Tigris is the storage layer for any cloud. Low latency & zero egress fees. Now with bucket forking.
San Francisco Bay Area
Joined March 2022
- Tweets860
- Following24
- Followers2.7K
- Likes818
๐ Dropbox plans to acquire Tigris, and everything you rely on today stays as it is: the product, the API, the pricing, the people.
Even more storage, free to move to any cloud.
Hereโs what it means and what comes next: tigrisdata.com/blog/dropbox-โฆ
One of the fun things about running FoundationDB in production is that when someone asks "what's hammering the cluster?", the answer is usually "give me a minute, I need to run a Python script."
1
3
563
The core parts of relational databases are transactions, tables, and indices. At Tigris we use a non-relational database (FoundationDB) but still have indices. How did we do it? We made our own!
1
1
492
Ram is inordinately expensive right now, so everything we can do to cut ram usage makes objgit better for everyone.
2
1
451
Hey all, we have more updates on objgit!
Looks like it's taking off amidst other companies making git services for some reason. Again, really have to wonder what that's all about.
9ms point lookups without running your own database in every region
That's what @goldskyio built with Tigris, @vortexdotdev, and SlateDB
Get our side of the story ๐
At Tigris we love databases, especially FoundationDB. FoundationDB is that: our foundation. It's the basis for everything we do, even our message queue. Turns out it's pretty great to do data and queueing in the same transaction!
1
8
1,180
If our objgit post was so good last time, where's objgit post 2?
Today we have an update on the epic saga of Xe reinventing Git storage on top of object storage: making seek-native packfiles. What does this mean? Well, git is complicated. Huddle into this thread!
1
3
733
"We didn't need a database" is normally the kind of catchy blogpost title that shows up months before "We migrated to Postgres".
Our friend JP runs AmpBase, and his control plane sits directly on top of Tigris' object storage without a relational database under it.
1
3
562
Chicago says an object is deleted. Singapore says it exists. Who's right?
FoundationDB uses a variant of Paxos for consensus within a cluster, but versionstamps don't compare across clusters. Since we run so many clusters all over the globe, we invented our own protocol for conflict resolution.
Read how we did it and how we replicate metadata at scale ๐
1
9
965
Walgit, from @tobi, is exactly what @martinkl was talking about with the classic line: "the log is the database, and everything else is cache or materialized views."
In walgit, the git repository is a write-ahead log in a bucket, and the SoT is the log itself, and the machines running it are throwaway caches.
well, if the log is the database and the database is just a cached view of replaying the log, then what is a git checkout but a cached view of replaying the commit history?
What's cool is we did the same thing, a layer down, to build Tigris: the object store IS the log. Every put is a new immutable version, every delete a tombstone.
We see WALs bolted-on to manage mutability, but when you make the log the source of truth, it inverts the model.
A HN commenter called out how walgit's core idea is the same as objgit-- git built on top of Tigris-- and we think there's some form of convergent evolution here that shows it's the right direction to make the log the SoT.
Give a couple of our blogs a read and see if you agree ๐
"Everything is a log or should be a log if you want to rely on it"
tigrisdata.com/blog/objgit/
2
467
An agent can run off a VM image that is larger than the disk it runs on.
AgentENV by @KVCache_AI keeps microVM disk images in object storage and pulls down only the blocks each machine actually needs. That is how image size stops being limited by local disk.
It also means the local disk is a cache with a size limit. Cold blocks get evicted, and when they are needed again they get fetched again. Every machine in the fleet does this independently, for as long as the cluster is running.
Those reads never stop. With Tigris, they never cost egress either.
Swapping in a different object store takes a few lines of config: kvcache-ai.github.io/AgentENโฆ
1
704
๐ช๐ต๐ ๐ฝ๐๐ ๐ฎ๐ด๐ฒ๐ป๐ ๐๐ฎ๐ป๐ฑ๐ฏ๐ผ๐
๐๐๐ฎ๐๐ฒ ๐ถ๐ป ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐ ๐๐๐ผ๐ฟ๐ฎ๐ด๐ฒ?
For one sandbox on one machine, a host mount is fine. It starts to matter ๐๐ต๐ฒ๐ป ๐๐ผ๐ ๐ฟ๐๐ป ๐ฎ ๐ณ๐น๐ฒ๐ฒ๐, where sandboxes are created and destroyed constantly and land on whichever node has room.
Agent sandboxes are ephemeral on purpose. However, everything the agent produced dies unless it lived somewhere else.
A bucket is reachable from every node.
1. With a volume plugin, ๐ฐ๐ฟ๐ฒ๐ฑ๐ฒ๐ป๐๐ถ๐ฎ๐น๐ ๐ป๐ฒ๐๐ฒ๐ฟ ๐ฒ๐ป๐๐ฒ๐ฟ ๐๐ต๐ฒ ๐๐ฎ๐ป๐ฑ๐ฏ๐ผ๐
and isolation is stronger. The keys stay in a root-owned config on the host and the microVM only ever sees a mounted filesystem.
2. Since mounts are reference-counted per node, ๐ผ๐ป๐ฒ ๐ฑ๐ฎ๐๐ฎ๐๐ฒ๐ ๐๐ฒ๐ฟ๐๐ฒ๐ ๐บ๐ฎ๐ป๐ ๐๐ฎ๐ป๐ฑ๐ฏ๐ผ๐
๐ฒ๐. Model weights mount read-only into every agent on a box from a single copy, instead of one pull per sandbox.
Our S3 compatible backend is now in CubeSandbox -Tencent Cloud's open source microVM runtime for AI agents.
Docs: cubesandbox.com/guide/integrโฆ
2
395
Delete doesn't have to mean the data is gone. Tigris lets you un-delete your most precious objects and buckets if you(r agent) didn't really mean to actually remove it for others.
Today we're going to dive into how the soft-delete feature works so you can learn what goes into making object storage that developers crave.
Read more on the blog: tigrisdata.com/blog/soft-delโฆ
122
Authentication seems simple. Signature-based authentication seems even more simple. Spoiler: it's not.
Today Xe has a deep dive for us into how SigV4 authentication works at Tigris so that your apps, agents, and CLI tools all work the way they should.
Link ๐
๐ค Made with AI
1
493
Learn about SigV4 auth here: tigrisdata.com/blog/sigv4/
111
Congrats to @skypilot_org on the Seed! We've been fans for a while, so we wanted to re-share our _thorough_ guide to launching a truly portable training job on Skypilot compute and Tigris storage.
The guide teaches you to stream datasets bigger than RAM, shard into Tigris, pre-tokenize, train a LoRA โ every step idempotent because GPUs fail. ~15 min on a T4, same YAML anywhere
Link ๐
1
303
Full guide: tigrisdata.com/blog/trainingโฆ
91
Our LangGraph checkpointer is now ๐ผ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฎ๐น๐น๐ ๐ถ๐ป ๐๐ต๐ฒ @LangChain ๐ฑ๐ผ๐ฐ๐.
docs.langchain.com/oss/pythoโฆ
If you build agents with LangGraph, the checkpointer is where your agent's past lives: ๐ฒ๐๐ฒ๐ฟ๐ ๐๐ต๐ฟ๐ฒ๐ฎ๐ฑ, ๐ฒ๐๐ฒ๐ฟ๐ ๐ฐ๐ต๐ฒ๐ฐ๐ธ๐ฝ๐ผ๐ถ๐ป๐, ๐ฒ๐๐ฒ๐ฟ๐ ๐ณ๐ฎ๐ฐ๐ ๐ฎ ๐๐๐ฒ๐ฟ ๐ฒ๐๐ฒ๐ฟ ๐๐ผ๐น๐ฑ ๐ถ๐.
Most teams keep that state in Postgres.
But a database copies rows to branch, so cloning an agent costs more the more history it has.
On Tigris Data, a fork is a pointer, not a copy. You can clone your entire agent, every thread and checkpoint, in constant time, whether it holds 1 MB or 1 TB.
That's what lets you evaluate a prompt change against real user state instead of three blank fixtures: fork prod, replay real threads through the candidate, judge it against the baseline, drop the fork.
Nothing touches live data.
Full walkthrough: tigrisdata.com/blog/eval-ageโฆ
417
If you use RAGFlow by @infiniflowai to answer questions from your documents, everything it knows lives in its storage backend.
By default, if you use MiniIO, RAGFlow re-reads your corpus every time you tune it.ย
When you change your chunking or swap embedding models, every document gets pulled from storage again.
If you point RAGFlow at a Tigris bucket - you can re-process your corpus as often as you want, for free.
If you snapshot the bucket first, a bad re-parse becomes a rollback instead of a restore-from-backup.
Full setup guide:ย tigrisdata.com/docs/guides/rโฆ
1
205