There are many cases where replication to a warehouse makes sense, but there are also plenty where it doesn't. Our philosophy at Splitgraph, where we're building something just like this blog post, is that querying should be cheap and easy in the experimentation phase. Analysts shouldn't need to ask engineers to setup a data pipeline just so they can query some data they might not even want. So why not start with a "data lake" (or "data mesh")? You can query any data connected to the mesh without any setup or waiting time. If you eventually decide that you don't want the tradeoffs of federation, then you can selectively warehouse only the data you need to optimize certain queries.
(I might go even further and suggest that the data industry has the idea of the "modern data stack" completely wrong. Replicating data to a warehouse is the ultimate act of centralization. Every other aspect of software is trending toward decentralization; it's the only real way to deal with increasing complexity and sprawl. It's only a matter of time before this paradigm shift hits the data stack too.)
However, I'm disinclined to believe there will be a trend away from a centralized data warehouse (or data lake).
There is inherent value in having a single source of truth for analytics. With modern cloud tools abstracting away the complexity distributed analytical databases, data warehousing is getting easier and more powerful.
It's true that there is added complexity in centralizing the data. But as the author of this article suggested, you're in a bad spot if your marketing team and your sales team can't agree on last month's revenue. I'm not sure how you'd solve that problem in an architecture where the data isn't getting centralized.
> I'm not sure how you'd solve that problem in an architecture where the data isn't getting centralized.
Isn't the entire point of this article that the author achieved a single source of truth without moving the data by using postgres's Foreign Data Wrappers?
My understanding is that FDW's act as a proxy. I don't believe the tables are materialized.
Can you link me where it says otherwise? I don't see anything in the docs [1].
[1] https://www.postgresql.org/docs/current/postgres-fdw.html
Federated data systems, like what is described in this post, are a kind of “nerd trap.” They’re fascinating and people keep building them, but they have little real world utility. Everyone is either big enough where it’s worth it to build a real data warehouse, or small enough that they shouldn’t be messing around with complex data infrastructure.
Decentralization can help with sprawl (aka scale) but it makes software more complex so it’s not a way to tackle complexity. Decentralization is only tolerated because, at some point, it’s the only method left to scale. As an example, distributed databases are significantly more complex to build than a single node database.
You’ve also ignored all of the downsides of using your OLTP database for OLAP queries. The most important is that adhoc queries can impact production. In my experience, that cost dwarfs the costs of maintaining an ETL pipeline as soon as you have your first outage. Another downside is that with a federated, pull based model you need to keep the mapping (FDW definition) up to date. In essence, the maintenance cost moves from an ETL pipeline to an FDW definition.
Re: maintaining the mapping. Sure that's a tradeoff. But I'd rather manage declarative FDW definitions than a programmatic ETL pipeline. Yes, handling schema changes is non-trivial (a naive starting point is periodic introspection of the remote table), but that's true in an ETL pipeline too. As you say, you're moving the cost from the ETL pipeline to the FDW definition. But what if we can lower the cost of the "FDW definition" (a.k.a. setting up and maintaining the data mesh)?
From our perspective, at least with the product we're building, we think we can make a data mesh that's low maintenance enough that it makes sense as a first default before ETL. But the two aren't mutually exclusive. A data mesh can connect multiple warehouses just as easily as it can your production databases. For example, you could move data (ETL) from your data sources to your Snowflake and BigQuery warehouses, and then connect them both to your data mesh.
Also, a mesh can add value as a unified access layer that proxies queries and applies data governance rules (e.g. sharing data, access control, column masking, query whitelisting, rewriting, rate limiting, auditing, firewalling, etc.). It would be hard to do this in your warehouse, because as soon as you have another one, or there's some data that you can't ETL into it, you lose any benefits of its access layer. (e.g. Maybe you have two warehouses because you haven't fully migrated to one, or for cost reasons like storing FireBase clickstream data in BigTable that you can't move to Snowflake with the rest of your data. Now you need to maintain two sets of accounts and ACLs for each warehouse.).
(And for what it's worth -- this is a bit off topic of the current thread -- we do recognize there is often a use case for warehousing when the tradeoffs of federation aren't worth it. We can actually act as a warehouse too, which makes it easy to start with a data mesh but have the option to bail out to warehousing for the specific data assets that need it. Ideally, you can just click a button that says something like "snapshot data" or "schedule query" or "replicate data source." To ingest and store the data, we can leverage the concept we introduced of "data images" for storing versioned snapshots of data in delta-compressed, content-addressable chunks of hot-swappable cstore_fdw files. You can create data images with a declarative, Dockerfile-like syntax called Splitfiles. They look something like FROM upstream:tag IMPORT { SELECT x, y FROM ... }). You get full provenance and can `sgr rebuild` an updated image only when its upstream changes, just like Docker.)
I have been actively championing this point of view for years but... I'm curious, are any paying customers actually interested in this? I don't know how you can market this successfully.
The common selling point across customers seems to be the integration of the access layer (Postgres proxy) and discovery layer (data catalog). Their data storage/access patterns are different. e.g. We have one who uses it as a data lake for periodically querying CSV files in a 20TB S3 bucket. But then another is more interested in the versioning features; they're building data images in Prefect workflows, tagging them with Semantic Versioning syntax, and pushing them to their Splitgraph deployment. For all of them, the discovery features of the catalog are important.
We still have a lot of work to do on the ACL / governance features, but we've had prospects specifically ask about them as priorities.
IMO, the best argument for our thesis is similar to GitLab's (unify the DevOps lifecycle). Even though GitLab didn't have the best-in-class version control or CI on their own, by offering them together they made it worth using GitLab. We want to unify the data stack by integrating the discovery, access and (optionally) storage layers. Even if each layer has some tradeoffs, the benefit of integrating them has a multiplicative effect on the platform itself.
It is probably more complex than replicating a single database to a Snowflake, but one huge benefit of the described solution is that it's real-time - and getting realtime replication to data warehouse is a very different challenge.
It allows you to create both batch and realtime data products just by creating views or materialized views, which is pretty simple and neat. You also get real-time ad-hoc querying, which is very valuable.
The issue with this approach is it keeps the data in row-based format so from certain size it isn't going to be practical. For the "daily report" use case, it's probably fine for a while, but if the source system grow to terabyte size, you will be spending a lot of time waiting on Postgres to do it's thing.
So as the data grows, you might lose first the real-time ad-hoc querying (if you ever had it), then the real-time reporting and then the daily reporting.
But maybe the data will never grow that much and everything is going to be fine for a long time? Maybe this gets you where you need to be for a few years and who knows how the data platforms will look like in a few years.
If you intentionally get rid of your original data then you're saying goodbye to provenance tracking and also open the door to lose information with each conversion you believe makes things "simpler to maintain".
I've used Postgres's FDWs for an analytical DB at a past job. The reason we used them was because we needed minimal analytics for the projects, and didn't want to deal with a more complex data pipeline solution. It sounds like the author was also trying to avoid a more sophisticated data pipeline solution for as long as possible.
Is this a typo or some kind of joke about BigQuery being buggy?
Anyway, agreed on this being needlessly complex and I stopped reading pretty much immediately when I saw the article continue to talk about postgres when it came to the "data lake". If you already have a postgres database backing every service and already have views defined for the data then just set up exports of those views to BigQuery/Snowflake/Redshift/etc. All of them have out of the box templates for doing exactly that.