Seeing a lot of posts that CI needs a lot of compute, so you can run your CI on X, Y, and Z sandbox platform and cut your bill by 1/3rd. I think you will be surprised how much other things matter too: 1. Good bandwidth to Azure/Microsoft if your code lives in GitHub. 2. Efficient VM snapshotting to do a “sticky disk” type thing. 3. Package caching on the network. 4. Carefully managing oversubscription, or not oversubscribing at all, because compilers often use all the CPUs you can throw at them. 5. A distributed file system for storing caches, because CI is all about caching expensive steps. CI is all about caching, so cheap compute without all of this would still slow your team down. The winning combination would have a fast distributed file system, cheap compute, no oversubscription of resources, vertical scaling to keep the sandboxes small when caches and being restored and scale up quickly when running compilers.

The next few months is going to be a ride.

Love what @arni0x9053 is cooking here!

Coming soon to @tensorlake ! We will manage windows licenses and offer Windows Sandboxes which agents can control remotely and work with Windows apps like Office, Powerpoint and Excel sheets.

Incredible team and opportunity !

You can now vertically scale running sandboxes on @tensorlake. This was one of the most requested features for building coding agents. Sandboxes can now adapt to the workload on the fly, instead of relying on static resource configuration. The live DuckDB workload in the demo scaled up by 7x in throughput when its CPU and memory were increased live. If the machine running the sandbox doesn’t have enough capacity, the sandbox is transparently migrated to a node that can meet the new resource requirements.

OpenAI dots just booked me a DMV appointment, blocked time on the calendar and it found all the personal details from my emails, didn't ask me a thing. Finally feeling the impact of AGI in my life 😂

This is cool!

Active CPU usage based billing makes it possible to use sandboxes at scale for background agents. I was looking into resource usage of some of our users. More often than not, most agents use between < 10% of CPU in aggregate of what users request for. I expect every platform to offer this shortly if they are not doing it already.

It would be interesting if coding agent harnesses defined some sort of "portable session format". Sometimes I am working on a problem on my laptop but I will need to go out, and at that point I want to move the session off to a Tensorlake sandbox. The workaround for me right now is using TLFS, which can be shared between laptops and many sandboxes as a shared file system. I spin up a sandbox with the shared file system and resume the agent session there.

I miss writing code manually. There was a lot of joy in learning a new programming language, learn the APIs of a new database, go through the journey of coming up with a cool idea and seeing it work after a couple of days/weeks of hacking.

Pretty excited to be partnering with @datadoghq! We have a solid roadmap to support all the cool products their teams are working on.

This is a huge issue. Our solution to this problem is a new content addressed file system that can share file chunks across sandboxes on the same machine thereby downloads very little file chunks in most cases. It was inspired by work from AWS Lambda a couple of years back. docs.tensorlake.ai/sandboxes… In most scenarios our new VMM resumes in a few seconds from sleep even if the sandbox snapshot/image is not cached. A provider we know asks their users to keep refreshing their sandboxes every couple of hours so that the images are cached on disk, etc. If you are evaluating a platform ask about p95 and p99 latency of real world sandbox create/resume latency without involving hacks that you would have to pay for.

I was looking at our AWS bill, we will spend 50k+ this month on just bandwidth for TLFS (Tensorlake versioned file system). We are not even pushing the product, it's just organic use from people who find it on the docs and start using it.

Two and a half years ago, I built an open-source project called Indexify during evenings and weekends before starting Tensorlake. It was a compute engine for building a serverless durable workflow engine for data-intensive applications. If you had a million PDFs and wanted to OCR and index all of them, you’d reach for something like this. What remains today is the underlying fast VM scheduler that ran lowe latency functions in MicroVMs. It has been powering Tensorlake’s sandbox platform and runs hundreds of thousands of sandboxes every day. We did Computer SDK's 100k challenge a couple of months back - burst from 0 to 100k sandboxes in 20 seconds on top of this service. As we scaled over the last two months by 20–30x, we built a new scheduler from the ground up to address some of the scaling challenges and requirement to be highly available even during partial failures. It’s unlike anything in the industry: the life cycle of sandboxes and clusters is still represented as replicated state machines, but it’s heavily sharded, with very little global ordering. Instead of building a specialized database using Raft, we now use FoundationDB to store the WAL, and use watches to replicate that WAL across scheduler replicas. Thanks to coding agents, we have been able to build a simulator which injects all kinds of failures into the system to prove the changes we make preserve reliability, speed and correctness. We are going to push this to production next week, we are hoping it would power the next chapter of the company. Long live Indexify!