Build your own “data lake” for reporting purposes
tech.fretlink.com
tech.fretlink.com
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
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.
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.
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.
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".
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.
There are two potential flaws, though.
One is that while using a read replica for reporting works, there be dragons. There is a reasonable amount of tuning possible here. It's important not to conflate the reporting replica with a standby replica for prod. You also might consider asynchronous replication over synchronous replication for the reporting DB. Lastly, there are (tunables) limits to how much you can get away with, but long running queries on either end can potentially cause queries to get cancelled should they hold up replication enough. In other words, it is still very valuable to have optimized, small queries on the reporting DB where possible.
Second is that this works fine until your data doesn't fit on an instance. Large RDS instances are not cheap, and at some large table size it begins to make sense to look at warehouse solutions, like Spark on S3/whatever or Redshift or Snowflake, which can scale out to match your data size. I'd be concerned to find a rats-nest of reporting database instances when a proper cluster should be used.
I'd simply suggest reading this top-to-bottom very carefully, until you understand what each option does. That should in turn help you understand how and where replication impacts the source and the sinks. The docs can be terse, but it's rare for anything to be omitted -- read with care.
Data Warehouses are used to store and process anything which looks like a structured table, or nested tables, so called semi-structured data.
Data lakes are for everything else. Your SQL warehouse can't process or is not ergonomic for processing of:
- purchased data, shared as a zip archive of csv files
- 10 level deep nested json files coming from api calls
- html files from webscrapes
- contents of ftp containing 100s of various csv and other files
- array data used for machine learning algorithms
- pickled python ml models
- yaml configs
- pdf documents such as data dictionaries
- materialized views over raw data
I think Data Mesh concept is proposing structuring your org as a bunch of data producing teams, each maintaining their own data lake, instead of having one huge ass lake mess in the middle.
Tools like delta.io and Databricks are giving data lakes full capabilities of a data warehouse, so the difference between data lake and data warehouse is diminishing. These days you can get away without a dedicated DWH and just store everything in a blob store and plug in short-lived processing engines as you wish without vendor lock-ins.
https://blog.openbridge.com/how-to-be-a-hero-with-powerful-p... https://stackoverflow.com/a/36831549
And with projects like Apache Iceberg you can also have ACID transactions which will make it easy to update or delete just using SQL. This really opens up the separation of compute and data and you can use any engine you want (spark, drill, hive, impala, athena, dremio, redshift spectrum etc) on top of your files.
Other people want to deploy robust, tested and tried solutions, that won't break in mysterious ways, just to make money.
I side with the later.
A data-warehouse is typically somewhat more structured and doesn't just collect data but also combines and links data from multiple sources. Typically with the goal of creating a set of tables that you can use for reporting without needing to know all the intricate details of how the source-data is linked.
A data-warehouse can be based on a datalake. You could also make a data-warehouse without first building a datalake but keeping the datalake part separate allows for better separation of concerns. You can also have datalake without building a data-warehouse on top of it, it depends on what you want to use it for.
> Data lakes and data warehouses are both widely used for storing big data, but they are not interchangeable terms. A data lake is a vast pool of raw data, the purpose for which is not yet defined. A data warehouse is a repository for structured, filtered data that has already been processed for a specific purpose.
# Data lake
Data is stored en masse with no schema applied, either unstructed or structured data can be dumped straight into the lake or can be transformed and then dumped in. Turns into a data swamp when it becomes unusable due to staleness or complexity.
Data lakes are basically an AWS S3 bucket business users can access and (attempt) to do reporting on.
# Data warehouse
Heavily structured schema applied to data used in reporting, usually defines the single point of truth for business purposes. Uses a star schema model (if you follow Kimball [0] methodology) to create dimension tables used to filter and aggregate raw measurements from the central fact tables (which contain your actual measures like £ made on 1 sale).
Kimball and Inmon [1] philosophies come with their own benefits and trade offs. See bottom of [2].
Edit: got methodolgies the wrong way round with initial costs, linked article has a useful table that I didn't see.
Data warehouses have a very concrete definition and are usually implemented via Kimball's or Inmon's method. When I've worked with them they've become the bastion of business reporting (excel users love a pivot table).
---
Just to confuse matters, there's also the data vault: https://en.m.wikipedia.org/wiki/Data_vault_modeling
0: https://en.m.wikipedia.org/wiki/Ralph_Kimball
1: https://en.m.wikipedia.org/wiki/Bill_Inmon
2: https://www.zentut.com/data-warehouse/kimball-and-inmon-data...
> We see customers creating big data graveyards, dumping everything into Hadoop distributed file system (HDFS) and hoping to do something with it down the road. But then they just lose track of what’s there. The main challenge is not creating a data lake, but taking advantage of the opportunities it presents.
https://en.wikipedia.org/wiki/Data_lake
> If you use your data lake as the source for your warehouse then you’re forced to keep it clean
So you apply a schema...? Determining what should and should not live in your storage? What format it should be in etc.?
That's not a data lake in my mind -- it's just the ETL staging area.
This is why I dislike the term data lake. It's so vague and loose that it could mean 300 different things to 300 different experts. Makes it a great term for snake oil marketing though.
Good! This is much better than building something well before you need it. So when do you know you'll need it? By having each team track trends in metrics for their services/work and predicting when limits will be exceeded. (This used to be mandatory back when we had to buy gear a year or two ahead) This catches persistently increasing data sizes as well as other issues. Each team should track their metrics, write a description for what its effect means for the whole product, and forward them to a product owner (along with estimates of when limits will be exceeded).
> do try to set up some streaming replication on a dedicated reporting database server for quasi-live data (or set up automated regular dump & restore process if you don't need live data)
Both of these are potentially hazardous to performance, security, regulatory requirements, customer contract requirements, etc. Replication can literally halt your production database, as can ETL, and where it goes and how it's managed is just as important as for the production DB. So look at your particular use case and design the solution just as carefully as if you were giving direct access to production.
As for the structure of it all, once you start adding outside data, you'll find it's easier to architect your solution to have multiple tiers of extracted data in different pipelines, and to expose them each to analysis independently. You can make all of them eventually land in a dedicated database, but allowing individual analysis allows you to later build purpose-fit solutions for just a subset of the data without having to constantly mutate one "end state" database (which may become vastly more complex than your production database).
Btw, the name for this kind of work is called Business Intelligence (https://en.m.wikipedia.org/wiki/Business_intelligence)
I would rather spend months to build a good platform rather than working on a over-complicated and low-level setup like this.
Our goal with Splitgraph is to provide a single interface to query and discover data. Our product integrates the discovery layer (a data catalog) with the query layer (a Postgres compatible proxy to data sources, aka a "data mesh" or perhaps "data lake"). This way, we improve both the catalog and the access layer in ways that would be difficult or impossible as separate products. The catalog can index live data without "drift" problems. And since the query layer is a Postgres-compatible proxy, we can apply data governance rules at query time that the user defines in the web catalog (e.g. sharing data, access control, column masking, query whitelisting, rewriting, rate limiting, auditing, firewalling, etc.).
We like to use GitLab's strategy as an analogy. GitLab may not have the best CI, the best source control, the best Kubernetes deploy orchestration, but by integrating them all together in one platform, they have a multiplicative effect on the platform itself. We think the same logic can apply to the data stack. In our vision of the world, a "data mesh" integrated with a "data catalog" can augment or eventually replace various complicated ETL and warehousing workflows.
P.S. We're hiring immediately for all-remote Senior Software Engineer positions, frontend and backend [5].
P.P.S. We also have a private beta program where we can deploy a full Splitgraph stack onto either self-hosted or managed infrastructure. If you want that, get in touch. We'll probably be in beta for 12-18 months.
[0] https://www.splitgraph.com
[1] We talked about all this in depth on a podcast: https://softwareengineeringdaily.com/2020/11/06/splitgraph-d...
[2] https://martinfowler.com/articles/data-monolith-to-mesh.html
[3] https://www.splitgraph.com/blog/foreign-data-wrappers
[4] https://www.splitgraph.com/docs/concepts/images
[5] Job posting: https://www.notion.so/splitgraph/Splitgraph-is-Hiring-25b421...