Just a heads up for people. This will not work (as far as I could get working) with a system running 0.3.x backups to s3 and trying to restore from them.

Oh hey this is using my go sqlite vfs module[0]. I love it when I find out some code I wrote is useful to others!

[0]: https://github.com/psanford/sqlite3vfs

It worked great! Thanks for your work on it.

that's all we really want in life.

> What we’re doing here is instantaneous point-in-time recovery (PITR), expressed simply in SQL and SQLite pragmas.

> Ever wanted to do a quick query against a prod dataset, but didn’t want to shell into a prod server and fumble with the sqlite3 terminal command like a hacker in an 80s movie? Or needed to do a quick sanity check against yesterday’s data, but without doing a full database restore? Litestream VFS makes that easy. I’m so psyched about how it turned out.

Man this is cool. I love the unix ethos of Litestream's design. SQLite works as normal and Litestream operates transparently on that process.

This is such a clean interface design:

  export LITESTREAM_REPLICA_URL="s3://my-bucket/my.db"
  export AWS_ACCESS_KEY_ID="your-access-key"
  export AWS_SECRET_ACCESS_KEY="your-secret-key"

  sqlite3

  .load litestream.so
  .open file:///my.db?vfs=litestream
  PRAGMA litestream_time = '5 minutes ago'; 
  select * from sandwich_ratings limit 3;

For macos users,

brew install sqlite3, then change the bottom part:

  /opt/homebrew/opt/sqlite/bin/sqlite3
  .load litestream sqlite3_litestreamvfs_init
  .open file:///my.db?vfs=litestream

you have to manually pass in the init function name

This is great... just got it working using bun:sqlite! Just need to have "LITESTREAM_REPLICA_URL" and the key id and secret env vars set when running the script.

  import { Database } from "bun:sqlite";
  Database.setCustomSQLite("/opt/homebrew/opt/sqlite/lib/libsqlite3.dylib");

  // Load extension first with a temp db
  const temp = new Database(":memory:");
  temp.loadExtension("/path/to/litestream.dylib", "sqlite3_litestreamvfs_init");

  // Now open with litestream VFS
  const db = new Database("file:my.db?vfs=litestream");

  const fruits = db.query("SELECT * FROM fruits;").all();
  console.log(fruits);

Neat! Would this mainly be used for JavaScript servers running bun (ie, not end users)?

Cool that you got this to work! How did you get the "dylib" location or file.

    brew list sqlite

gives you the installed path, works for any formula.

Neat. What I wasn't able to find was the dynamic library, just the `litestream` executable. Was there some secret you used for the litestream dylib? Thanks in advance!

Looks like you need to build it yourself: https://litestream.io/guides/vfs/

Got it....you beat me to it. I had just figured that part out!! I didn't understand that part. ALSO, and this might be helpful for others, in the releases section, if you expand the little blue link at the bottom that says something like "see the other 35 assets", then the VFS extension will be downloadable from there!

Thanks for humouring me! :D

The litestream one, from the litestream github releases page!

For anyone following this example, one thing to note that I figured out the hard way is that this line from @zackify SHOULD BE HIGHTLIGHTED...

* "Just need to have "LITESTREAM_REPLICA_URL" and the key id and secret env vars set when running the script"

... and that attempting to load the variables using `dotenv` will not work!!

This is awesome. Especially for sqlite db’s that are read only from a website user perspective. My use case would be an sqlite DB that would live on S3 and get updated by cron or some other task runner/automation means (eg some other facility independent of the website that is using the db), and the website would use litestream vfs and just make use of that “read only” (the website will never change or modify the db) db straightup. Can it be used in this described fashion? Also/if so, how will litestream vfs react to the remote db updating itself within this scenario? Will it be cool with that? Also I’m assuming there is or will be Python modules/integration for doing the needful around Litestream VFS?

Currently on this app, I have the Python/flask app just refreshing the sqlite db from a Google spreadsheet as the auth source (via dataframe then convert to sqlite) for the sqlite db on a daily scheduled basis done within the app.

For reference this is the current app: (yes the app is kinda shite but I’m just a sysadmin trying to learn Python!) https://github.com/jgbrwn/my-upc/blob/main/app.py

Author here. Litestream VFS will automatically poll for new back up data every second so it keeps itself up to date with any changes made by the original database.

You don't need any additional code (Python or otherwise) to use the VFS. It will work on the SQLite CLI as is.

Ok, yeah I think litestream vfs isn’t suitable to do as I described in the intended scenario.

Perhaps you would find ZeroFS [0] useful. It works great out the box with SQLite [1] and only depends on S3 as an external service.

[0] https://github.com/Barre/ZeroFS

[1] https://github.com/Barre/ZeroFS?#sqlite-performance

Forgot to say, thanks for posting this, looks quite useful for various projects that have been on my mind. At one point I was looking for a git vfs for Python (I did find one for caddy static serving specifically, but I needed it for Python) but couldn’t find much that wasn’t abandoned—- an s3 vfs might do the trick for a lot of use cases though.

My pleasure, this project has been a lot of fun :).

Yes this approach might be better. Sounds like litestream vfs won’t really do what I wanted in my described scenario

I also have this implemented and ready to go in my Go SQLite driver: https://github.com/ncruces/go-sqlite3/blob/main/litestream/e...

Slightly different API (programmatic, no env variables, works with as many databases as you may want), but otherwise, everything should work.

Note that PRAGMA litestream_time is per connection, so some care is necessary when using a connection pool.

ncruces/go-sqlite3 was the first thing I thought of when I saw .load litestream.so. That's awesome that you've implemented this. Was it a big lift to make it work with wasm?

Not really, no.

Litestream is made in Go, and I have the VFS API well covered.

The bigger issue is Litestream is not really meant to be used as a library.

It depends on the modernc driver, and some bits on mattn, APIs not very stable, just got updated to require Go 1.25 when 1.24 is still a supported version, brings a bunch of non optional dependencies for monitoring, etc.

Eventually I had to fork to make these more manageable. I'd still hope Litestream can be made more modular, and I can depend directly on upstream.

Love the progress being made here. I've been really enjoying learning about another embedded database - DuckDB - the OLAP to SQLite's OLTP.

DuckDB has a lakehouse extension called "DuckLake" which generates "snapshots" for every transaction and lets you "time travel" through your database. Feels kind of analogous to LiteStream VFS PITR - but it's fascinating to see the nomenclature used for similar features. The OLTP world calls it Point In Time Recovery, while in the OLAP/data lake world, they call it Time Travel and it feels like a first-class feature.

In SQLite Litestream VFS, you use `PRAGMA litestream_time = ‘5 minutes ago’` ( or a timestamp ) - and in DuckLake, you use `SELECT * FROM tbl AT (VERSION => 3);` ( or a time stamp ).

DuckDB (unlike SQLite) doesn't allow other processes to read while one process is writing to the same file - all processes get locked out during writes. DuckLake solves this by using an external catalog database (PostgreSQL, MySQL, or SQLite) to coordinate concurrent access across multiple processes, while storing the actual data as Parquet files. It's a clever architecture for "multiplayer DuckDB.” - deliciously dependent on an OLTP to manage their distributed multiple user OLAP. Delta Lake uses uploaded JSON files to manage the metadata skipping the OLTP.

Another interesting comparison is the Parquet files used in the OLAP world - they’re immutable, column oriented and contain summaries of the content in the footers. LTX seems analogous - they’re immutable, stored on shared storage s3, allowing multiple database readers. No doubt they’re row oriented, being from the OLTP world.

Parquet files (in DuckLake) can be "merged" together - with DuckLake tracking this in its PostgreSQL/SQLite catalog - and in SQLite Litestream, the LTX files get “compacted” by the Litestream daemon, and read by the LitestreamVFS client. They both use range requests on s3 to retrieve the headers so they can efficiently download only the needed pages.

Both worlds are converging on immutable files hosted on shared storage + metadata + compaction for handling versioned data.

I'd love to see more cross-pollination between these projects!

As a sandwich enthusiast, I would like to know more about these sandwich ratings.

This sounds pretty cool, but I’m confused about what software being announced. Is there a new release of Litestream?

Author here. Yes, Litestream v0.5.3 has been released with a new read-only VFS option: https://github.com/benbjohnson/litestream/releases/tag/v0.5....

so how would i connect from a separate machine, i can't figure out from the post or release notes or current litestream website docs, how would i use the extension to do that?

Edit:

need to set LITESTREAM_ACCESS_KEY_ID, LITESTREAM_SECRET_ACCESS_KEY, LITESTREAM_REPLICA_URL

then the module works

I noticed the new release also includes "directory replication support for multi-tenant databases", great addition as well!

From a Litestream user’s perspective:

Litestream continues to work as always, making continuous backups to S3.

Like always, I can restore from those backups to my local system.

But now I have the option of doing “virtual restores” where I can query a database backup directly on S3.

Does this work with sqlite extensions? If I were using e.g. sqlite-vec or -vss or some other vector search extension would I be able to use litestream to back it up to S3 live, and then litestream-vfs to query it remotely without downloading the whole thing?

Yeah I wonder how it would work with things like custome application functions etc. But I guess the query is run locally and it's the pages that are fetched from S3? So it just works? That would be awesome.

I guess there's only one way to find out.

because liters works at the lowest level (virtual file system) all other things should just work.

Yes. (that's all, really)

I remember when Litestream not being a VFS was a plus https://news.ycombinator.com/item?id=29461406 ;)

It still is! I don't know how much clearer we could be about this point. :)

Litestream does not require a VFS to work. It still does all the cool stuff it did before; in fact, it does those things much better now, even without the VFS.

I'm glad they did this! I've always thought VFS was a better fit for the objectives of Litestream than the original design.

SQLite VFS is really cool tech, and pretty easy to work with (IMO easier than FUSE).

I had made a _somewhat similar_ VFS [1] (with a totally different set of guarantees), and it felt pretty magical how it "just worked" with normal SQLite

[1] https://github.com/danthegoodman1/gRPSQLite

Been tinkering with litestream... the read-only VFS is neat but I'm curious about eventual write capabilities... using VFS for distributed DBs could unlock some interesting patterns.

ALSO I'm thinking about mixing this with object store caching... maybe combining memfs with remote metadata; would love to see more details on performance.

BUT I might be overthinking it... just excited to see SQLite exploring beyond local files...

Opens up a whole new can of worms. Transactions come to mind. Who would be responsible for coordinating? If two nodes wrote to the table conflicting information at the same time, who wins?

Author here. We've done some proof-of-concept work on creating distributed leases using S3. We have some use cases internally where we've considered adding write capabilities to the VFS but we haven't started any work on it yet.

It’s a tricky problem that goes beyond the fs as you know. Since it’s cloud and since it’s distributed, a manager wouldn’t be that far fetched that could issue CRDT like messages across the cluster of nodes to issue a “write” or utilize the sync mechanism you have to propagate a “master” db that you write to (aggregate or designate, either way). I did some work on this on a go based database graph and ended up doing a gossip sync crdt message bus.

If latency isn't important. Something like this could work. For applications where latency or linearizability is important, it probably won't work.

Is this VFS for read-only databases? Or can I query a database that has a single litestream writer somewhere continously making updates and backing them up to S3?

The VFS is read only but it will continuously poll for new updates so if you have a writer somewhere else using regular Litestream then it will pick up those updates automatically.

I work with many distributed, often offline, hosts with varied levels of internet speeds. Does this do any offline caching? Like if I load a vfs litestream database on one of my nodes and it goes offline can it still query or will it fall over unless the data was recently fetched?

Author here. It will cache pages in memory right now but it doesn't do disk-based caching currently. That's a good idea though and probably not too difficult to implement. Feel free to post a feature request as a GitHub issue on the project.

I dont fully understand this, would this be useful for scaling sqlite on systems that have really high read needs and a single writer? I thought that was what LiteFS was for, or am i off on that too?

No, you're right: scaling "out" SQLite is what LiteFS is about, and this is about (significantly) improving operational capabilities for single-server (or multi-independent-server, like distributed cache) SQLite deployments.

I have to say I'm similarly not really sure I'm getting it, and I've been following litestream pretty closely over the years (not an expert but certainly familiar).

I think what we're getting here is a way to just spin up a local shell / app and run arbitrary queries from any point in time over the network without having to sync the full prod database. I guess with LiteFS you would have to do this, or pre-plan to do this, it's not totally on-demand.

Or said another way, do things locally as though in prod without having to ssh to prod and do it there (if you even can, I guess if 'prod' is just s3 you can't really do this anyway so it's an entirely new capability).

@benbjohnson is this right? I humbly suggest adding a tl;dr of the main takeaway up top of the post to clarify. Love your work on litestream, thanks for what you do!

Your second paragraph is correct, with the added proviso that you can build features in your own application tooling to take advantage of the same capability.

One reason you're not getting such a clear usage statement at the top of this post is, it's an early feature for a general-purpose capability. I think we might rather get other people's takes about what it's most useful for? There are some obvious use cases, like the one you just identified.

very fair! Thanks for confirming

Does this mean that I can run an application in K8s via one or many horizontally scaled pods all running off DB in s3? No StatefulSet required?

I was doing something similar just the other day and came across sqlite-s3vfs[0]. It is likewise a SQLite VFS that translates IO to S3 api calls. However, that project is only for python and seemingly abandoned. Additionally, if you want more than one writer, you'd have to coordinate that yourself, afaik.

[0]: https://pypi.org/project/sqlite-s3vfs/

Author here. The VFS support right now is currently read only so it's useful for something more like a shared cache of data.

more goodies nice!

I am going to integrate Litestream into the thing I am going to building[1]. I experimented with a lot of ways, but it turns out there is WebDAV support recently merged, not in the docs.

[1]: https://github.com/blue-monads/potatoverse

dumb question: can this be used for versioned tables then ? what to see the state of a table 1 hour ago ?

Author here. You can query the state of a table from an hour ago with Litestream VFS. It won't give you versioned tables in the sense that every time you update a row that it writes a new version in a table somewhere though.

Would this work with other object stores or is it s3 specific?

Works with other object stores.

So much fun streaming/sync/cdc stuff happening, all so cool. Having an underlying FUSE driver doing the Change Data Capture is really neat. This looks like such an incredibly lightweight way to remote-connect to sqlite. And to add a sort of exterior transaction management.

Different use case, but makes me think of sqlite Rewrite-it-it-Rust Turso announcing AgentFS. Here the roles are flipped, sqlite is acting as a file store to back FUSE, to allow watching/transaction-managing the filesystem/what agents are doing. Turso also has a sick CDC system built in, that just writes all changes to a cdc table. Which is related to this whole meta question, of what is happening to my sqlite DB. https://turso.tech/blog/agentfs

Just to be clear, the underlying FUSE thing is LiteFS, not Litestream; nothing described in this post needs a FUSE filesystem, just a small SQLite VFS plugin library.

Thanks Thomas.

To just drop the relevant paragraph that addresses my un-clarity/in-correctness (and which is super fun to read):

> Litestream v0.5 integrates LTX, our SQLite data-shipping file format. Where earlier Litestream blindly shipped whole raw SQLite pages to and from object storage, LTX ships ordered sets of pages. We built LTX for LiteFS, which uses a FUSE filesystem to do transaction-aware replication for unmodified applications, but we’ve spent this year figuring out ways to use LTX in Litestream, without all that FUSE drama.

You got it. Yeah, Ben built LiteFS a year or two ago, which is why he designed LTX. But using LiteFS required people to set up FUSE filesystems, which was too complicated for a lot of people, so Litestream got way more uptake. This past year he's been harvesting all the good stuff from LiteFS that doesn't require FUSE and building it into Litestream.

The easiest way so far to understand the split between Litestream and LiteFS: Litestream is an operational tool, for backup and restore. LiteFS is a method for doing online leader/follower replica clusters.

Now do this with DuckDB.

Use iceberg tables for that in duckdb

Really nice. We should have this as an add-on to https://app.codecrafters.io/courses/sqlite/overview It can probably teach one a lot about the value of good replication and data formats.

If you are not familiar with data systems, havea read DDIA(Designing Data Intensive Applications) Chapter 3. Especially the part on building a database from the ground up — It almost starts with sthing like "Whats the simplest key value store?": `echo`(O(1) write to end of file, super fast) and `grep`(O(n) read, slow) — and then build up all the way to LSMTrees and BTrees. It will all make a lot more sense why this preserves so many of those ideas.

I use Litestream for near real-time backups. Does not change how SQLite is used on the server, just a replacement for .backup

No, it's still pretty cool, easy to use with low operational complexity in low volume read-mostly projects: CMSs, blogs, ecommerce platforms.

It's got crazy write throughput too if you hold it right.

I am. Super simple. Super cheap. Great dev experience. Want to know whether the migration is going to work? Just download the prod db locally and test it. I'm happy.

Works for very small prod databases, I guess.

We use it internally for some rather large databases. It's not database size that matters, it's usage pattern.

What is a very small database? At what size do you think read queries stops working in SQLite?

Handles billions of rows just fine. Can take you unreasonably far on a single server.

For things like config management I feel like it makes all the sense in the world. Whomever the primary is can soak some infrequent-ish write-load. Then the whole DB can quickly copy to where it's needed, or, in lite stream VFS 's case, even less needs to be shipped.

People are building DBMSes and, instead of writing the engine from scratch, are choosing an off-the-shelf solution that integrates into a DBMS with ease.

A better question to ask is why the world needs yet another DBMS, but the reasons are no doubt valid.

Nope, still going strong.

https://news.ycombinator.com/item?id=46124205