For the past few months, as part of my job at Supabase, I have been working on pg_replicate. pg_replicate lets you very easily build applications which can copy data (full table copies and cdc) from Postgres to any other data system. Around six months back I was figuring out what can be built by tailing Postgres' WAL. pg_replicate grew organically out of that effort. Many similar tools, like Debezium, exist already which do a great job, but pg_replicate is much simpler and focussed only on Postgres. Rust was used in the project because I am most comfortable with it. pg_replicate abstracts over the Postgres logical replication protocol[0] and lets you work with higher level concepts. There are three main concepts to understand pg_replicate: source, sink and pipeline.
1/ A source is a Postgres db from which data is to be copied. 2/ A sink is a data system into which data will be copied. 3/ A pipeline connects a source to a sink.
Currently pg_replicate supports BigQuery, DuckDb local file and, MotherDuck as sinks. More sinks will be added in future. To support a new data system, you just need to implement the BatchSink trait (older Sink trait will be deprecated soon).
pg_replicate is still under heavy development and is a little thin on documentation. Performance is another area which hasn't received much attention. We are releasing this to get feedback from the community and are still evaluating how (or if) we can integrate it with the Supabase platform. Comments and feedback are welcome.
[0] Postgres logical replication protocol: [https://www.postgresql.org/docs/current/protocol-logical-rep...)
its also inherently much nicer than listen/notify, since you don't have to go back and figure out what data was associated with the event
With MySQL I saw the interesting use-case of the black hole storage engine to scale out replication logs but ultimately the only usage I’m aware of was for scaling other mysql read replicas.
The idea of scaling an application by tailing logs from a database sounds very interesting to me, and I’m curious if you’ve explored this at all. There’s of course things like Kafka (and then things like Debezium), but it’s hard to beat direct!
I'd expect more clients to put more pressure on the resource usage on the db. But it's not clear whether the relationship between number of clients and resource usage is linear, quadratic or something else since I haven't done benchmarking yet.
> The idea of scaling an application by tailing logs from a database sounds very interesting to me, and I’m curious if you’ve explored this at all. There’s of course things like Kafka (and then things like Debezium), but it’s hard to beat direct!
It doesn't exist yet, but I was thinking of creating a sink which exposes Postgres WAL via websockets. This way the number of clients might scale much better.
There's external tooling like his project, but postgres extensions in Rust are exciting.
Full extensions via pgrx have been cool to see, but plrust + pg_tle is also starting to show up.
If you aren't familiar with TLE (Trusted Language Extensions), it is a postgres extension from AWS that created some privileged interfaces for procedural languages (used for user-defined functions) to do some extra stuff. Right now it's mostly auth-related hooks but my hope is that it expands in the future.
Plrust is a procedural language extension for Rust, allowing user defined functions written in Rust.
The combination of those two could open up a world of rich extensions usable in managed hosted environments like RDS.
Full disclosure: I'm a Rust Engieer at Supabase.
[0] https://github.com/supabase/pg_graphql [1] https://github.com/supabase/pg_jsonschema [2] https://github.com/supabase/wrappers
Windmill (https://windmill.dev) used to only support webhooks to trigger code and flow jobs. We have just added email support building our own MX server, and wanted to add CDC change. We were gonna do it on Debezium but this will allow us to remove the need for a third-party service and just add this as a crate. Thank you supabase for open-sourcing this.
My specific use-case would be: Single Postgres in my cluster, replicated via something based on pg_replicate running as a side-car and writing to my NAS running Minio.
> equivalent to SQLite: a very cheap option for simple deployments
FWIW, we will launch something tomorrow which will cover this part of your use-case. You will be able to find it on our blog in the morning (PT)
Might be a good idea to have it documented or have the default level set to info for the stdout example.
Maybe this is common Rust knowledge and I just don't know what I'm doing though.
Most applications will have something like
tracing_subscriber::registry()
.with(
tracing_subscriber::EnvFilter::try_from_default_env()
.unwrap_or_else(|_| "my_app=info".into()),
)
.with(tracing_subscriber::fmt::layer())
.init();
The issue we've ran into is some team at work decides to re-write an entire table and things get backed up until they stop updating rows.
> Have you thought about / have solved sharding consumption across multiple slots / multi consumers to increase throughput?
Not yet, there has been not performance yet, as the project is still quite young.
I'm exploring an alternative way to run logic asynchronously after db operations without the overhead, and I think using cdc to export jobs into an external queue is the way to go here. Essentially a lightweight alternative to Debezium with a better developer experience that is easier to manage. This crate could serve as the core of such a service.