Did you create a multi-tier Elastic Beanstalk deployment? Did you tie it to CodePipeline to deploy out of Github? Has it been working well until just recently?
…did you accidentally leave RDS attached to your worker tier?
This post is for you.
I built an Elastic Beanstalk for a customer with those characteristics. It's been working great for about a year, until suddenly… the developer of the application reports that he's no longer able to deploy his code changes. It keeps failing and rolling back all of the changes to the last known good state, which includes older versions of his code. This was bad news for everyone because we had a Monday-morning deadline to demo code changes to a new customer.
Sunday morning offered me a chance to sit and focus on this. I've been trying to understand this problem for a few days. It looks like I was finally able to understand the issue after some focus and coffee.
First, let's cover what was actually happening. When the developer pushed his code updates through CodePipeline, Elastic Beanstalk was working through its “magic” (cough) to update the config to its “known good state” (which was wrong) and failed to apply the changes because of CloudFormation problems. This triggered a rollback on CloudFormation, CodePipeline, and Elastic Beanstalk config changes. Hence the failure.
How did it all get out of whack?
There were several mistakes committed, most of them on my part. Some of them are just problems with Elastic Beanstalk itself. But I'll make the no-no list:
Don't let Elastic Beanstalk manage your RDS instance. Remove all references to RDS in all tiers before you build your RDS instance. Even AWS tells you to not to do this. I missed the one in the worker tier.
If you proceed forward with RDS tied to your EB, do NOT use the RDS console to make any changes to the RDS instance. EB won't know about the changes and will get really angry when they don't match. In our case, we did some performance testing and modified the RDS instance size from db.t2.micro to db.m4.large. We also changed the storage setting from 20gb to 100gb. We made those changes in the RDS console and not the EB console. Don't do that.
You should change one setting in the RDS console. Turn off automatic version upgrade. In our case, RDS was upgrading the minor version of the database and once again, EB got angry. Worse yet, you can't change the minor version in EB's console. It's locked. That's EB's fault. But whatever.
Those three items led to a huge bag of fail whenever our developer pushed changes. Elastic Beanstalk would initiate changes, but see that RDS’ configuration was out of whack from its understanding. It would fail and roll everything back.
But wait - there's more!
Elastic Beanstalk was also using some very old CloudFormation to make changes to the RDS instance. It was still using DBSecurityGroups, which apparently is illegal to use now… at least for our case. We were using postgres and minor version 9.6.6. It looks like the RDS team has moved on from DBSecurityGroups and now enforces the use of VPC Security Groups. Therefore, any changes to RDS would completely fail with the error:
Updating RDS database named: <instance> failed Reason: DB Security Groups can no longer be associated with this DB Instance. Use VPC Security Groups instead.
How do you fix all of this mess?
Let's go over how Elastic Beanstalk actually works. I'll be describing some of the simple concepts that are covered in documentation on the AWS site. Bookmark it and keep it handy.
First thing's first. You need to understand that Elastic Beanstalk is really driven by a simple YAML file. This YAML file is specific to the “environment”, which is a child of the “Application” in Elastic Beanstalk. This always confuses me because I think of an “Environment” as being a place to put an “Application,” but in Elastic Beanstalk it's backwards of how I think. AWS has a pretty good document on how you can look at this YAML file and see what's going on.
In this case, I was able to save the configuration as described in the AWS document. I then visited the S3 bucket and was able to see a few things that was making my life difficult. There was also a clue left in this document about how EB was driving changes to the RDS instance via CloudFormation. I knew this was happening. If you're using Elastic Beanstalk, take a few minutes to go look at your CloudFormation console. You'll see a template in there - one for each EB “environment” you have deployed. The top of your EB environment dashboard has an “environment ID” displayed in a very small font. This environment ID corresponds to the CloudFormation template ID in the CloudFormation console. You can see the nitty-gritty of what it's trying to do in there.
But Elastic Beanstalk is coughing up some invalid CloudFormation. How do I know? That security group error that was coming up is actually coming out of CloudFormation. I can see the error event in there. CloudFormation is the service that actually triggers the rollback. CloudFormation and RDS is enforcing the change away from DBSecurityGroups to VPCSecurityGroups. But when Elastic Beanstalk creates the CloudFormation template to initiate the change, it uses DBSecurityGroups.
I used one troubleshooting session to manually fix the CloudFormation JSON that Elastic Beanstalk is spitting out. I pushed it through by hand and it worked. I made the changes to the security groups in the way that CloudFormation and RDS expect - however, if I initiated a change through Elastic Beanstalk or the developer pushed a code update, it would fail with invalid CloudFormation once again.
I'll take a quick break to break down what's happening here. When you make a change in Elastic Beanstalk, my new understanding is that this happens:
Elastic Beanstalk console writes new YAML config file to S3 –> Elastic Beanstalk parses the config file and decides what changes should be made –> Elastic Beanstalk generates a CloudFormation JSON template –> Elastic Beanstalk saves the CloudFormation JSON to S3 –> Elastic Beanstalk pokes CloudFormation and asks it to update –> CloudFormation updates… if a failure is encountered, it rolls back and tells Elastic Beanstalk that everything is hosed –> Elastic Beanstalk rolls back the version of code that was deployed to a known good state.
Now I understand the root cause here. RDS made a change to enforce the security group update. Elastic Beanstalk can't seem to figure that out.
Here's how to resolve this.
Look at the AWS documentation on Elastic Beanstalk's config above. Follow their steps to save the configuration file from the console. Then, get your favorite code application out. Download the file and manipulate it by hand.
I changed the RDS properties to reflect reality. EB still thought it was postgres, version 9.6.2, on a db.t2.micro with 20gb of storage. I updated these properties to reality.
Then, I saw it. At the bottom of the file, there is a block of YAML that tells Elastic Beanstalk where to pick up the CloudFormation JSON and feed parameters. The default value was:
Take a look at that URL. Go ahead. I'll wait.
It's the bad CloudFormation template.
How did I resolve this? Well, I took that template and downloaded it. I modified it in my code editor to change the DBSecurityGroup resources into VPC Security Group resources. I had to manually add the SecurityGroupIngress information too, but because I speak CloudFormation this wasn't too hard. It's cheating a little bit, but not a big deal.
I created a new S3 bucket and uploaded my new CloudFormation JSON template into that bucket. Then, I revisited this YAML config and changed the URL to point to my new private copy of the CloudFormation template.
Go back to the Elastic Beanstalk console and “load” the configuration template and wham, it worked. Everything was fine.
Now I know how Elastic Beanstalk really works, and I figured out some super advanced ways to manipulate it to my bidding.
I hope this helps you understand Elastic Beanstalk a little more - it certainly helped me. Now I know how to trick Elastic Beanstalk into working if it hoses up again.
Since it's working, turn off minor version upgrade in RDS to prevent this from happening, then use your AWS support plan to tell them that Elastic Beanstalk has a bug with CloudFormation and RDS security groups :)
Happy cloud days.