I disagree. Every AWS Lambda function I've ever written can be ran as a regular node/python process. The lambda-specific part is miniscule. If I wanted to run these on Azure or Google only the most inconsequential parts of the function would need to be changed.
In my experience, having started and abandoned side projects in both aws lambda and google app engine, half your project becomes:
* Well, obviously we use a hosted database
* And obviously, AWS provides our logging and all our analytics.
* Obviously when people call our lambda functions, they do so either through an AWS-specific API, or one constrained to a very limited set of forms.
* Of course, we can't blindly let everything access everything, so naturally we have IAM roles and permissions for every lambda function.
* Well, the cloud provider will look after secrets and things like that for us, no need for us to worry about database passwords.
* Naturally, with all these functions and IAM roles to look after, and we need tagging for billing. We should define it all with CloudFormation scripting.
* Well, the nosql database the they provide comes with their specific library. And as it shards things like this, and doesn't let you index things like that, you've got to structure your data this specific way if you want to avoid performance problems.
* You don't want your function to take 200ms+ to respond, your users will notice how slow it is. So no installing things with apt-get or pip for you, let me get you a guide on how to repackage those into the vendor-specific bundle format.
* You want to test your functions locally, with access to an interactive debugger? You're living in the past, modern developers deploy to a beta environment and debug exclusively with print statements.
* And so on.
In this case, a lot of the 'complexity' one hoped to eliminate has just been moved into XML files and weird console GUIs.
This, but it was told by serverless experts, on stage, in front of hundreds of people. If that was their sales pitch, their reality was likely even less impressive.
This sounds a bit like the posh workshop I went to on WAP (wireless access protocol) years ago when I worked for BT.
It was a complete omnishambles - to the point I avoided the fancy lunch and went to the Pub for a ploughman's lunch, In case I suddenly blurted out "this is all S*&T" and caused a political row with the mobile side of the company I worked for.
Never tried something marketed as Serverless but Google App Engine. If you wanted any performance, you had to follow the Guidelines really closely. Which could be legit, if it's worth the effort but I think it isn't. I think people under-estimate the effort and that the code will require much more not so nice optimizations than expected. That includes very verbose logging. It's sold as carefree and elegant but it only works when using patterns that nobody enjoys. I really liked the log browser and Dashboard though ;) It's like a stripped down version of New Relic and Elastic Search combined.
Most of your points are not relevant to my original statement of the code being generic - they are talking more about the architectural decisions. If I want to use DynamoDB I can do so in EC2 or Lambda, serverless doesn't dictate that. You also seem to believe one chooses Lambda because of complexity reduction and that's not really the only reason. I can very easily port a node/express API backend that connect to RDS to any other cloud provider. What about serverless makes you think that's not the case?
> Every AWS Lambda function I've ever written can be ran as a regular node/python process. The lambda-specific part is miniscule.
Of the actual function code, sure.
Of course, if you aren't manually configuring everything and are doing IaC, that isn't generic. And if you are supporting the Lambda with any other AWS serverless services, the code interfacing with them is pretty AWS specific.
I was only talking about function code. It should be painfully obvious to anyone that if you opt to leverage other AWS services that those are things you would ultimately have to replace but those decisions have nothing to do with serverless.
> but those decisions have nothing to do with serverless.
Sure they do.
Because if you need a DB for persistence you are either setting up a server for it (and therefore not serverless, even if part of your system uses a Lambda) or consuming a serverless DB service. And so on for other parts of the stack. On AWS, for DB, that might be Dynamo (lock-in heavy) or, say, Aurora Serverless (no real lock-in if you don't use the RDS Data API but instead use the normal MySQL or Postgres API, but that's higher friction—more involved VPC setup—to use from Lambda than RDS Data API is, so the path of least resistance leads to lockin.)
Lambda or other FaaS is often part of a serverless solution, but is rarely a serverless solution by itself.
