These just seems like over engineered solutions trying to guarantee their job security. When the dataflows are so straight forward, just replicate into pick your OLAP, and transform there.
I came from traditional engineering into data engineering by accident and had a similar view. But every time I tried to make a pipeline from first principles it always eventually turned out something like this for a reason. This is especially true when trying to bridge many teams and skillsets - everyone wants their favourite tool.
What is the current state of the art (open source) when doing oltp to olap pipelines in these days? I don’t mean a one-off etl style load at night but a continuous process with relatively low latency?
Idk what the state of the art is, but I’ve used change data capture with Debezium and Kafka, sink’d into Snowflake. Not sure Kafka is the right tool as you don’t need persistence, and having replication slots makes a lot of operations (eg DB engine upgrade) a lot harder though.
I know it's easy to be critical, but I'm having trouble seeing the ROI on this.
This is a $20k/year savings. Perhaps, I'm not aware of the pricing in the Indian market (where this startup is), but that simply doesn't seem like a good use of time. There's an actual cost of doing these implementations. Both in hard financial dollars (salaries of the people doing the work) and the trade-offs of de prioritizing other other.
The biggest issue IMO is that engineers who work on projects like these inevitably get bored and move on, and then the company is stuck trying to add features, fix bugs and generally untangle the mess, all taking away time and resources from their actual product.
Yeah, but you can always make this argument and build nothing - dealing with all the problems of every 3rd party SaaS / PaaS under the sun. Sometimes it's much easier to just build the thing and then you know where its' limitations are and you can address them over time.
When working with ETL, it really helps to not conflate the letters or worry about them in the wrong order. A lot of the most insane complexity comes out of moving too quickly with data.
If you don't have good staging data after running extraction (i.e., a 1:1 view of the source system data available in your database), there is nothing you can do to help with this downstream. You should stop right there and keep digging.
Extracting the data should be the most challenging aspect of an ETL pipeline. It can make a lot of sense to write custom software to handle this part. It is worth the investment because if you do the extraction really well, the transform & load stages can happen as a combined afterthought [0,1,2,3] in many situations.
This also tends to be one of the fastest ways to deal with gigantic amounts of data. If you are doing things like pulling 2 different tables and joining them in code as part of your T/L stages, you are really missing out on the power of views, CTEs, TVFs, merge statements, etc.
Not all data lives in a SQL database. Much of the extraction code I write does things like loading flat files from unusual sources and querying APIs.
If the source data is already in a SQL store, then the solution should be obvious. You don't need any other tools to produce the desired view of the business at that point. Transforming for an upstream schema is a select statement per target table. This doesn't need to be complicated.
Yeah I extract a lot of data out of Dynamo. It’s still the easiest part. Change capture just isn’t complicated. You need some basic constructs and then you’re golden. The data mart design phase is orders of magnitude more effort.
We use BQ and Metabase heavily at work. Our BQ analytics pipeline is several hundred TBs. In the beginning we had data (engineer|analyst|person) run amock and run up a BQ bill around 4,000 per month.
By far the biggest things was:-
- partition key was optional -> fix: required
- bypass the BQ caching layer -> fix: make queries use deterministic inputs [2]
It took a few weeks to go through each query using the metadata tables [1] but it worth it. In the end our BQ analysis pricing was down to something like 10 per day.
So you saved just $20k per year? Not sure the context of your company but I’m not sure if this turns out to be a net win given the cost of engineering resources to produce this infra gain
> But generally rolling your own has other benefits.
Not for startups it doesn't. The only rolling-your-own they should be doing is their main product.
Once you get bigger with many hundreds of employees, and existing software starts becoming a measurable blocker, then you can gradually build your own stuff once the tradeoffs actually make sense. But it generally takes a while for that to become the case. For startups, premature optimization isn't the root of all evil, but it's the root of a lot of it.
You might argue that it is never worth it to roll anything on your own (which is already an extreme proposition), but to argue that it has no benefits (other than cost?), I think it's either bad reading comprehension or overzeal to jump into the keyboard and type dogma that you should never roll your own and that you should download 1000 dependencies and a dependency manager to manage all those version conflicts.
If you're a startup and you need something and it already exists, why would you write your own instead? Even if it's clunky, it saves you time.
And for startups, speed and runway is everything. Rolling your own is not a luxury you have.
Especially when you consider how much of the startup experience is changing requirements, pivoting, throwing out code, etc.
And I've never seen 1,000 dependencies in my life. Sure maybe you have 50. That's entirely manageable. Think how much time you'd lose writing and testing all that stuff yourself...
Rolling your own generally has mainly downsides in the context they are in.
1. This is clearly a small team with very little spend
2. Tomorrow someone leaves and next engineer will have to manage all of this.
3. I don't think they realize that they actually increased cost of this service not decreased it. Now they need to manage their own Kafka monthly. Engineering time is expensive.
We did have a discussion on Self vs Managed and TCOs associated with it.
1> We have multi regional setup so it came up with Data Sovereignty requirements.
2> Vendor Lock ins - Few of the services were not available in that geographic region
3> With managed services, you often pay for capacity you might not always use. our workloads were often consistent and predictable, so self managed solutions helped in fine tuning our resources.
4> One og the goal was to keep our storage and compute loosely coupled while staying Iceberg-compatible for flexibility. Whether it’s Trino today or Snowflake/Databricks tomorrow, we aren’t locked in.
You’d think that pushing all of the data into any ldap database, but especially some of the newer postgres based ones would give you all the performance you need at 10% of the costs? Let alone all the maintenance of the mind boggling architecture drawing.
Is Debezium the only good CDC tool out there?
I have a fairly simple data stack and am looking at integrating a CDC solution but I really dont want to touch Kafka just for this. Are there any easier alternatives?
Except when a new column is added upstream and DMS inserts the new column not at the end but before the dms timestamp fields, ruining the index based loads of DMS files
How much did this cost in engineering time and how much will it cost to maintain? How about when you need to add a new feature? Seems like you saved roughly 1.5k per month which pays for a couple days of engineering time (ignoring product,mgmt and costs related to maintaining the software)
No idea how many hours to build, but I maintain something similar, (different stack though) and its so trivial I don't even count the hours, so it's probably about 1/2 a day to maintain that every 3 months?
Even if you needed to invent a new feature, you could invent a months worth of features every year and still save money.
I know this is off the topic of the actual post but I'm confused as to why I've been downvoted, seems other people have since made similar comments but my comment has been downvoted. Help me out here, why?
I do have a question on the BigQuery. i f you were experiencing unpredictable query costs or customization issues, that sounds like user error. There are ways to optimize or commit slots for reducing the cost. Did you try that ?
As for BigQuery, while it's a great tool, we faced challenges with high-volume, small queries where costs became unpredictable as it is priced per data volume scanned. Clustered tables, Materialised views helped to some extent, but they didn’t fully mitigate the overhead for our specific workloads. There are ways to overcome and optimize it for sure so i wouldn't exactly put it on GBQ or any limitations.
It’s always a trade-off, and we made the call that best fit our scale, workloads, and long-term plans
I am not sure if managing kafka connect cluster in too expensive in long term. This solution might work for you based on your needs. I would suggest to look for alternatives.
This is a little bit of a word soup. Its hard to see why the various redesigns were done without a set of requirements. I dont get why you'd trigger Airflow workflows for doing CDC. These things were designed for large scale batch jobs rather than doing CDC on a some Google sheets. Either way without scale numbers its hard to why PG was used or why the shift to BigQuery. Anyways the site uses Hugo, which actually sticks out for me.
I'm not a data engineer but was tasked with building an ETL pipeline for a large company. It's all just Step Functions, looping through file streams in a Lambda, transforming, then putting the data into Snowflake for the analytics team to view. My pipeline processes billions of rows from many different sources. Each pipeline runs daily on a cron job (maybe that's the key differentiator, we don't need live streaming data, it's a lot of point of sale data).
Whenever I hear actual data engineers talk about pipelines there are always a million different tools and complicated sounding processes. What's am I missing?
Hi, data engineer here (of sorts). I'll take a crack.
> It's all just Step Functions, looping through file streams in a Lambda, transforming, then putting the data into Snowflake for the analytics team to view.
So, you're right. Data going from A -> B is the most important part of the job. Like, the pipeline you've got is like the 60-80% critical bit. Get data from systems A through Y and put it into system Z. Without it, there's no point talking about anything else.
But what happens when you need to rollback all of the data from system F for the last week? What if you need to rollback system C for the last 2 days as well? How do you keep track of whether data from systems G, H and I is still "good" quality? How do you handle frequently changing source systems, which could do any one of drop, rename, add columns on a daily basis? Do you ignore those columns forever? Or do you siphon that data off to a special place for manually review? Do you process all previously seen data every single night? If you only process the newest subset of data, what do you do if you've noticed a bug in the pipeline code? Process everything again and spend a whole bunch of money every time we fix buggy pipeline code? Or do a "migration" on already processed data? How do you track which subsets were processed with which version of your pipeline? How did that run get triggered -- was it a glue crawler scheduled cron job, or did someone trigger it manually? Does any of this actually matter? Does anyone even look at the dashboards any more?
Getting the pipeline running is the first bit. With non-volatile/slow-changing source data/systems, you're fine with just the pipeline most of the time. Depending on the volume of the data, you could probably get away with just rerunning the pipeline over all previously seen data when there's been a bugfix and just eat the cost. You don't need anything more complicated than that.
But if the source data/systems are volatile [0], then keeping it working, keeping it accurate and keeping it cheap comes next. And the complexity of that depends on how complex X -> Y is.
A simple "back of a napkin" way to determine complexity is to sit down and ask some questions about the 6 V's: Volume; Velocity; Variety; Veracity; Value; Variability [1]. It sounds like [2] you have high volume, some velocity (nightly), some variety, low veracity, high value and low variability data. There's a lot of data coming from multiple source systems, and it's important data. But it doesn't sound like the changes that much and is all probably all in the same or similar formats [3] and good quality. So... a simple pipeline triggered nightly works in that case.
When things get more complicated, that's when the system moving data from A -> B gets more complicated.
---
Having said all that, there has been a lot of resume driven development in the industry. $LAST_JOB i worked at the previous team thought it would be a great idea to roll their own job orchestrator product [4]. from scratch. in django. not a library that django called, not some django API endpoints that forwarded requests to some Airflow backend. nope. a job orchestrator written from scratch in django.
Because "bare metal" apparently.
They were also obsessed with this idea of reproducibility, but what they actually wanted was idempotency... but that's a story for another reply.
---
[0]: even if each upstream is slow to change, the cumulative effect could be there is at least one upstream change breaking something every day, so the cumulative effect is that the upstreams together are volatile even if individually they are not
> "We are a fintech startup helping SMEs raise capital from our platform where we provide diverse financial products ranging from Term Loan, Revenue Based Financing to Syndication, we face one unique data challenge: Our data comes from everywhere."
Reading carefully: the result of this work yields an expected $21,000 USD in annual operating cost savings for infrastructure services.
Is this resume driven development?
What was the opportunity cost of this work? Is the resulting system more or less maintainable by future employees/teammates?
reply