▲ 151 points
back
31 comments
Clickhouse really nailed this with the acquisition of peerdb. I used it with many terabyte databases and I essentially never thought about it. The only thing we really had to watch for was trying to replicate too much at once (because of the physical compute/io capacity of the postgres or clickhouse clusters).
This was basically Vertica's party trick for quite a long time to have a WOS and ROS formats for the same row and anti-caching between those two.
You could've built a similar system with dezebium and delta lake for quite some time but it would fail compactions, if you run it fast enough. I've seen Oracle GoldenGate 12c do this trick in 2014 or so, using Mysql as the cheap replica. But they are all fragile to schema updates in some direction.
The closest batteries-included equivalent to this is the Aurora -> Redshift bridge[1].
Aurora zero etl was a nightmare for us. Almost any schema changes require a VACUUM FULL for it to continue functioning. On a few occasions, it just stopped running without an obvious explanation, requiring slow and lengthy back and forth threads with AWS support. If it worked as advertised, it would be great, but I can't recommend it for any serious production system.
And this is, when it doesn't delete data randomly: https://github.com/trinodb/trino/issues/28885 (some enterprise open source was slop before AI even it seems)
Although pg_lake is open source, worth noting that it heavily refers to but is missing CDC capabilities.
There's a bunch of comments/links to a closed https://github.com/snowflake-eng/sfpg-extension-pg_lake_repl...
I'm interested in pg_lake so I wanted to check out your link, but it seems to be internal to snowflake?
Apologies, I could have introduced that better.
The article links https://github.com/Snowflake-Labs/pg_lake but if you go looking for CDC, it's not there, and all you end up with is links to the private/closed repo that I just linked, in various corners and spots. The point is that all the CDC stuff is in the repo we don't get access to, that isn't open source: https://github.com/snowflake-eng/sfpg-extension-pg_lake_repl...
pg_lake is an open source Postgres extension based on work done at Crunchy Data prior to the acquisition by Snowflake - you can find the repo here [1] and a blog post with more context on the project here [2]
[1] https://github.com/Snowflake-Labs/pg_lake
[2] https://www.snowflake.com/en/blog/engineering/pg-lake-postgr...
I see you are a co-founder! Thanks for pg_lake. I am actually already heavily using it, was just interested to read about CDC in the context of pg_lake.
It's interesting to watch how different companies that offer both Postgres and warehousing solution under 1 roof approach the same problem:
- ClickHouse focuses on traditional CDC (ClickPipes) and just make it blazingly fast
- Databricks leans on their unified storage architecture (LTAP) to avoid copying data (though you can argue there is still a copy in the cache)
- Snowflake uses a data mirroring CDC as extension so it runs directly on Postgres
I'm still waiting for a Postgres provider to just let me mirror data directly to Iceberg, so I can plug in my own stateless query engine.
The challenge is converting primary key updates/deletes to row offsets in a columnar table. That requires maintaining an expensive mapping or doing expensive scans, and is not something you want Postgres itself to do. It's also this bursty, memory-intensive workload that you'd rather not have a lot of dedicated infrastructure for.
At Snowflake we use Snowflake to do the apply work. Hence end-to-end mirroring has more pieces than just Postgres, but the capturing of changes is cheap enough to do in Postgres directly.
(Author)
Yes I’m not suggesting to do this inside Postgres. I’m hoping that a Postgres provider can provide this mirroring capability out of the box, similar to how they provide a connection-pooled endpoint out of the box so I don’t have to self-host pgBouncer. I just want to be able to check a box somewhere and have a table in Postgres automatically mirrored to Iceberg, with guarantee that no data is lost. They can charge more for it, I will gladly pay.
Makes sense, we just shipped it
https://www.linkedin.com/posts/craigkerstiens_barely-over-2-...
> I'm still waiting for a Postgres provider to just let me mirror data directly to Iceberg, so I can plug in my own stateless query engine.
The issue that each of those providers above has recently adopted Postgres as a secondary product aimed at supporting their main product, an OLAP database or engine, so they don’t want you plugging in your own query engine.
I’d bet you’re likely to see this from a Postgres-specific provider first, like Supabase.
> The issue that each of those providers above has recently adopted Postgres as a secondary product aimed at supporting their main product, an OLAP database or engine, so they don’t want you plugging in your own query engine.
Disclosure: I work on Supermetal
You don't need to wait for a provider, and the provider is arguably the wrong place for this. They all have an incentive to make their own OLAP engine the happy path. A dedicated CDC tool that writes Iceberg to your own storage and catalog keeps the tables and the engine choice yours.
We built exactly that, a native Iceberg destination with Merge on Read. Since Snowflake and Databricks reject equality delete files, there's also a positional deletes only mode that writes deletion vectors instead, so the tables are readable from whatever engine you use.
> mirror data directly to Iceberg
> I’d bet you’re likely to see this from a Postgres-specific provider first, like Supabase.
we deprecated this feature in our ETL tool[0]. The functionality is still in there but we can't support some of the production features we'd need for data/schema guarantees
Iceberg is still nascent - only supporting single-table transactions (at least when we tried). A lot of important CDC/transactional semantics were "a work in progress" upstream. We shifted our focus to ducklake, which stores the catalog in Postgres
I’m waiting for Cloudflare R2 to eventually support mirroring Postgres into R2 catalog. It seems like a nice fit because they already have R2 SQL.
Snowflake/Crunchydata comes close to doing that with the pglake extension.
There’s not a mirror function like what Snowflake offers directly but you can come close with a pgcron to upsert changes to the iceberg tables every so often.
You can also purge the table put to the iceberg version every so often too depending on your data needs. Then you can create a query unions the results of both.
> I'm still waiting for a Postgres provider to just let me mirror data directly to Iceberg, so I can plug in my own stateless query engine.
If you count AWS as Postgres provider, DMS into Kinesis into Firehose can do that.
There was preview of just Firehose doing it directly, but AWS have pulled it because it was too unreliable. Maybe they rebuilt it since?
DMS is so unreliable though.
Google: Charges out the wazoo for a half baked product built on top of Dataflow
Snowflake is a really amazing product. It's been a delight using it the last few years.
as much praise as some people give to it, I feel deeply uncomfortable with an idea of SaaS-only DB tech that you don't have an option to self host
Agreed. I personally am very uncomfortable with the idea of SaaS-only DB tech. Databases I deal with are very large. And we are a small company. Self-hosted works for us. We cannot afford hosted database solutions as they charge by storage and some also by data transfer. From my point of view, 3rd party hosting of databases solves problems we don't have. Particularly with AI tools managing our services using ansible/terraform, I think we'd be worse off if we switched to a SaaS product.
Seeing Postgres articles from Snowflake surprises me a lot though given how there's zero relation between Snowflake the product and Postgres itself
EDIT: I now see it's mainly to do with pushing data out of customer's postgres systems into snowflake
They acquired Crunchy Data and now have some of the most prominent Postgres developers working there.
What did Center for Disease Control use before? Mysql?
In this case CDC = Changed Data Capture.
Now fivetran has more competition. Good.
Excellent that Control Data is contributing! Oh wait, I'm a few decades out of sync
I'm afraid nobody (else) here remembers Control Data Corporation. But I'm happy to see the Centers for Disease Control using Postgres.