- One tool to ingest data
- Another one to transform it
- If you wanted to run Python, set up an orchestrator
- If you need to check the data, a data quality tool
Let alone this being hard to set up and taking time, it is also pretty high-maintenance. I had to do a lot of infra work, and while this being billable hours for me I didn’t enjoy the work at all. For some parts of it, there were nice solutions like dbt, but in the end for an end-to-end workflow, it didn’t work. That’s why I decided to build an end-to-end solution that could take care of data ingestion, transformation, and Python stuff. Initially, it was just for our own usage, but in the end, we thought this could be a useful tool for everyone.
In its core, Bruin is a data framework that consists of a CLI application written in Golang, and a VS Code extension that supports it with a local UI.
Bruin supports quite a few stuff:
- Data ingestion using ingestr (https://github.com/bruin-data/ingestr)
- Data transformation in SQL & Python, similar to dbt
- Python env management using uv
- Built-in data quality checks
- Secrets management
- Query validation & SQL parsing
- Built-in templates for common scenarios, e.g. Shopify, Notion, Gorgias, BigQuery, etc
This means that you can write end-to-end pipelines within the same framework and get it running with a single command. You can run it on your own computer, on GitHub Actions, or in an EC2 instance somewhere. Using the templates, you can also have ready-to-go pipelines with modeled data for your data warehouse in seconds.
It includes an open-source VS Code extension as well, which allows working with the data pipelines locally, in a more visual way. The resulting changes are all in code, which means everything is version-controlled regardless, it just adds a nice layer.
Bruin can run SQL, Python, and data ingestion workflows, as well as quality checks. For Python stuff, we use the awesome (and it really is awesome!) uv under the hood, install dependencies in an isolated environment, and install and manage the Python versions locally, all in a cross-platform way. Then in order to manage data uploads to the data warehouse, it uses dlt under the hood to upload the data to the destination. It also uses Arrow’s memory-mapped files to easily access the data between the processes before uploading them to the destination.
We went with Golang because of its speed and strong concurrency primitives, but more importantly, I knew Go better than the other languages available to me and I enjoy writing Go, so there’s also that.
We had a small pool of beta testers for quite some time and I am really excited to launch Bruin CLI to the rest of the world and get feedback from you all. I know it is not often to build data tooling in Go but I believe we found ourselves in a nice spot in terms of features, speed, and stability.
https://github.com/bruin-data/bruin
I’d love to hear your feedback and learn more about how we can make data pipelines easier and better to work with, looking forward to your thoughts!
Best, Burak
As I read through the documentation, Do you have a mode in ingstr that lets you specify the maximum lateness of a file? (For late-arriving rows or files or backfills) I didn't see it in my brief read through.
https://bruin-data.github.io/bruin/assets/ingestr.html
Reminds me a bit of Benthos / Bento / RedPanda Connect (in a good way)
Interested to kick the tires on this (compared to, say, Python dlt)
as per the lateness: ingestr itself does the fetching itself, which means the moment you run it it will ingest the data right away, which means there's no latency there. in terms of loading files from S3 as an example, you can already define your own blob pattern, which would allow you to ingest only certain files that fit into your lateness criteria, would this fit?
in addition, we will implement the concept of a "sensor", which will allow you to wait until a certain condition is met, e.g. a table/file exists, or a certain query returns true, and continue the pipeline from there, which could also help your usecase.
feel free to join our slack community, happy to dig deeper into this and see what we can implement there.
Using dbt at $JOB, and building a custom dbt adapter for our legacy data repos, I've slowly developed a difficult relationship dbt's internals and externals. Struggling with the way it (python) handles concurrency, threading, timeouts with long running (4hr+ jobs), and the like. Not to mention inconsistencies with the way it handles Jinja in config files vs SQL files. Also it's lack of ingestion handling and VSCode/editor support, which it seems like Bruin considers very well! Since starting poking around on the inside of dbt I've felt like Go or Rust would be a far more suitable platform for a pipeline building tool, and this looks to be going in a great direction, so congrats on the launch and best of luck with your cloud offering.
That being said, I tried starting the example bruin pipeline with duckdb on a current data project, and I'm having no luck getting the connection to appear with `bruin connections list` so nothing will run. So looks like I'm going to have to stick with dbt for now. Might be worth adding some more documentation around the .bruin.yml file; dbt has great documentation listing the purpose and layout of each file in the folder which is very helpful when trying to set things up.
Your point on .bruin.yml documentation is spot on, and we’ll make improving that a priority. If you’re still running into issues, please don’t hesitate to reach out—I’d be happy to help debug this with you directly. Thanks again for giving Bruin a try!
I am sorry to hear that it didn't work, we do have a dedicated page for duckdb specifically here: https://bruin-data.github.io/bruin/platforms/duckdb.html
Would this help with it? I'd love to see how we can improve if you'd like to share your thoughts on that. Please feel free to join our slack community as well, we can talk directly there too.
I love the idea, effectively allowing going towards a direction where the right platform for the right job is used, and it is very much in line with where we are taking things towards. Another interesting project in that spirit is sqlframe: https://github.com/eakmanrq/sqlframe
I also noticed this pattern where library authors sometimes do a bit extra in terms of discussing and even promoting their competitors, and it makes me trust them more. A “heres why ours is better and everyone else sucks …” section always comes across as the infomercial character who is having quite a hard time peeling an apple to the point you wonder if this the first time they’ve used hands.
One thing wish for is a tool that’s essentially just Celery that doesn’t require a message broker (and can just use a database), and which is supported on Windows. There’s always a handful of edge cases where we’re pulling data from an old 32-bit system on Windows. And basically every system has some not-quite-ergonomic workaround that’s as much work as if you’d just built it yourself.
It seems like it’s just sending a JSON message over a queue or HTTP API and the worker receives it and runs the task. Maybe it’s way harder than I’m envisioning (but I don’t think so because I’ve already written most of it).
I guess that’s one thing I’m not clear on with Bruin, can I run workers if different physical locations and have them carry out the tasks in the right order? Or is this more of a centralized thing (meaning even if its K8s or Dask or Ray, those are all run in a cluster which happens to be distributed, but they’re all machines sitting in the same subnet, which isn’t the definition of a “distributed task” I’m going for.
I like the comparison page in Hamilton, and in their examples they operate in the asset level, whereas Bruin crosses the asset level into the orchestrator level as well, effectively bridging the gap there. What Bruin does is beyond a single asset that might be a group of functions, it is basically being able to build and run pipelines of that.
In terms of distributed execution, it is in our roadmap to support running distributed workloads as simple as possible, and Postgres as a pluggable queue backend is one of the options as well. Currently, Bruin is meant as a single-node CLI tool that will do the orchestration and the execution within the same machine.
If you’re doing data analytics in Python it’s well worth a look.
Does Bruin support specifying and visualizing DAGs? I didn't see that in the documentation via a quick look, but I thought to ask because you may use different terminology that can be a substitute.
Do you mean like Airflow or Pachyderm? I am also very interested in new tooling in this space that has these features.
I did look into CUE in the very early days of Bruin but ended up going with a more YAML-based configuration due to its support. I am not familiar with their flow package specifically, but I'll definitely take a deeper look. From a quick look, it seems like it could have replaced some of the orchestration code in Bruin to a certain extent.
One of the challenges, maybe specific to the data world, is that the userbase is familiar with a certain set of tools and patterns, such as SQL and Python, therefore introducing even a small variance into the mix is often adding friction, this was one of the reasons we didn't go with CUE at the time. I should definitely take another look though. thanks!
I’m not (necessarily) motivated to switch tooling because of the language it is written in. I’m motivated to switch tooling if it has better ergonomics, performance, or features.
I agree with you 100% on the language part, I think it is an interesting detail for a data tool to be built in Go, but we have a lot more than that, a couple of things we do there is:
- everything is local-first: native Python support, local VS Code extension, isolated local environments, etc
- very quick iteration speed: rendered queries, backfills, all running locally
- support for data ingestion, transformation, and quality, without leaving the framework, while also having the ability to extend it with Python
these are some of the improvements we focused on bringing into the workflows, I hope this explains our thinking a bit more.
I really want to know how this is going to benefit me before I start putting in a lot of effort to switch to using it. That means I need to see why it is better than ${EXISTING_TOOL}.
I also need to know that it is actually compatible with my existing data pipeline. For example, we have many single tenant databases that are replicated to a central warehouse. During replication, we have to attach source information to the records to distinguish them and for RBAC. It looks like I can do this with Bruin but the documentation doesn't explicitly talk about single tenant vs multi-tenant design.
I guess a more comparable alternative would be Meltano + dbt + Great Expectations + Airflow (for Python stuff), whereas Bruin does all of them at once. In that sense, Bruin's alternative would be a stack rather than a single product.
Does that make sense?
Bruin is effectively going a layer above individual assets, and instead takes a declarative approach to the full pipeline, which could contain assets that are using Ray internally. In the end, think of Bruin as a full pipeline/orchestrator, which would contain one or more assets using various other technologies.
I hope this makes sense.