I am a fan of this concept, I've thought a lot about it actually. Having a proper open source option is awesome and there is no doubt I'll be trying this on some of my home/hobby infra to see what it has to offer. If nothing else, I really think we need more exploration into the space of adding some basic engineering and ops on top of things like this.
Not just for servers and infrastructure, but does it bother anyone else that for example, the Linux desktop has robust APIs for managing powerful service daemons like systemd, but relatively few applications that use them and provide useful UIs/control panels/dashboards? Somehow for many of the tools we DO have, the Linux desktop is stuck in the age of using popen and regex against command output even though we really ought to be able to do better. It's an odd blindspot in a lot of open source work, maybe it is deceptively difficult!
That's mostly nothing to do with this project though, but the two problems are intertwined in my head as problems where a little bit of engineering could really go a long way.
There's limits on either end of that spectrum (pile of bespoke pipes vs. Lotus Notes), that we can see a demonstrated interest in, somewhere in the middle: the success of tools like Docker, and arguably even something as banal as the concept of an OS (distribution). Any user-facing software, etc.
You don't want everyone wasting too much time handrolling slightly-differing monstrosities instead of a cohesive design if it's not a business or engineering differentiator.
I am not really asking for tools that do a lot of jobs, though. A simple program that manages systemd services or provides a nice UI around journald logs seems completely reasonable to me.
That said, I don't take the UNIX philosophy to be gospel, and frankly, very little of the Linux kernel or desktop really seems to. That seems to be closer to what the *BSD derivatives would steer towards. I just want a functional operating system, and the specific ideology or design of packages isn't too big of a concern so as long as the end result seems solid.
I thought the philosophy was to just build tools you want. So we have, eg systemd and OpenRC. Where one is many-tools and the other is composed of many-small-tools.
The best part about windmill is the team behind it, join the discord server to see how fast they fix issues and add features. The blog is also top notch: https://www.windmill.dev/blog
Hello, founder here. Thank you for submitting Windmill. We're really close from hitting v2 (adding multiplayer with yjs is our last milestone) and would have waited a bit before submitting but here we go.
We are fully open-source (AGPLv3) with just one ee plugin for syncing the cache of the workers at large scale (more on that below). You can deploy it on a small ec2 instance with our docker-compose or on very large clusters with our helm charts: https://docs.windmill.dev/docs/advanced/self_host
We are an onion of 3 layers, which are usually separated verticals and the closest services we replace in each vertical are: Lambda + n8n/temporal + Retool:
- Workers implemented in Rust that can run any scripts from source in python/typescript(deno)/go/bash. They are extremely efficient, with most typescript scripts being able to run in 40ms e2e. This can be seen as a self-hosted lambda except it works very differently. Lambdas spawn microvms using firecracker and launch http servers within those micro-vms. We run the scripts bare (with isolation using nsjail on our multi-tenant instance) but cache very aggressively the dependencies. For instance, in python we analyze the dependencies using the AST: https://github.com/windmill-labs/windmill/blob/main/backend/..., infer the dependencies from there, have them go through pip-compile to generate a lockfile (and cache the result), then create a pathset similar to how virtualenv behave with the version of each dependency locked. For typescript, we leverage the very well made immutable deno cache. The cache works well but if you scale your cluster to more than 10 workers, the likelihood of having your worker see dependencies for the first time increase, which is why we reimplemented a pip scheme backed by s3 that is extremely fast (we do not compress, just tar individual wheels and s3 is very fast on local networks). For all the supported languages, we parse the signatures of the main function to infer the corresponding json schema of the payload to trigger them, and from that jsonschema we can then generate automatically the UI and infer the input shape required in the flows for that particular step.
- A workflow engine re-implemented from scratch which can be seen as a distributed FSM and whose state is stored in postgresql and every transition done through postgres transactions. Every worker is picking a job and progressing the fsm at completion. The full spec is at: https://docs.windmill.dev/docs/openflow. We support branches, for loop, suspend/sleep, approval steps, sharing the same folder between steps and everything you would expect from a temporal like engine (retries, error handlers).
- A dashboard/app builder that is very similar to Retool but much faster because it's implemented in Svelte and which can also run python scripts directly leveraging the 1st layer. We also support full react apps: https://github.com/windmill-labs/windmill-react-template
We have a CLI that allows you to sync from github directly: https://github.com/windmill-labs/windmill-sync-example or your local filesystem. Our scripts are absolutely normal scripts so there is no lock-in, and they are executable locally https://docs.windmill.dev/docs/advanced/local_development so you do not have to use our webeditor if you do not want to but our editor supports pyright and deno lsp through our dedicated backend communicating to monaco using websockets.
We also support worker groups to run some jobs on hardware accelerated machines (GPU) and a lot of other features (oauth syncing of the resources, schedules, granular permissioning, groups and folders).
Every script/flow deployed has a dedicated webhook endpoint to run synchronously and asynchronously. Scripts are never overwritten, they keep a lineage of hash for each version and have per-hash dedicated webhook to ensure their behavior never change.
We also have a hub for sharing re-usable scripts https://hub.windmill.dev to be used in flows to have some of the same convenience you can find in Pipedream or Zapier.
Our closest alternative would be Airplane.dev, but they're not open-source, not self-hostable air-gapped, less performant, less-featured and much more expensive. You can self-host us fully free.
We are used in production at scale in clusters of 50 nodes+ and thanks to the reliability of Rust and Postgresql have had no downtime to report.
I'm a solo founder, YC S22. The MVP got some attention at the very beginning: https://news.ycombinator.com/item?id=31272793 but was very bare. Now we're ready to bring this to the next level and enable all engineers to spend more time building and less time reinventing the wheel thanks to it :)
One nitpick: In my humble opinon, you should switch from Discord chat to Zulip.
It's completely open-source (Apache 2, self-hostable, free full cloud plan for open-source projects) and has an excellent threads-first (called "topics" model): https://zulip.com/why-zulip/
It can run fully in the browser, but it has desktop and mobile clients as well.
It also supports public channels which AFAIK discord doesn't. So you can link to specific messages inside a GitHub issue and anyone who doesn't even have an account can see the relevant messages and context.
I'm very tempted and agree that Zulip sounds much better than Discord. The only thing I'm a little bit worried about is that I have noticed that most people hate having multiple chat clients open and that discord seem to be the de factor standards for open source projects. I apologize to be contributing to the network effect and will strongly consider switching to Zulip.
I would suggest against it, Discord is the de facto standard, especially with the younger crowd, millennials and Gen Z. Everyone already has it installed and everyone knows exactly how it works, and it also "just works". I've never even heard of Zulip and I hang out on HN almost every day.
Just because everyone knows how it works doesn't mean alternatives shouldn't be tried, especially when they're superior in the aspects that matter for the audience you're interested in.
Besides, how do you think a competing product should rise then?
I'm not trying to start an argument here, but I consider less friction and more familiarity to be superior for this use case.
I was trying to provide a contrary opinion because I often see the HN hivemind hate on Discord heavily for no reason, while it's become a ubiquitous piece of software used by younger people that the older generations have been completely sleeping on and still think it's some niche version of IRC for gamers. Discord has hundreds of millions of users, it's a behemoth rivaling the big social networks.
As a random user, if I were trying to ask a question about this piece of software, I want to open the Discord server and ask it, I don't want to start researching whatever Zulip is and how to use it and figuring out how a new interface works
It it against the Discord Terms of Service to extend the accessibility of the client past what is available in web standards (and what of that they implement). The regularly ban accounts of people who try to make it more accessible (and often of the users too).
These might be good qualities for a gamer guild or YouTuber community, but it's a real barrier to participating in open source software communities.
Zulip is open, and doesn't have these restrictions (and already had many bridges and accessibility integrations). I participated in the Oil Zulip, and it was less of a hassle tHan to join a new Discord guild and navigating their adhoc rules bots.
Discourse[0] is a perfect fit for Windmill, IMO. Maybe I'm biased because I don't like Discord for this kind of project. Let the search engines leverage your project.
As a sidenote: is the windmill hub open-source too? Does it make use of the base product to operate? I think it might be a good candidate for showing what your product is capable of, as an actual example of use in production
Windmill Hub is not open-source, not because we want to keep it private but because I wrote it a bit quickly using svelte kit and it's not ready for the world to see yet. It does not use windmill to operate but every actions trigger a windmill webhook and alert the discord.
Hey Ruben! The product looks pretty interesting and I love the underlying technical choices.
One suggestion: I think you could do improve the docs by (a) providing a direct comparison to Airplane and (b) providing more code-oriented documentation.
Most of you docs highlight the web editor which, frankly, I have little interest in. I want to write my scripts and views in my IDE and store them in GitHub. This is something Airplane really nails: their docs are clearly about how to use this as a developer. Compare your flows quick start [0] with their tasks quick start [1] and it really feels like Airplane is more suited to a developer audience. I'm worried I will be locked into static control flow with yours, whereas Airplane makes it clear I can do ~any code I want, as long as it runs in Temporal.
Of course, totally fair if you are trying to target more no-code use cases.
You're absolutely right. We're pretty weak in docs for developers that only want to use the CLI and the reason why I wanted to wait for the V2 is that we wanted to do a full cycle on showing the potential of the CLI for local dev and airplane-like kind of development.
For the temporal thing, that's absolutely fair and completely understand the worry but since most workflows are about the steps and not the glue between them, you will be able to reuse most of your code. However, temporal is quite an heavy format for writing workflows so of course it will still be a bit of work.
Our views support our retool like builder but also full react apps, or svelte, or vue, or anything that can compile to an iife. The example provided above can be iterated on with just a simple 'npm run dev': https://github.com/windmill-labs/windmill-react-template
I love open-source and wanted it to be open-source but given that's our whole product and not just a part of a bigger ecosystem, and given that we intend to commercialize mostly the large enterprise audience, we needed to have a reason for the legal departments of those enterprises to push to buy our commercial license. So otherwise said, we do not want to employ dark patterns and want it to be as easy as possible to self-host but also need to be sustainable as a company. Most features that you would expect to be premium plugins are baked-in and open-source (sso, audit logs, smart autocompletion on the webeditor, etc etc)
> In addition, a commercial license grants you a dedicated engineer to transition your current infrastructure to Windmill, support with tight SLA, audit logs export features, SSO, unlimited users creation, advanced permission managing features such as groups and the ability to create more than one workspace.
From the README. It sounds like a commercial license is needed for SSO and that the free version has a limited number of users.
Sorry that's obsolete, we will remove it from the Readme!
The updated paragraph:
> In addition, a commercial license grants you a dedicated engineer to transition
your current infrastructure to Windmill, support with tight SLA, and our global
cache sync for high-performance/no dependency cache miss of cluster from 10+
nodes to 200+ nodes.
Thinking about it as an alternative to scripted workflows with Rundeck, it seems like the main gap here is getting scripts to run on a specific machine without needing to have set that machine up to run worker processes.
(I could be misunderstanding here, because I haven't spotted anything in the docs that talks about the minimum deployment necessary just to allow a particular machine to run workers.)
Really simple example: there's an administrative command on a mailserver, which can be used to create a new mailbox. I want to make that operation part of some workflow. With Rundeck, I can script this and I just have to make sure Rundeck has ssh credentials onto the mailserver; at execution time it'll deploy the script, substitute in whatever variables, run it and capture the output.
I'm not clear on how I would achieve that with Windmill. Would it involve installing Windmill (or some component of it) on the mailserver, and setting it up as an individual worker group?
> And you would do exactly what you suggest in the last paragraph
Assuming you mean this:
> Would it involve installing Windmill (or some component of it) on the mailserver, and setting it up as an individual worker group?
I was looking at Windmill as a replacement for Rundeck as well and this is exactly the point I was trying to figure out too.
I get why you need to install the worker. And there are lots of tools out there to automate this, from putting it in a custom AMI to tools like chef to building it as part of your base container image.
But one thing that's really nice about Rundeck (maybe the only thing) is that you can have it connect to any host and run commands with just some SSH credentials.
It also feels like worker groups maybe work for this use case but aren't the actual intention. Worker groups feel like "this script does some work and I want to make sure it runs on a machine with the resources that it needs to support its work". Whereas Rundeck scripts on remote hosts are more about "I want to run this command on this specific host because I'm doing maintenance on this host."
I guess you could have a script keep a long lived ssh deamon and do this
You could also use the inline script endpoints to ask the worker to execute bash commands directly. But this probably does not compare well to rundeck. If you had time, I'd be very interested to chat and understand what is the optimal experience you'd be looking for: ruben@windmill.dev
Thanks for the reply. Winmill is definitely something that I'll need to take a closer look at, when time permits. I do suspect though that it's going to be a bit too heavyweight to be a direct replacement for Rundeck.
My first questions with any new tool of this kind is immediately:
- Can I run/validate branch versions of the scripts non-destructively, for example as part of pre-merge checks in an IaC CD workflow?
- Can I run my scripts locally without a lot of drama, while developing?
If the answer to either of these questions is no, then the tool is the moral equivalent of directly editing PHP files on the prod machine, unless you're committed to deploying an entire second instance of the tool for staging/dev purposes.
1) since you can run them locally, you can also run tests as part of your CI. You can also deploy a full windmill as part of your CI, deploy to that local instance, run some e2e tests by calling the webhooks, and then only deploy once those work.
Note that we have a full drafting system baked in into the webeditor as well: https://docs.windmill.dev/docs/core_concepts/draft_and_deplo... and have support for folders with different permissions and you can create workspaces so you may implement dev/staging/prod in any fashion you like (different folder, workspace, instances)
You would run your CI in github like you would for normal codebase, the draft aspect is something that is only for the frontend editor. Every time you push from the the CLI (which you could do upon merge to main), it is akin to do a deployment of the new versions.
I have been following this space for really long time. I think closest competitor is Airplane.dev [1]. One of the most difficult thing to sell as a sales person is to sell an open ended workflow and UI engine and ask me to segment and prospect customers. It is not that you cant sell it, it is just too broad even for your customers to figure out a usecase, because theres always a specific tool to solve the usecase. For instance, stripe refund is one most common usecase posted by all of these platform, but there are actual tools that just does stripe refund and they solve it in depth. You can argue these tools are value for money but then if it involves build time, in long run it becomes expensive. Retool is actually captured the market really well. It is basically become the verb in the internal tool market, it is hard to compete with that even as an opensource project.
Its hard to predict how a good open source tool that solves hard and painful problems will evolve in a marketplace. I don't think Linux developers would have remotely dreamt of a cloud revolution when they were creating an open source kernel. Maybe windmill will enable a marketplace of automation tools that can be leveraged across multiple business for multiple usecases. Never underestimate a open source solution solving painful problems.
You are argument is fair but being a VC backed company puts a deadline on the org. They have to not only prove that it solves a pain point but also make a ton of money and growth.
Was hoping it allowed my runtime. I have a bunch of scripts in NodeJS, Python 2, C#, etc. I don't care to rewrite them and/or write scripts tied to a weird sandbox runtime.
We could but decided against it because we have no way to run those efficiently given our architecture and prefer to support a few languages and ensure we can run them blazing fast than all languages but with no guarantees of performance. In windmill, you have the guarantee to have no cold-start thanks to leveraging the great design of the deno runtime. For python, we had to do a LOT of engineering to make it work well and did not have the bandwidth to do the same for all languages. Ultimately, we could decide that some steps could be slow and it doesn't matter but it wasn't requested enough to prioritize it.
To be clear, the workers can run in docker, but do not run docker jobs.
Looks like you're a new y combinator startup. Kool. I would suggest you find away to allow existing code to be reused. Deno market share is very small and experienced/larger teams are not going to want to rewrite and learn that stack. They would have to depend on you, a small new company. Too much risk.
It depends on how efficient you need to be. There is the option to just use "shell" scripts where you run the command. I haven't done much with that, but it's there.
airplane.dev is a commercial offering very similar
in spirit; it is not free but not expensive. It has native support for node, python, SQL, shell scripts and a couple other options as well as docker images for everything else. It’s been a game changer for me.
This looks cool as a concept, but I've already got script-server [0] and it works pretty great as a no/low code solution for UI-ifying shell scripts.
What does windmill do that script-server can't? Or is windmill targeting a completely different area, and is thus irrelevant to me and my pile of 500+ python, shell, Golang, and java bin dir.
We do automatic parsing of the parameters in typescript/python/go/bash, we infer your dependencies automatically.
And we have workflows and a retool alternative/app builder baked-in. And granular permissions and a lot of other things... The UI of script-server looks a bit bare.
But like with a lot of those tools, if you're happy with script-server then that's probably the most important!
We were just looking for something similar to this to run workflows at $JOB and decided to build our own in the end, because nothing we found fulfilled our requirements.
A few questions, just in case:
1. Does Windmill support long-running workflows? Our use case has our customers being in a workflow for years (as long as they have a relationship with us) and events moving them around states.
2. Does it support rule-based state transitions? Everything I've seen is "when this webhook comes, go to that state", but we want something that will trigger the evaluation of a set of rules (with an event or webhook), and the rules will contact APIs to judge whether everything is satisfied for a transition to some next state.
3. Does it support synchronous calls? We usually want the customer to ask "OK where am I now?" and the system to trigger evaluation, transition them to the state they should be in, and tell them where they are. This is for use cases like the website asking "the user has just submitted the data we wanted, where should I redirect them now?".
1. Yes there are no time limits for workflows and there is no cost to keep a workflow suspended. It will just be scheduled to be reactivated in the future
2. The webhook can contain a payload, and you can but the suspend step into a for-loop. You could literally say, for-loop forever the suspend until the result of the suspend step (which can be the result of the evaluation of the webhook payload by your rule is true). For that, you will need to use another primitive that we expose which is "Break" that will break out of the for-loop
3. I'm a bit unsure of the term synchronous call. We have a long list of APIs https://app.windmill.dev/openapi.html#/ and you fetch the status of the flow or trigger a flow in a synchronous http webhook to wait for the result. As to force the state of the workflow. We do not expose something like that but it could be added easily.
Happy to chat if we could be a good fit for your needs: ruben@windmill.dev
Thank you! It doesn't sound like an exact fit but Windmill looks great in general, well done!
For the second point, the idea is that we don't want to rely on webhooks at all, except as evaluation triggers. The rules are built in to each state, so even if we've missed a webhook/event, the customer can still proceed.
For an example, one of our flows requires the user to have filled out a consent form in the last three months. As I understand it, with your design, we need to send a webhook when a user has given consent, and then we can proceed to the next state, so we'd have to somehow check for previous consent in some different state? With our design, when the user reaches the state, the rules will be evaluated, and "has given consent within the last three months" will be true, and the user will proceed to the next state without ever seeing the consent form.
For the synchronous bit, it kind of follows from the above feature, and it is so we evaluate state transitions while the user is waiting. We run the workflow synchronously, in the request (since no workflow needs external events to continue), and return the new state to the user.
Since this is synchronous, it avoids races or long waits for events to be delivered.
Thanks for listening, maybe you'll find some of the above helpful in your design!
Makes sense. I understand the synchronous part now. I think there are ways to achieve the same benefits with an asynchronous system but if what you have works, no point changing it!
I'm confused as to how this compares to tools in the workflow manager space such as Airflow/Perfect/Dragster. The words describing the software are almost the same but it seems to come from alternate universe of tools I'm not familiar with. Could someone fill me in?
It's similar to airflow/prefect/dragster in that it can run workflows, and those workflows can do ETL and ops and can run arbitrary code.
It's different in those fashions. Windmill can run python but also typescript, go and bash. Airflow/Prefect etc uses code annotation which mean the code is actually specific to the platform. We actually have the steps be directly scripts/code with a main function, and then have a low-code builder to compose them into powerful workflows. That low-code builder produces a yaml like that one: https://github.com/windmill-labs/windmill-sync-example/blob/... which you could theoretically write by hand.
The low-code builder is intuitive and similar to n8n except it has the feature-set of the above framework. It's easy to test the individual steps or to test the whole flow. The workflow engine is very efficient and running 50 steps will essentially have about 1s overhead total, and the steps are executed bare so it will take as much time as if you were running them without windmill on your node directly. They can be hardware accelerated easily in that fashion. We also support some features borrowed from temporal like the ability to suspend at no-cost flows until they are resumed by an external event such as an approval from slack.
I like to think that we target hybrid teams made of low-code/semi-technical people AND engineers that expect to have the same characteristics seen in temporal/airflow. Abstraction without compromise, or zero-cost abstractions as the rust community like to say
How does it compare to something like apache airflow? I like the native Go integration, and that it is specifically mentioned that the overhead is minimal to start a job. Airflow can take quite a varying amount of time to schedule your job.
Passing information between steps is a pain in Airflow (https://docs.astronomer.io/learn/airflow-passing-data-betwee...) whereas it's quite trivial in windmill, the performances/latency is better, and it supports typescript. Overall, the biggest improvement is the support for typescript/go/bash and the low-code builder that will allow you to build extremely complex flows that would be hard to visualize with just code annotations but with exactly the same featureset and more (retries, error handlers, suspend/sleep). We also make it really easy to test individual steps and the whole flow since web apps allow you to build advanced UX fit for big graphs. But Airflow is a great framework that we took a lot of inspiration from.
That's because you're looking at the TaskFlow syntax and a trivial example of passing all the results from a Task A as all the inputs of a Task B that follows it immediately. TaskFlow works mostly in trivial case. The true syntax for it is in the tab "Traditional syntax". For a complex graph where you're passing data from step A to a node that is far down the tree and deeply nested and you will see that it's quite cumbersome. Even in their trivial example:
res = requests.get(url)
ti.xcom_push(key="cat_fact", value=json.loads(res.text)["fact"])
...
cat_fact = ti.xcom_pull(key="cat_fact", task_ids="get_a_cat_fact")
in this interview Ruben, the founder of Windmill, talks about his background, shares how the project got started (bootstrapped solo founder, went through YC last summer) and encourages more software engineers to start their own companies
OP here... Oh dang this blew up. Found this project while trying to find a low-code / PaaS for my home infrastructure. Found this project and decided to share. Pretty easy to selfhost as I deployed it on a Alpine Linux server using Podman + Docker Compose binary. Really nice piece of software.
No, just don't modify the software itself without making your changes open-sources and you will be good to go. Having scripts on the platform is not modifying the software.
Without looking in detail, I've had some experience with internal business process/app development, and I always thought Temporal would be a great technology to build a platform on. Great choice.
Can this be used to host external services? Is this only for faas that are inaccessible to the outside world? I'm curious if you could use this to build web hook tools specifically.
Webhooks require to be authentified through tokens passed as bearer tokens or query args. This is because we have a tight permissioning system around visibility of scripts/variables/resources/flows, etc.... But nothing prevent you to add a proxy on top that call every webhook with a sufficiently permissioned token.
We provide a generous free-tier of 1000 executions a month and as a very small team I'd rather rely on github/google bot abuse prevention than reinvent the wheel there. Email verification is not strong enough.
it should be trivial to push that to a self-hosted setup if you need it.
I have not tried to take my entire workspace from self-hosted to cloud. But I have been exporting and importing individual scripts/flows/apps from self-hosted to cloud.
I actually got codeium to work on windmill today trough their chrome extension.
the extension lets you whitelist sites to work on. They do not offer a chat interface yet for the chrome extensions but they do have it on other integrations they provide. So I would not be surprised if they got it on the chrome extension also, at least if it turns out to be popular.
Yes, we expose webhooks for all scripts and flows so you can integrate it for anything that can send webhooks, including Gitlab, and our workflows can be used as CI/CD.
I get what they mean -- it has definitely gained a foothold in the tool/utility world thanks to easy cross-compilation and single-binary output. Given their target, I definitely understand why they would put it in that group.
Not just for servers and infrastructure, but does it bother anyone else that for example, the Linux desktop has robust APIs for managing powerful service daemons like systemd, but relatively few applications that use them and provide useful UIs/control panels/dashboards? Somehow for many of the tools we DO have, the Linux desktop is stuck in the age of using popen and regex against command output even though we really ought to be able to do better. It's an odd blindspot in a lot of open source work, maybe it is deceptively difficult!
That's mostly nothing to do with this project though, but the two problems are intertwined in my head as problems where a little bit of engineering could really go a long way.