The future of Terraform CDK
github.com134 points by mfornasa 3 days ago
134 points by mfornasa 3 days ago
It's odd to always say "Hashicorp, an IBM company". Looks like they want to assign blame.
I did try Pulumi a while back, but the compatibility with Terraform modules was not great, so I've switched to CDKTF, which can handle unmodified modules. Dunno if I'll switch back to Pulumi or just use OpenTofu directly.
> It's odd to always say "Hashicorp, an IBM company". Looks like they want to assign blame.
All their branding does this now, including the HashiCorp logo on their website [0]. There's gotta be a name for this specific branding pattern, but I don't know it.
It’s endorsed branding. Basically when a parent company “endorses” its subsidiaries’ brands, but keep their own name (as opposed to renaming everything to IBM, like eg Google would do).
I was recently working for a company which got acquired by IBM and we had to do it too. It’s an IBM thing. I bet most people at HashiCorp hate it, at least that was the case for us.
Makes IBM look really bad. Do they also force people to bow when the CEO of IBM enters the room, and address them as sir or your highness?
They used to have their employees sign songs praising the company...
Granted, that was in the 1930s or something, but still.
I remember it being a part of the movie Pirates of Silicon Valley and thought it couldn’t be true. But apparently it was.
I have absolutely nothing good to say about Pulumi. Stay far, far away.
My experience with Pulumi is you can write bad pulumi code and good pulumi code and just like everything else, it's easy to end up in a codebase where one poor soul was tasked with writing it all and they didn't do the best job with it.
Please expand on your experiences, because I've had great luck with Pulumi at my company since October 2021. No engineer liked HCL, our demographic was engineers who were familiar with programming languages who wanted to self service basic infrastructure (AWS SecretsManager, IRSA roles, Databricks Service Principals, etc). We were pretty easily able to shim in a RunAtlantis inspired system that displayed previews that required explicit approval when a PR was raised, performed apply on merge to main, and ran drift checks periodically.
Their stack builds a lot of abstractions on top of each other and this works only well as long as you don't deviate from the beaten path.
One example:
You can't really build custom TS providers for AWS resources.
Why?
Because this feature is built using the compilation magic that makes inline lambdas work.
But the compilation step omits the AWS SDKs since these are present in a lambda anyways. So you can't use the AWS SDK in custom providers.
For me, the ideal is each team owns its own config/lifecycle mgmt, and does it in the language they wrote the rest of the system in.
Why? I’ve had nothing but good experiences, but I don’t run it and the team that does is extremely competent
Strange, I have a lot of good things to say about both it and Terraform.
Probably some specifics might be more useful there...
My experience is that by stealing providers from Terraform, they failed to properly handle statically typed languages (Go) with certain providers (HCloud); I had problems with their ID type and had to abandon my Pulumi setup.
Have a look at https://github.com/purpleidea/mgmt/ and tell me what you think. We don't have enough docs though. Tough being an open source thing that you want to keep open.
Running SST with Pulumi and it's been a great experience. Infrastructure and maintenance has been pleasant and SST's pre-fabs really make things easy to spin up resources.
We use OpenTofu it’s pretty seamless
Now more will be using a combination of OpenTofu and Terraform, and there will probably be some tacit endorsement of OpenTofu by Hashicorp folks in their communication with those who are using both. Good to see!
Does it do ephemeral values yet?
Yep, as of yesterday’s 1.11 release it’s supported!
That also includes a new “enabled” meta argument, so you don’t have to hack around conditional resources with count = 0.
[0]: https://opentofu.org/blog/opentofu-1-11-0/
Disclaimer: affiliated with the project
How do you migrate from count/for_each to `enabled` ?
You can just switch from `count = 1` to `enabled = true` (or vice-versa, works back-and-forth) for a resource and tofu will automatically move it next time you apply.
It's pretty seamless.
That's cool! We'll still need to change all of the references to `resource[0]`, right? Or does tofu obviate that need as well?
I’m not sure I understand. You refer to the conditional resource fields normally - without list indices. You just have to make sure the object isn’t null.
There’s some samples in the docs[0] on safe access patterns!
[0]: https://opentofu.org/docs/language/meta-arguments/enabled/
And you don't get the annoying array form for the resulting resource with the `enabled` syntax, right?
EDIT: Oh just realized the sibling asked the same, but the doc doesn't state that clearly, although it seems to me that the doc implies that yeah, it doesn't use the array form anymore.
Yes indeed! It does not use the annoying array form.
Worth switching to Opentofu only for this, then! I fuckin hate the count pattern for conditional present/not present that leads to an array of size == 1.
Damn, might finally be able to use it. The lack of ephemeral values was a major blocker.
I was thinking the same thing about the "an IBM company". My guess is that it's a lazy find/replace.
I assume it's a matter of branding and making IBM look more modern by associating with the Hashicorp brand.
It’s one thing to say it once but 3 times in the same paragraph seems weird for sure!
> It's odd to always say "Hashicorp, an IBM company". Looks like they want to assign blame.
Or it's legal trying to preempt a risk.
If it was the author just wanting to point at IBM, they'd mention it just once or twice, but using that awkward phrase throughout the text makes me think it was an edit mandated by a careful lawyer.
"Hashicorp, an IBM company"
Common sense would be IBM mandating that branding, as opposed to Hashicorp.
They should have renamed it first to HashiCorp, an IBM Company CDK, then shut it down
It’s how Red Hat identifies themselves too
It's common when corps buy large enough companies that they don't want to kill the original brand. That's why you get hotels like "(something) by Hilton".
Do you mean Red Hat identifies itself using the phrase "Red Hat, an IBM Company"? Because I don't see any use of this on redhat.com (including that website's corporate "about" content) and if any Red Hatters are using this phrasing (I'm a current Red Hat employee) I haven't been aware of it.
I have seen it in several articles. (I don’t work for HashiCorp or Red Hat subsidiaries so no idea what is said or done inside those subs)
1. https://business.adobe.com/customer-success-stories/red-hat....
2. https://www.openpr.com/news/4100338/linux-operating-system-m...
You said "It’s how Red Hat identifies themselves too" but those two articles are not by Red Hat. So ... it's not how Red Hat identifies themselves, though it seems to be how Hashicorp identifies themselves.
This is particularly frustrating as I've spent the last year writing many thousands of lines of CDKTF Python.
HCL just does not have the modularity and expressiveness that Python, or other languages CDKTF supports.
I guess I'll spend another year migrating to Pulumi now..
The lack of expressiveness of HCL is the point and what makes it so good
Being able to inherit from Ingress and add a parameter of say public=True/False and then it change annotations, middleware, etc and then being able to re-use that across 100s of stacks is very powerful. DRY is not something HCL is good at.
Getting too clever with an imperative language in what is inherently a declarative domain, is an idea bad enough that they invented a whole new language to avoid you doing it. But some lessons have to be learned the hard way I guess
The problem is they did an exceptionally poor job at designing their language. A reasonably large Terraform codebase is almost universally hard to read for one of two reasons: it's either unexpressive (read: verbose to the point it's hard to read) or modularized but hard to read because it's fragmented into a bajillion reusable modules.
SQL is also declarative, but incredibly expressive. A thousand character query contains enough complexity that it's hard to reason about. A thousand characters of Terraform will barely stand up a CRUD app on AWS.
Designing a language from first principles for this was a mistake. HCL is awful; they should have gone the Starlark route and made a stripped-down version of an existing language instead of making their own language from scratch. This feels like the worst of both worlds. The language is practically imperative, but it has its own syntax that isn't useful outside of this one single domain.
> reasonably large
Anyway you shouldn't have too many resources in a single Terraform workspace, for performance reasons. The real issues with Terraform come when you start to want to orchestrate different workspaces triggering each other, and trying to write that orchestration language, which itself would be declarative.
Terraform built a Stacks feature, but support is Terraform Cloud-only. OpenTofu has issues in the area that have been open for years: https://github.com/opentofu/opentofu/issues/931 https://github.com/opentofu/opentofu/issues/2860 and progress is slow, in part (IMO) because a genuine solution requires server-side evaluation (i.e. triggering applies as Kubernetes Jobs) and the open-source implementation of Terraform Enterprise/Cloud is a completely separate project with a completely different group of maintainers, Terrakube.
I'd argue the real issue with Terraform is that workspace orchestration is necessary in the first place. If they addressed the performance issues with large workspaces, then we wouldn't need to split up workspaces and Terraform could just orchestrate changes naturally.
The performance issues in large workspaces are due to needing to refresh status on all the resources in the large workspace before coming up with a plan. Actual apply time is either negligible or the inherently long amount of time it's supposed to take.
You split the workspace into smaller workspaces precisely to tell Terraform that you haven't made any changes to the networking layer, so don't bother trying to refresh the status of the networking layer to see if any changes are needed, it's not relevant when you're trying to scale up your Kubernetes cluster or whatever.
Declarative vs. imperative doesn't have anything to do with power or expressiveness. Some general purpose programming languages are declarative, and some declarative DSLs are Turing-complete.
I worry that comments like this lead the average newbie to overlook (or worse, avoid) declarative languages (both among DSLs and among general-purpose languages) because they will associate the term with hacky, confining, gotcha-ridden messes like Terraform's HCL, Azure DevOps' standards-breaking "YAML" DSL, etc.
Incidentally I agree that a language like Python is a terrible fit for this domain, but it's also plain to see that HCL is a shitty tarpit. It's not hard to understand why people want to get away from HCL.
And concretely, you can use Pulumi in a pure functional style with F# or Scala.
They invented a language to avoid you imperatively updating infrastructure, but that's not what CDKTF does; it just makes it easier to materialize that declarative output.
It also makes it easier to reason about that output as you can avoid awkward iteration in your declarative spec.
Yet said language continues to add imperative-inspired constructs to make up for its limitations..
The end result is still declarative, your just using an imperative language to keep your IaC DRY.
If you have the expertise and restraint to not go off the rails, I agree, imperative is more powerful. That plan does not survive teams of sizes over 2 in the majority of cases.
But it's not even imperative. Your code runs, declares all its resources up front and then normal terraform runs on it. With cdktf you can even have it output the HCL.
At the point where we are templating Terraform files we've already lost the plot. You might as well get to use a real programming language.
I have used Terraform, Puppet, Helm, and Ansible (although that's not strictly declarative), and all of them ran into problems in real-world use cases that needed common imperative language features to solve.
Not only does grafting this functionality onto a language after-the-fact inevitably result in a usability nightmare, it also gets in the way of enabling developer self-service for these tools.
When a developer used to the features and functionality of full-featured language sees something ridiculous like Terraform's `count` parameter being overloaded as a conditional (because Terraform's HCL wasn't designed with conditional logic support, even though every tool in this class has always needed it), they go JoePesciWhatTheFuckIsThisPieceOfShit.mp4 at it, and just kick it over to Ops (or whoever gets saddled with grunt work) to deal with.
I'm seeing the team I'm working with going down that same road with Helm right now. It's just layers of templating YAML, and in addition to looking completely ugly and having no real support for introspection (so in order to see what the Helm chart actually does, you essentially have to compile it first), it has such a steep learning curve that no one other than the person that come up with this approach wants to even touch it, even though enabling developer self-service was an explicit goal of our Kubernetes efforts. It's absolutely maddening.
That is... not a good idea at all imo. It's very, very easy to over-DRY infrastructure config and it sounds like you're well past that point.
That's very subjective. Concepts like iterations are inevitable, and they don't look great in a declarative language like HCL.
I also find refactorings considerably harder in a declarative language, since configurations have a rigid structure.
"Will be sunset on Dec 10"... commit date: Dec 10.
That seems like rather short notice.
As I said here [0] there's more of this coming.
[0] - https://news.ycombinator.com/item?id=46192130#46198058
That’s a real shame. It seems like Pulumi is the only alternative for internal DSLs for IaaC? I always found HCL to be quite terrible, slowly becoming less painful, but not really refactoring-friendly.
Terraform CDK had promise as a blessed infrastructure-as-actual-code solution from the official maintainer of Terraform, so easier to sell internally rather than something from a new vendor like Pulumi. I feel sorry for those teams who have migrated to TF CDK.
Internal vs external DSLs explained in the middle of this page: https://martinfowler.com/bliki/DslQandA.html
Rug pulls on infrastructure components seem even worse than other rug pulls as they can hit your entire infra codebase at once
This is why infrastructure people are conservative by nature, it's so damn much gruntwork to migrate without downtime
And it happens while we are all very enthusiastically dedicated to migrating off Kubernetes ingress-nginx. Just as planned.
As an Infrastructure Engineer who used it: I blame people who didnt help fund/maintain it (including ourselves)
Sad to see it go. The philosophy of CDK has been to offer a shared ecosystem between IaC, backend code and frontend code, allowing to share configuration, data structures and libraries between all of them. It has made development more unified and have less redundancy and manual work. Personally I don't want to repeat some stuff in a special Terraform language, if I can find a way to manage the whole application in TypeScript.
Pulumi
Thanks, will definitely look into it. I first used Pulumi when it was just a cloud platform but seems it is a more general devops tool now.
Damn, what are the best alternatives here? For pure AWS I guess CDK directly is okay, but locks you in.
I went with CDK, I'm locked into AWS already and it means my major dependency for IaC is my cloud vendor and not a third party.
If I really need to migrate off of AWS at some point I'll throw an LLM at it.
IaaC code is one of those use cases just throwing LLM is painful for a refactor.
In my experience claude/codex to wrangle CDK constructs be complicated, it frequently hallucinates constructs that simply do not exist, options that are not supported etc.
While they can generate IaaC component mostly okay and these problems can be managed, Iterations can take a lot of time, each checkpoint, goes the deploy/ rollback cycles in CF. CloudFormation is also not particularly fast, other IaaC frameworks are not that different.
Running an agent to iterate until it gets it right is just more difficult with IaaC refactor projects. Hallucinations, stuck loops and other issues, can quickly run the infra bill up not to mention security.
I had Gemini ingest our huge aws cloudformation repo . I had it describe each infrastructure component and how it related to others and creation hierarchy and IAM.
I got a nice and comprehensive infrastructure requirement document out of this.
Now I am using it to create Terraform repo , deploying it via OpenTofu and comparing it to my existing AWS cloud formation . This part is still a WIP .
I have used ChatGPT to generate perfect IaaC using the CDK and Terraform. I give it my labelled descriptive design diagram that I have to do anyway.
I am very detailed about all of the security group requirements, tell it that I don’t need Internet access and tell it which VPC endpoints. I don’t do “agentic coding”.
Exactly. It's just so much cleaner to do it in the Cloud provider's native tooling. The impedance mismatch from Cloud-agnostic abstractions always just makes thing shitty enough that in the long run you spend more time dealing with weird edge cases.
Besides, actual full-scale Cloud migrations are exceedingly rare.
Terraform is not an abstraction on top of multiple cloud providers, you work with aws, azure etc explicitly. It is , however, agnostic in the sense that you can provision aws, azure, gcp, etc resources within the same iac project
I always hated this meme. Using Terraform no more makes you “cloud agnostic” than using Python to script AWS services and calling boto3 than using bash and calling the AWS CLI.
AWS's native tooling is Cloudformation, and CDK is actually just a wrapper around that that generates cloudformation code (as CDKTF is a wrapper for terraform). And I like to avoid cloudformation as much as possible.
writing HCL is so much more enjoyable than writing CF, even if HCL is fairly verbose (hey, it's not as bad as XML!). CF feels like a series of PM requirements dutifully codified with no dogfooding whereas HCL/TF feels like a tool that was developed by people who actually wanted to use it.
> but locks you in
Not picking on you personally but having had this conversation many times over many years with many clients I find it confounding. Oftentimes TF itself was heavily promoted as a way of "avoiding lock in".
Well guess what? Now you're locked-in to IBM, whose motivations may not be perfectly aligned to you goals of simply and efficiently using your cloud provider of choice to deliver your business outcomes.
What we refer to as lock-in is simply an expression of risk, with one axis being the cost of getting off $solution and the other being the likelihood of needing to do so. Having stepped through this exercise a few times, the cost of rewriting your e.g. AWS API Gateway + Lambda + SQS + RDS + CloudWatch etc architecture invariably vastly dwarfed the cost of changing the IaC language it is expressed in.
Anytime you feel the urge to overbuild on a cloud provider's services, stop, and do a really rigorous cost/benefit analysis. If you truly have unique drivers the data should tell the story.
I made https://github.com/andrewbaxter/terrars ! It's great! You get more benefits if you're in a Rust project (obviously) but it has some things that make it a good alternative anywhere:
- More accurate types/type safety than the CDK (for static feedback on required parameters, etc)
- No CLI required - just plain Rust (provider definitions can be published as normal rust packages so you don't have to generate them yourselves, and I've published a bunch of common ones - docker, aws, etc)
- Simpler: Terraform CDK had this crazy flow where it (go code) generated typescript code then used some transpiler to generate target language code. The output wasn't pretty, and there were bugs. Your project directory would get filled with boilerplate generated files.
It generates tf json files and has a fairly safe way for handling variable interpolation and escapes - I haven't hit any weird bugs with it.
Hm, we have a few very repetitive terraform projects to setup structured infrastructure clusters. For those, we just use ansible with a bunch of templating to generate a configurable, HCL-based terraform module and version that.
It's a bit of an "Caveman solve problem with rock" approach, but for very regular projects it's great. A new cluster is some group vars, larger changes to the structures can be easily reviewed - and if you really really have to, you can also just modify the generated code by hand to fix something your generation code can't deal with right now.
Just use Terraform?
The value of CDK was always that it allowed you to write in a fully orthogonal language rather than the poor pseudo-language of hcl.
When writing stacks you need normal language features: loops, yes, but also if statements, reuse (functions), being able to do stuff like complex string parsing and re-formatting, etc etc.
HCL supports loops, modules can be kind-of used as functions with lots of footguns, there are awful hacks for some other things, and some stuff just couldn't be done.
Yeah I'm struggling to see the value here.
The value for TFCDK was Developers don't have to learn another language, they can just continue to use existing language they already know.
Downsides are doing infrastructure in a programming language was always problematic unless developer was skilled at Ops which most who used TFCDK were not.
I ought to have phrased it I guess as "I don't agree with the value proposition", mainly because of the downside you point out. This seems superior to Pulumi, though, in that the abstraction is (was) at least owned by Hashicorp so there was less likelihood of it falling out of date and giving you footguns.
That might have been the promise but never the real value. As you say in practice the engineer needs to know ops & terraform along side their language of choice.
The real value of cdktf was more dynamic infrastructure provisioning while still having the plan / apply pattern.
Probably Pulumi
I'll be honest Pulumi is pretty cool but I'm a little worried by how high on the stack it is. I wonder if the same thing won't happen to them that's happening to CDKTF here.
Terraform is ugly but it works well enough for me and seems ingrained enough to be durable to this kind of thing (i.e. I bet for sure the community would pick it up (I wish I could say that I'm part of that community but I can't say I use it quite that often))
> I wonder if the same thing won't happen to them that's happening to CDKTF here.
This is clearly a business decision rather than technical.
Pulumi is meant to be semi-automated (in generating the bridges) so perhaps is slightly better off in maintenance.
Terranix? ;-)
Not gonna lie Terranix has been working great for us, all our configuration is in Nix files anyway so it's so easy to just pass stuff in rather than using Tf variables etc
At least they gave us some notice, that’s much appreciated.
Hashicorp, an IBM company
As far as corporate mercy killings goes archived under mozilla license is better than a pivot to "you now pay per core" or whatever
CDKTF works beautifully, all the complains here seem to be from salty devops who got pissed the developers wanted something more powerful than the garbage HCL with its even more garbage module system.
CDKTF stacks are great and the construct pattern gives you modularization without all the baggage.
Not stated in the most diplomatic way, but I do agree. Having used CDK (not cdktf) and now being forced back to Terraform feels like going back to the stone age. It is absolutely obvious to me that generating infrastructure definitions from a regular, testable language using all the same tools, techniques and distribution mechanisms that you use for all your other software development is the superior way. Being able to piggyback off of the vast ecosystem of Terraform providers was a really clever move, although I understand it led to some rough edges.
I kind of like it but I always found it kind of clunky how it's ultimately just generating JSON/HCL anyway. For instance, you can't data source then use code to transform and send it to a resource since it has to transpile first.
That also means you end up with things like the language's native JSON not doing what you expect and having to use a special Terraform function call.
This is a bummer. I don't particularly like Pulumi but use it anyways because for my use cases being able to write actual code is really impactful. Sucks to see fewer options in that space
The often excluded option is dynamically generating JSON and feeding that to TF instead of HCL.
You can combine it with tools like Dhall or my personal preference Jsonnet instead of imperative languages for an interesting experience for reusable pieces outside of module concepts.
Any particular libraries you use to generate TF-JSON from jsonnet?
I wrote a generator a little while ago that can create jsonnet libraries from the TF schemas: https://github.com/Duologic/soysonnet
Example lib here: https://github.com/Duologic/soysonnet-aws
I only needed it for AWS so I didn't spend more time on it.
[Pulumi founder here] Sorry to hear you don't particularly like Pulumi---any/all feedback welcome. If nothing else, we do listen and we do try to get better. -Joe
What is it that you don't like about Pulumi? As I mentioned in another comment, my team of backend-engineers who took over an infra team went from Cloudformation -> CDK -> Terraform -> Pulumi and honestly find it the most approachable for other engineers familiar with normal programming languages (sorry HCL). We've been using it since 2021 and have a "what's on main is what's deployed" philosophy and adopted a RunAtlantis inspired workflow where previews are run as status checks on PRs and require explicit approvals, apply is run on merge to main and periodically, and drift checks run preview+refresh and alerts if what's checked in doesn't match what exists. We don't really use stacks, we just use a separate project for everything and write code to encapsulate modules (and luckily we can easily write unit tests and runtime assertions).
This is so sad. It’s a great project. Needs to be forked and maintained. If anyone forks please email me I’ll contribute.
OpenTofu is already the de facto fork.
OpenTofu is a fork of terraform, not CDKTF.
Although, I would hope a fork of cdktf would target opentofu instead of terraform.
As an alternative is anyone considering https://sst.dev/ (which uses Pulumi under the hood)? We use it at work and I’ve been quite happy with it
It’s not an alternative at all. Terraform CDK is basically TypeScript transpired to HCL. You can codegen TypeScript bindings for any provider. And then write normal TypeScript.
It would be great to have an alternative to Terraform that uses a bit more advanced provider (at last for AWS). Does OpenTofu use that same provider?
The providers for tofu are by design the same as for terraform.
Also, for large providers like AWS, GCP, Azure, etc - these are often largely authored by the hyperscaler themselves, for better or worse.
It does. What are you looking for in a more advanced AWS provider?
Have you ever debugged what goes on with lets say the aws provider during a tf apply?
we're using cdk since 100% of our stuff is in aws but will soon need to hook up some external resources like cloudflare. looked at tfcdk a while back but didn't think it was a good idea (glad). still trying to figure out a good way forward and hoping it's not to rip the bandaid and migrate everything to terraform / pulumi
Well that sucks for me.
Good move. They clearly didn't have the resources they needed. The design of the CDKs was atrocious.
What was the point of it? Terraform supports AWS anyway.
[dead]