use Global Tables.
A simpler solution could be - depending on your application - to use DynamoDB Global Tables (see https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/GlobalTables.html). However, clearly DynamoDB is not a relational database so it may not fit all cases. Nevertheless, DynamoDB works generally very good with Lambda and is also easier for cross-region replication. Note: if you encrypt your data using AWS KMS CMK (recommended) then you need to have this key also available in all regions where you plan to use Global tables (see https://docs.aws.amazon.com/kms/latest/developerguide/multi-region-keys-overview.html).
Another solution could be AWS RDS Aurora Global Tables (https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/aurora-global-database.html) - those are available in multiple region and failover is thus easier (cf. https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/aurora-global-database-disaster-recovery.html).
In the Aurora case you have to detect a region failure yourself (e.g. you could have a lambda in both regions that regularly tries to connect to the current active cluster for writing) and automatically promote the cluster in the new region as primary if it is not available in the original region.
Do not forget: You need to regularly test the failover otherwise it is almost ensured that it will not work when you need it.
Generally having databases cross-regions implies transfer costs and additional resource costs compared to a single region - not only during failover, but all the time data is written.
With this configuration, I recommend failing over the entire stack (to another Region), rather than failing over individual tiers (components) of the architecture. (This is what you seem to be saying in your question, but just making sure we are on the same page).
Your question comes down to how to configure the health check, and specifically how to implement shallow versus deep (checking dependencies like RDS) health checks.
There is an AWS Well-Architected lab that covers these concepts Implementing Health Checks and Managing Dependencies to improve Reliability.
This really wasn't clear for me in the Docs. And the console configuration is very confusing.
Will a Docker Cluster running in Fargate mode behind a Load Balancer shutdown and not charge me while it's not being used?
What about cold starts? Do I need to care about this in Fargate like in Lambda?
Is it less horizontal than Lambda? A lambda hooked to API Gateway will spawn a new function for every concurrent request, will Fargate do this too? Or will the load balancer decide it?
I've been running Flask/Django applications in Lambda for some time (Using Serverless/Zappa), are there any benefits in migrating them to Fargate?
It seems to be that it is more expensive than Lambda but if the Lambda limitations are not a problem then Lambda should always be the better choice right?
Will a Docker Cluster running in Fargate mode behind a Load Balancer shutdown and not charge me while it's not being used?
This will depend on how you configure your AutoScaling Group. If you allow it to scale down to 0 then yes.
What about cold starts? Do I need to care about this in Fargate like in Lambda?
Some good research has been done on this here: https://blog.cribl.io/2018/05/29/analyzing-aws-fargate/
But the takeaway is for smaller instances you shouldnt notice any more and ~40seconds time to get to a running state. For bigger ones this will take longer.
Is it less horizontal than Lambda? A lambda hooked to API Gateway will spawn a new function for every concurrent request, will Fargate do this too? Or will the load balancer decide it?
ECS will not create a new instance for every concurrent request,any scaling will be done off the AutoScaling group. The load balancer doesnt have any control over scaling, it will exclusively just balance load. However the metrics which it can give can be used to help determine if scaling is needed
I've been running Flask/Django applications in Lambda for some time (Using Serverless/Zappa), are there any benefits in migrating them to Fargate?
I havent used Flask or Django, but the main reason people tend to migrate over to serverless is to remove the need to maintain the scaling of servers, this inc managing instance types, cluster scheduling, optimizing cluster utilization
#abdullahkhawer i agree to his view on sticking to lambdas. Unless you require something to always be running and always being used 99% of the time lambdas will be cheaper than running a VM.
For a pricing example
1 t2.medium on demand EC2 instance = ~$36/month
2 Million invocations of a 256MB 3 second running lambda = $0.42/month
With AWS Fargate, you pay only for the amount of vCPU and memory resources that your containerized application requests from the time your container images are pulled until the AWS ECS Task (running in Fargate mode) terminates. A minimum charge of 1 minute applies. So, you pay until your Task (a group of containers) is running, more like AWS EC2 but on a per-minute basis and unlike AWS Lambda where you pay per request/invocation.
AWS Fargate doesn't spawn containers on every request as in AWS Lambda. AWS Fargate works by simply running containers on a fleet of AWS EC2 instances internally managed by AWS.
AWS Fargate now supports the ability to run tasks on a scheduled basis and in response to AWS CloudWatch Events. This makes it easier to launch and stop container services that you need to run only at a certain time to save money.
Keeping in mind your use case, if your applications are not making any problems in the production environment due to any AWS Lambda limitations then AWS Lambda is the better choice. If the AWS Lambda is being invoked too much (e.g., more than 1K concurrent invocations at every point of time) in the production environment, then go for AWS EKS or AWS Fargate as AWS Lambda might cost you more.
I am having long cold start times on our Lambda functions. We have tried "pinging" the Lambdas to keep them warm but that can get costly and seems like a poor way to keep performance up. We also have an EC2 instance running 24/7. I could theoretically "mirror" all of our Lambda functions to our EC2 instance to respond with the same data for our API call. Our Lambdas are on https://api.mysite.com and our EC2 is https://dev.mysite.com.
My question is could we "load balance?" traffic between the two. (Create a new subdomain to do the following) To have our dev subdomain (EC2) respond to all requests up until a certain "requests per minute" is hit. Then start to route traffic to our dev subdomain (Lambda) since we have enough traffic coming in to keep the Lambdas hot. Once traffic slows down, we move the traffic back over to our EC2.. Is this possible?
No, you can not do that with AWS Load balancer. What you can do is setup cloudwatch trigger to the lambda function which will map api.mysite.com to the lambda load balancer dns. Similarly, you can add trigger for low traffic which will redo dns entry.
If you are aware of rise in traffic, you can scheduled instances. Else you can also try AWS Fargate.
Hope this helps you.
Cloudwatch allows you to choose the number of triggers you wanna set for a lambda i.e. whenever an API is called, cloudwatch will trigger those many of the lambdas. This is one way you could achieve it. There's no way yet, to load balance between lambdas and instances.
You can configure your API to use Step functions which will orchestrate your lambda.
In AWS Lambda, there is no need of provisioning done by us. But I was wondering how AWS Lambda might be provisioning machines to run for the requests. Is it creating a EC2 server for each request and execute the request and then kill the server? Or it keeps some EC2 servers always on to serve the request by executing the lambda function? If its doing the former point, then I am wondering it would also affect performance of AWS Lambda to serve the request. Can anyone guide me on this?
Lambdas run inside of a docker-like container, on EC2 servers (using Firecracker) that are highly, highly optimized. AWS has thousands of servers running full time to serve all of the Lambda functions that are running.
A cold start Lambda (one that's never been run before) starts up in a few seconds, depending on how big it is. An EC2 server takes 30+ seconds to startup. If it had to startup an EC2 server, you'd never be able to use a Lambda through API gateway (because API Gateway has a 30 second timeout). But obviously you can.
If you want your Lambdas to startup super duper fast (100ms), use Provisioned Concurrency.
AWS Lambda is known to reuse the resources. It will not create an EC2 server for each request so that will not be a performance concern
But you should note that the disk space provided for your function sometimes not cleanup properly. As some users reported
You can read more on the execution life cycle of Lambda here: https://docs.aws.amazon.com/lambda/latest/dg/running-lambda-code.html
I have a web app which runs behind Amazon AWS Elastic Load Balancer with 3 instances attached. The app has a /refresh endpoint to reload reference data. It need to be run whenever new data is available, which happens several times a week.
What I have been doing is assigning public address to all instances, and do refresh independently (using ec2-url/refresh). I agree with Michael's answer on a different topic, EC2 instances behind ELB shouldn't allow direct public access. Now my problem is how can I make elb-url/refresh call reaching all instances behind the load balancer?
And it would be nice if I can collect HTTP responses from multiple instances. But I don't mind doing the refresh blindly for now.
one of the way I'd solve this problem is by
writing the data to an AWS s3 bucket
triggering a AWS Lambda function automatically from the s3 write
using AWS SDK to to identify the instances attached to the ELB from the Lambda function e.g. using boto3 from python or AWS Java SDK
call /refresh on individual instances from Lambda
ensuring when a new instance is created (due to autoscaling or deployment), it fetches the data from the s3 bucket during startup
ensuring that the private subnets the instances are in allows traffic from the subnets attached to the Lambda
ensuring that the security groups attached to the instances allow traffic from the security group attached to the Lambda
the key wins of this solution are
the process is fully automated from the instant the data is written to s3,
avoids data inconsistency due to autoscaling/deployment,
simple to maintain (you don't have to hardcode instance ip addresses anywhere),
you don't have to expose instances outside the VPC
highly available (AWS ensures the Lambda is invoked on s3 write, you don't worry about running a script in an instance and ensuring the instance is up and running)
hope this is useful.
While this may not be possible given the constraints of your application & circumstances, its worth noting that best practice application architecture for instances running behind an AWS ELB (particularly if they are part of an AutoScalingGroup) is ensure that the instances are not stateful.
The idea is to make it so that you can scale out by adding new instances, or scale-in by removing instances, without compromising data integrity or performance.
One option would be to change the application to store the results of the reference data reload into an off-instance data store, such as a cache or database (e.g. Elasticache or RDS), instead of in-memory.
If the application was able to do that, then you would only need to hit the refresh endpoint on a single server - it would reload the reference data, do whatever analysis and manipulation is required to store it efficiently in a fit-for-purpose way for the application, store it to the data store, and then all instances would have access to the refreshed data via the shared data store.
While there is a latency increase adding a round-trip to a data store, it is often well worth it for the consistency of the application - under your current model, if one server lags behind the others in refreshing the reference data, if the ELB is not using sticky sessions, requests via the ELB will return inconsistent data depending on which server they are allocated to.
You can't make these requests through the load balancer, So you will have to open up the security group of the instances to allow incoming traffic from source other than the ELB. That doesn't mean you need to open it to all direct traffic though. You could simply whitelist an IP address in the security group to allow requests from your specific computer.
If you don't want to add public IP addresses to these servers then you will need to run something like a curl command on an EC2 instance inside the VPC. In that case you would only need to open the security group to allow traffic from some server (or group of servers) that exist in the VPC.
I solved it differently, without opening up new traffic in security groups or resorting to external resources like S3. It's flexible in that it will dynamically notify instances added through ECS or ASG.
The ELB's Target Group offers a feature of periodic health check to ensure instances behind it are live. This is a URL that your server responds on. The endpoint can include a timestamp parameter of the most recent configuration. Every server in the TG will receive the health check ping within the configured Interval threshold. If the parameter to the ping changes it signals a refresh.
A URL may look like:
/is-alive?last-configuration=2019-08-27T23%3A50%3A23Z
Above I passed a UTC timestamp of 2019-08-27T23:50:23Z
A service receiving the request will check if the in-memory state is at least as recent as the timestamp parameter. If not, it will refresh its state and update the timestamp. The next health-check will result in a no-op since your state was refreshed.
Implementation notes
If refreshing the state can take more time than the interval window or the TG health timeout, you need to offload it to another thread to prevent concurrent updates or outright service disruption as the health-checks need to return promptly. Otherwise the node will be considered off-line.
If you are using traffic port for this purpose, make sure the URL is secured by making it impossible to guess. Anything publicly exposed can be subject to a DoS attack.
As you are using S3 you can automate your task by using the ObjectCreated notification for S3.
https://docs.aws.amazon.com/AmazonS3/latest/dev/NotificationHowTo.html
https://docs.aws.amazon.com/cli/latest/reference/s3api/put-bucket-notification.html
You can install AWS CLI and write a simple Bash script that will monitor that ObjectCreated notification. Start a Cron job that will look for the S3 notification for creation of new object.
Setup a condition in that script file to curl "http: //127.0.0.1/refresh" when the script file detects new object created in S3 it will curl the 127.0.0.1/refresh and done you don't have to do that manually each time.
I personally like the answer by #redoc, but wanted to give another alternative for anyone that is interested, which is a combination of his and the accepted answer. Using SEE object creation events, you can trigger a lambda, but instead of discovering the instances and calling them, which requires the lambda to be in the vpc, you could have the lambda use SSM (aka Systems Manager) to execute commands via a powershell or bash document on EC2 instances that are targeted via tags. The document would then call 127.0.0.1/reload like the accepted answer has. The benefit of this is that your lambda doesn't have to be in the vpc, and your EC2s don't need inbound rules to allow the traffic from lambda. The downside is that it requires the instances to have the SSM agent installed, which sounds like more work than it really is. There's AWS AMIs already optimized with SSM agent stuff, but installing it yourself in the user data is very simple. Another potential downside, depending on your use case, is that it uses an exponential ramp up for simultaneous executions, which means if you're targeting 20 instances, it runs one 1, then 2 at once, then 4 at once, then 8, until they are all done, or it reaches what you set for the max. This is because of the error recovery stuff it has built in. It doesn't want to destroy all your stuff if something is wrong, like slowly putting your weight on some ice.
You could make the call multiple times in rapid succession to call all the instances behind the Load Balancer. This would work because the AWS Load Balancers use round-robin without sticky sessions by default, meaning that each call handled by the Load Balancer is dispatched to the next EC2 Instance in the list of available instances. So if you're making rapid calls, you're likely to hit all the instances.
Another option is that if your EC2 instances are fairly stable, you can create a Target Group for each EC2 Instance, and then create a listener rule on your Load Balancer to target those single instance groups based on some criteria, such as a query argument, URL or header.