Cloud infrastructure should be fun! Co-founder at burla dot dev
Boston, MA
Joined March 2020
- Tweets107
- Following131
- Followers28
- Likes97
"The reason you got that wall of text is that I wrote it anyway."
- sonnet 5.5
Fixing bullshit wordslop is more valuable than more intelligence imo.
It feels like every single dev tool is converging on the same value prop:
"make your agent is X% faster and cheaper"
Probably with some visibility layer involved.
Burla has an all new website!
Let me know what you think.
burla.dev
I feel an urgent need to thank my coding agents when they do a good job but don't want to waste tokens.
I had a dream I had infinite claude tokens, is was so nice 😭
If i ever make it i'm starting a token charity.
I think you can summarize all workflow durability as "two VMs crashing is less likely than one".
This means that if you can easily abstract hardware, you can easily create durable workflows.
yes, clearly you need resumable checkpoints.
AI writes my code now and simply has a rule to build resumable checkpoints everywhere.
No framework is necessary for this.
And therefore:
- Any framework that has "tasks" is unnecessary.
- The only framework you need to build durable workflows is a hardware abstraction framework.
Burla is:
* the simplest
* the fastest
* the most compute efficient
of any distributed computing framework.
I think it might be the most scalable too but I haven't benchmarked other frameworks.
Does anyone know how big the biggest single Ray cluster can be?
Burla is a self-hostable compute platform that vertically scales hardware available to each task live while the program is running.
This frequently more than doubles compute efficiency, and eliminates out of memory errors!
Here's how it works:
docs.burla.dev/blog/dynamic-…
The answer is automatic vertical hardware scaling.
When memory pressure goes up, the available memory should too!
The best part is you can do this today for free self-hosted in your cloud with Burla.
Here's how our system works:
docs.burla.dev/blog/dynamic-…
Currently we average <20% utilization across all actively rented cloud compute.
Entering an era of compute constraints, efficiency will be more important than ever.
This is why we've been so focused on dynamic hardware. You can achieve >90% utilization automatically!
People have no idea that scaling Python to 1000s of VMs in their private cloud is so trivial a total beginner can do it now.
Just got off a call with a bioinformatician who was struggling to scale past one big VM.
We got their pipeline running on 100VMs within the 30min call!
It only takes 2 commands to install our OSS parallel computing framework in your GCP environment:
- `pip install burla`
- `burla install`
Then you can scale any Python function to 1000s of VMs using one line of code: burla.dev
Not only is this much easier,
(no more guessing, no more memory errors)
It's often more than twice as resource efficient, on what are often some of the most expensive jobs happening inside a company (big data processing)
Full technical breakdown in the blog post:
docs.burla.dev/blog/dynamic-…
And if you're sick of dealing with with complex configs, memory errors, and inefficient frameworks please send me a DM!