Pulumi – A new open-source cloud development platform
joeduffyblog.com
joeduffyblog.com
A thing I notice about the space is that existing tooling tends to be very focused on a specific persona (e.g., ops), and when you want to do anything outside those use cases, it starts to feel like you're jamming a square peg into a round hole. So, e.g., Go templates/sed/awk work great if all you want to do is template a couple YAML files, but if your use cases is even slightly more demanding (sharing, dependencies, transforming), it gets painful quick.
Using a "real language" makes all of this much simpler. You can write tooling that captures the needs you actually have. If you have an ops team that wants a mostly-declarative or mostly-hermitic API, you can write that, and you don't have to map it to CloudFormation, ARM templates, Terraform, etc. Huge win. You can write JS/TS/Python libraries that provide secure defaults for your core infrastructure components. You can use standard dev tools to audit and refactor code you've written.
In Terraform, modules have proven to be very useful. For Pulumi, I'm interested to see what the tooling ecosystem will look like as the product becomes more mature.
>Using Pulumi, you author cloud programs using your favorite language, spanning low-level infrastructure-as-code to highly productive and modern container- and serverless-powered applications.
I'm enthusiastic about the prospect of using a real programming language to deploy and control a distributed system. It would be fantastic to have a library that gives me a nice, clean API for performing low-level distributed system tasks, manipulating and controlling hosts at the level of individual processes and sockets.
But I was disappointed to see that Pulumi doesn't seem to be designed to support deployment onto bare metal. In fact it seems to be just another layer of abstraction over various SaaS providers, or a huge dependency like Kubernetes. Why add another layer?
Today's systems are already heavily abstracted. For me, the value of using a really expressive programming language to implement a system, is that you can keep more of the abstraction in the language, and thereby avoid complexity in the underlying architecture.
To me, it is not very interesting to layer yet another library on top of the Kubernetes API. I'd be much more interested to see an attempt to replace the Kubernetes API with a library. This seems to be the former, and not oriented towards the latter, which is a bit of a disappointment.
In principle, there is no reason we couldn't support bare metal, provided we have an orchestrator. Pulumi bottoms out on raw CRUD operations that can do anything. It turns out that, when targeting existing cloud providers, that means invoking their APIs. As you say, this is just a layer of abstraction, but the key aspect at the center here is diffable immutable infrastructure.
I was, in fact, inspired to pursue Pulumi in part because of the experience of building a distributed OS that ran on bare metal. Many of the same challenges Pulumi solves around configuration, composition, and the application model, were inspired by this experience, and we've taken some of those lessons and applied them to cloud programs.
> you can keep more of the abstraction in the language, and thereby avoid complexity in the underlying architecture.
I look at other ways we've integrated with systems as inspiration here. For instance, we can productively code using tasks, and async await, these days, and reap the benefits of concurrency, despite there being massive differences between threading, I/O, and scheduling behavior across different systems. There was abstraction in the language, but it bottoms out on a set of primitives.
In a sense, I view the current APIs we're using as the syscall layer for a given cloud, Kubernetes included. (Also note, we don't require depending on Kubernetes.) That's not to say everybody will want to eschew the syscalls -- there's always a time and a place -- but we can be more productive by using the abstractions.
I believe that frameworks are languages, in that they give you a set of nouns and verbs you can use to express yourself. I do think there's opportunity for language-level innovation here, but if there's one thing I've taken away from our industry experiences building PaaS's, it's that it's a losing battle to bet against the incredible momentum of the public cloud vendors.
> I'd be much more interested to see an attempt to replace the Kubernetes API with a library.
We are just getting started, so expect to see a lot more progress in this direction in the weeks to come. The foundation of languages is a huge enabler for us, and we already have several projects in the works that I think will make you happier than this initial release.
YMMV, this is just my opinion, and I really love hearing the feedback. Thank you for sharing. This is a vibrant space and time and I'm enjoying the diversity of perspective.
I see the wisdom in it, though. Most developers these days aren't shipping their own servers off to a data center. And for those that want bare metal, there are providers that offer bare metal servers provisioned through a CRUD API, so Pulumi should be able to work with those.
If Terraform works for you, then definitely continue using it. Pulumi is just another option and works well for building libraries and components. I did a rundown of different tools for serverless apps at Velocity SF last week. Slides are here: https://cdn.oreillystatic.com/en/assets/1/event/270/Tooling%... and GitHub samples are here: https://github.com/lindydonna/velocity-examples
There are a ton of tools in this space, all with overlapping functionality, so there tends not to be a clear comparison of tool X vs tool Y.
My colleague wrote a blog post with more of the technical motivation: https://medium.com/@lukeh/programming-the-cloud-e795cafffc2b, and there's also this Twitch live-coding video: https://youtu.be/DM8Wd4f1MNA
It sounds like the value add for this tool is being able to describe your infrastructure as code rather than configuration, with all the benefits of code over configuration (diffs, loops, package managers, variables, etc).
So if, for example, you're running an app on GKE, then you can tell the Kubernetes API about the resources you want, but there's no system for building your app (Docker build + push to GCR), no way to do supervised rolling updates, and so on. If you're doing CI/CD, then presumably you have to build, push, then let the Pulumi code inject the right image and tag.
I like the idea of declaring the infrastructure with extensible code (as opposed to a DSL like Terraform where you're completely dependent on upstream support for all the objects you want to manage), but I'm a bit skeptical about crossing the bridge from ops to applications.
Typically applications aren't self-contained enough that they can, or should, declare all their resources. So presumably then you have to centralize common stuff (e.g. your central Postgres server) in a shared module that all your apps import. Now you get into versioning hell as your app depends on an old version of the "common" module and "pulumi update" tears down your Postgres 10 install and creates a 9.6 install instead. (Presumably it asks first. But still. Versioning has got to be a challenge here.)
> I may be wrong, but from what I can tell this is really a souped-up framework for working with clouds. Most of it looks like primitive for immutable infrastructure, so Terraform but as a library. As far as can tell, there's no deployment system, no orchestration etc. -- a "Pulumi app" doesn't know how to deploy itself. Is that accurate?
Not quite. A Pulumi app is always deployed through the Pulumi CLI, which deploys and manages Pulumi programs. Note that most Pulumi code runs at deployment time, not runtime. Instead of specifying resources in a configuration language, you write them in code. The Pulumi CLI turns this code into a declarative plan, and updates your infrastructure when you run `pulumi update`. You may find this doc page helpful: https://pulumi.io/reference/how.html
Docker builds, plus provisioning of a container registry instance (and deploying to it), are handled automatically. This blog post walks through the e2e container scenario (but on AWS, rather than GCP): http://blog.pulumi.com/deploying-production-ready-containers...
> Typically applications aren't self-contained enough that they can, or should, declare all their resources. So presumably then you have to centralize common stuff (e.g. your central Postgres server) in a shared module that all your apps import. Now you get into versioning hell as your app depends on an old version of the "common" module and "pulumi update" tears down your Postgres 10 install and creates a 9.6 install instead. (Presumably it asks first. But still. Versioning has got to be a challenge here.)
This is just one way to architect an app, and it's unlikely to work well for a database, as you say. In this case, you'd likely use a `database.get` call, where you reference an existing database, that may be managed in a different Pulumi program or stack, or even outside of Pulumi.
I'll keep trying to understand it, anyway.
Pulumi lets you describe cloud resources using code instead of a config language. It's not like Heroku, it's more like a deployment tool (e.g. Serverless Framework, Terraform, Claudia.js, Chalice, etc).
The difference compared to other deployment tools is that you use regular code, but it's turned into a declarative plan when you run `pulumi update`. So, you get the benefits of a regular programming language, while still following best practices of immutable infrastructure.
Does that help?
Having used a variety of tools in the space, this is the first time I ever had an engineer say to me: "I really like it, I hope we can keep using it."
But for a complete newbie. Say someone who can write an app and push it to Heroku, what is this and why might it be worth trying out?
If you are writing (e.g.) Rails code that you `git push` to a PaaS like Heroku, then they're provisioning this infrastructure on your behalf. So this is a bit lower level than that.
1. One of my pet peeves with CloudFormation is the disconnect between the state of CF and the "actual" state if someone decides to update resources by not using CF. Does Pulumi do checks to verify that the actual state matches its assumed state?
2. Is there a guide for contributing a new language and/or provider?
Also, I tried getting an invite for the Pulumi Slack community and was greeted with "invite_limit_reached".
We're hitting Slack's rate limiting on invites. Please DM your email @PulumiCorp on Twitter and we'll get you added!
There is a top-level CLI verb called `refresh` that explicitly checks Pulumi's idea of the state of the world against the actual state of the world. If there's a diff, Pulumi reconciles the diff by updating its own state. `pulumi update` does not do this by default, but you can run `pulumi refresh` at any time to verify that your state has not drifted from reality.
Also, this reminds me a bit of metaparticle, which I’m quite fond of.
I don't mean this as a joke, I am seriously having issues looking at HN fonts after reading the blog. Even my emacs fonts are broken to me now.
How can a site just steal my ignorant bliss away and ruin my entire system with something as simple as fonts!
An approach that goes back to relying on procedural programming languages as a primary means of configuration seems like it defeats the purpose.
I am working in this space, mainly focused on configurations: https://youtu.be/4yepPOznakk
The approach we've adopted is to maintain the declarative presentation while providing a way to plug in procedural language programmability without using templates or other similar concoctions (ksonnet..etc).
You should be able to read the source of the script directly at https://get.pulumi.com/ or even download it yourself, inspect it and then decide if you want to run it or not. You can also follow the manual install instructions which are listed a little further down on our install page at https://pulumi.io/install/, if you want total control over where it is installed.
I also just opened a pull request to our docs website: https://github.com/pulumi/docs/pull/430 to explain in a little more detail what we're doing with the scripts.
That way I'd have a convenient way of keeping it updated to the latest version
I'm worried that I might run into a case where it might provision something and I wouldn't know what it did to my AWS account necessarily.
I'll 10/10 give this a go for my next AWS project. Also, no YAML woo!
The `pulumi update` command first shows you a preview of the exact resources it will create. Once you've deployed, you can view all your resources on the pulumi.com console, so you know exactly what has happened.
What is difference in the use case for Pulumi compared to scripting infrastructure using vendor provided libraries e.g. New-AzureRmVm et c
Thanks -- we're very excited to be out and talking to the community!
> What is difference in the use case for Pulumi compared to scripting infrastructure using vendor provided libraries
The major difference is that Pulumi does immutable infrastructure. The code describes a goal state and our system manages your environment to ensure that it matches the goal state. This means you can preview changes before making them, and that, if you do make them, you've got a full audit trail of who changed what and when. Rollbacks are trivial because you just point us at an old goal state worked and we can chew on it until your live environment matches. We can even detect drift, so if someone manually updates in the cloud console, we can tell you about it. As a result, we always have a full object graph of your cloud resources, and can tie it back to the source code which created it, opening up some unique diagnostics capabilities.
The difference with scripting libraries, like AWS's Boto, or the Azure PowerShell SDK, is that they mutate cloud resources directly, and in an ad-hoc manner. So, you don't know what they are going to do before you run them. And in the event of failure, you're more likely to be in a corrupt state and unable to recover. Rollbacks are difficult. There's also no audit trail beyond the cloud access logs, and information like dependencies are lost, so resources end up disassociated from the code and workflow that created or updated them. Many people encounter these problems and need to build complex systems on top to address them. Or they end up using a solution like CloudFormation or Azure Resource Manager, which eschews code in favor of JSON/YAML/templates.
The hope here is that we've landed on a sweet spot between immutable infrastructure and fully programmable languages.
(Disclosure: I'm a member of the Pulumi team).
Strictly a timing issue. We brought TypeScript/JavaScript online first and most of our focus was there while building up the supporting tooling. Fortunately, the design of the system makes it pretty easy to add new languages (we've since then done Python and Go). Most of the team is ex-.NET folks and I've been itching to build out the .NET support. Now that we've launched, it will at least become /my/ spare time project.
I'll be really excited when we have it well. Do keep your eye out!
On the other hand, it looks quite interesting and is good to finally be able to know what Pulumi is all about.
Looking forward how it evolves.
In the meantime, as someone who had to learn TypeScript to start using Pulumi when I started here, I do encourage you to look at some of the examples we have and play around with them. It's actually a pretty nice language and with VSCode it's pretty easy to get started.
(Disclosure: I still work at Pulumi)
Pulumi looks like clever technology but the complete detachment from the underlying services and associated costs seems like the perfect storm for an extremely expensive disaster.
How does Pulumi keep track of which services are launched, especially during testing/development, and how does it ensure those are shut down once they are no longer needed? How does it determine the optimal size of instances/volumes/etc to launch?
> How does Pulumi keep track of which services are launched, especially during testing/development
Each Pulumi program is ran within the context of "a stack". The stack is essentially a collection of cloud resources. So when the Pulumi program runs, it will create resources that aren't in the stack, or update existing ones.
So if you create any resources during dev/testing, you just need to `pulumi destroy` those stacks and all of the cloud resources will be reclaimed.
This, IMHO, is one of Pulumi's best features. In that it makes it super-easy to create your own instance of a cloud application. For example, I have my own development instance of app.pulumi.com by just creating my own Pulumi stack and rerunning the same application.
> How does it determine the optimal size of instances/volumes/etc to launch?
It doesn't. The Pulumi program ran determines what resources to create. So you are left to configure, tune or tweak that as makes sense.
Obviously whatever program is actually creating the cloud resources will need credentials to do so. However, they aren't part of the Pulumi program.
When you run `pulumi update` on your machine (or on a CI/CD server) Pulumi will pick up whatever ambient credentials are on the machine. (e.g. ~/.aws/credentials.) So if you to restrict the credentials used to update a particular Pulumi stack, you just need to swap out whatever the current credentials are. (e.g. an AWS_ACCESS_KEY_ID env var.)
> How do you make sure the app only has, say, read access to an S3 bucket it needs to listen to, and can't accidentally delete it? And how does that then allow it to declare the bucket?
There are a lot of good questions there, so let me show you a quick example:
```typescript const imagesBucket = new aws.s3.Bucket( "images", { bucket: "example.com-images", acl: "private", }); ```
This snippet will create a new AWS S3 bucket named "example.com-images". It also sets the default ACL for the bucket to "private". Nothing too surprising there.
If you wanted another resource to have read access to that bucket, you would need to configure AWS to grant access. The Pulumi programming model is about how you declare/describe/create resources, but not actually define policy for how they work. So when using AWS, you would potentially need to create an `aws.iam.Role` / `aws.iam.RolePolicyAttachment` object and hook them up. (Or, if using Azure or GCP, configure access using some other method.)
So in short, to configure what _cloud resources_ can read/write other _cloud resources_, it's a matter of how the cloud resource provider exposes that.
When it comes to matters like preventing you from accidentally deleting the resources when you run `pulumi update` on a program, there are a few features that can help you with that. You can mark a resource as `protected`, so that any update that would delete that resource would produce an error. (Until you update the program again, making that resource as not protected.) Also, the `aws.s3.Bucket` type has a `forceDelete` parameter, that does something very similar. Unless set to true, the Bucket object cannot be deleted. (Thereby preventing some accidental dataloss.)
Does that make sense?
Is Pulumi stateful, then? If you create resources with "pulumi update", change the declarations without updating, and run "pulumi destroy" or whatever, it will only delete the stuff you created in the first step? (That is what I would expect. I would also expect it to support a dry run mode with a diff showing what operations would be executed.) If so, where is this state stored?
> That makes it sound like Pulumi only runs the infrastructure declarations when you run "pulumi update", and that those things don't run when your program runs. That's confusing to me, because your examples (like the thumbnailer) seems to have the program and the declarations in the same file.
This is an optional way to do it, by combining the runtime code and infra code. The runtime code doesn't run when you deploy with "pulumi update," but it is packaged and sent to AWS.
You can also put the runtime code in a different file, as in this example: https://github.com/lindydonna/velocity-examples/tree/master/...
> Is Pulumi stateful, then? If you create resources with "pulumi update", change the declarations without updating, and run "pulumi destroy" or whatever, it will only delete the stuff you created in the first step? (That is what I would expect. I would also expect it to support a dry run mode with a diff showing what operations would be executed.) If so, where is this state stored?
Yes, the state is stored on pulumi.com. The state is list of resource IDs that you provisioned. The Pulumi CLI does indeed have a dry run mode that shows a diff: whenever you run "pulumi update", it first shows a preview.
Yes, Pulumi does mutate resources in place, if the cloud provider supports it. For most resources, it will create a new one (such as a new ECS task), and wait for it to be ready before deleting the old one.
Still, this looks very cool, in that it's a real programming language and not YAML/JSON (which is another of our problems).
Concrete example: Our framework pulls in remote service dependencies via a link to an ELB in order to set remote HTTP endpoint URLs (yes, we know service discovery is a thing, but that's not where we were when we started). Some projects have 15+ dependencies, and it would take literally hours for it to walk the dependency tree. As a workaround, someone built the capability of passing in those dynamic URL endpoints and then the deployments were revised to build the remote URLs via string interpolation. Deployment time dropped to 10 minutes once we walked away from the concept of deploying stacks from the top down.
2nd concrete example: A developer used an incorrect argument during a deployment and deployed a second full stack of his application rather than replacing a single service. (I understand most other tools have diffs/change sets, but this particular developer isn't the sharpest knife in the drawer...) Rather than fix it immediately, he manually fiddled DNS entries and launch configs to create a mishmash stack. Naturally, he didn't tell anyone, so it took weeks (and lots of EC2 $$$) before we found and fixed it all.
I do see some value in an automated full stack deploy with all dependencies, but it should be the exception and not the rule.
Tangentially touching human ego and how it can spoil the intentions. Unless one finds the way to work with it.
Sorry if that's too intense.