ozkatz

costs?

1PB of EBS volumes on sc1-class HDD is roughly $15k/month. Let's say you need at least 2 copies of each byte for cross-AZ availability and durability, that's 30k per PB for JUST the drives.

These disks also need compute attached to them, coupling their scaling needs. Now you have a fleet of (expensive) VMs, regardless of how many queries end up hitting that database.

Of course, those VMs have to run pretty complex software such as PlanetScale, which has to take care of replication, backups, concurrency control, sharding, indexing, vaccuuming, and a whole bunch of other things that add operational complexity and require expertise. That's the fixed costs of an OLTP database. A lot of that cost is derived from the need to, well, process transactions.

Thing is, they already have a data lake. So the fixed cost to store a PB of data on S3 is already paid. They mention the index is roughly 3 orders of magnitude smaller than the data set (1PB of data is roughly 1TB of index). With a marginally less complex compute stack, I believe this becomes cost effective pretty quickly.

Curious if you’ve had a chance to try lakeFS?

It was designed with large-scale environments in mind. I’m aware of several deployments managing hundreds of petabytes of data and billions of objects, which is why lakeFS does not use Git’s Merkle tree / directory tree approach.

Disclaimer: I’m one of the project’s co-creators.

That’s the current RBAC implementation: agents by default can make no API calls. the only way for them to contact the outside world is through a forward proxy configured in the sandbox. that proxy only allows making requests to destinations explicitly allowed (host, path, method)

Appreciate the honest feedback. I agree there's a lot to improve there.

very very fast: proportional to the count of objects modified, but not their size. Every commit represents a snapshot - an immutable listing of objects that represents the repository. reverting is essentially applying the inverse of the diff introduced by the reverted commit.

This is metadata only as the objects themselves are immutable.

S3 offers versioning at the single file level.

Imagine an agent dropping a directory with 1m images in it. just figuring out what happened and what got dropped, restoring it one by one, etc. - doable, but ergonomics are a bit lacking.

Regarding pricing - that's indeed a great question and we don't have an answer yet. It will very likely be based on consumption and should be competitive to similar solutions.

Atomic commits are based on snapshotting done by lakeFS under the hood. Each sandbox run produces a new atomic commit to a hidden "main" branch. Updating that branch is optimistically concurrent, with lakeFS checking for conflicts - multiple writers updating the same object.

It uses lakeFS under the hood, so the unit of conflict would be a single file (object, under the hood). Resolving conflicts requires "picking" a winning side, or rerunning a conflicting job. Would you see a use case for merging changes into the same file? Interested to hear about your use case!

It provides a filesystem abstraction, which agents are really good at interacting with. Because it's just a POSIX filesystem - you can put a sqlite database directly on it and get those same transactional capabilities for that too.

it should absolutely scale to that. The filesystem is backed by lakeFS, where every sandbox automatically branches out, and mounts that branch. so you get isolation from lakeFS and the scale of an underlying object store (S3, in Tilde).

If that database is stored on the transactional filesystem available to the sandboxes, yes! Instead of backing up, it utilizes an efficient snapshot mechanism (lakeFS under the hood).

Part of the appeal (subjective, I know) of versioning is stuff like human-in-the-loop approvals. Think of a pull request: a change is requested by an agent, a human approves, changes get merged atomically. Even if other changes were applied since creation.

Hey! It doesn't necessarily have to be "data heavy", but any form of state (from code to binary files) that an agent might use for automation.

Agents are really good at interacting with files and directories (text in, text out!). This adds a layer for those that allows managing that state in a transactional, versioned way.

Oh VMS, How I miss thee

In a perfect world, every system and external API would expose a standardized interface for versioning its own immutable state, so you'd be able to rollback and time travel across multiple such systems.

Not sure what else we can do in this world other than tightly control outbound requests and provide enough visibility into those requests for a human|agent to try and undo changes.

Happy to hear your thoughts - what would you like to see us take this?

oz dot katz at treeverse.io would be best. ICP is SMB/mid-sized ISVs.

not at the moment. You can use lakeFS directly with Fuse-Mount to do something similar with your own compute.

Currently a static 2 cores and 4GB RAM, no GPU. Will be configurable soon!

Exactly that!

Much appreciated! and good luck with your project

https://www.youtube.com/watch?v=fDR8tmes020 - a 2 minute hands-on demo!

indeed - this only applies to the filesystem managed by tilde. Existing version control is fine if you're only managing code. For data (Think large parquet files, millions json files, images and videos, etc), git doesn't scale well for that.

The repository itself get fuse-mounted into the running sandbox - no skill or MCP required to interact with data: an agent can simply `cat <file>` and use whatever tools they are already good at using.

Sure! and it's not either/or - you can either import code from GitHub (or any other git remote) into a Tilde repository, or simply clone a repository directly inside the sandbox if you want full control over the git commit/branch semantics.

That's fair. What makes this unique is the versioned, composable filesystem. It's built on top of lakeFS (https://github.com/treeverse/lakeFS) so it scales really well, unlike other solutions that try and do this with Git directly.

the repo acts as a source of truth for agents. think memory, data & code. If an agent decides to change any of those, version control allows:

1. to have a human in the loop to approve certain changes 2. rollback changes that end up being incorrect 3. allow reviewing the timeline and history to figure out what changed and how

Fair enough - the underlying technology is indeed open source (https://github.com/treeverse/lakeFS) - the service provides the hosting and tooling to make it easy for consumption by agents.

Good question - the filesystem is Fuse-mounted into the sandbox, not copied into it. This way agents can modify data directly simply by interacting with the "local" files.