Tools like Fivetran and Airbyte exist to solve one narrow problem. They handle the extract part of ETL. Or more accurately today, ELT. Getting data out of source systems and landing it in a warehouse or data lake.
They do this through managed connectors. You authenticate, point at a destination, and within minutes data is flowing. They also handle schema drift. When source data changes over time, new columns, changed fields, they adapt without you having to write and maintain custom code.
For teams running on platforms like Databricks, this can be especially attractive. Integrations with Unity Catalog make governance and access control easier, and you get something usable very quickly.
That speed is the real value. The question is when that tradeoff stops making sense.
The Real Tradeoff Is Cost Over Time
At small to medium scale, these tools are often a no-brainer.
Take something like syncing HubSpot data. You plug it in, land sales data in your warehouse, and move on. For modest volumes, the cost can be very low. Internally, we've had connectors that cost tens of dollars per month, even after price increases. At that level, it's not worth building and maintaining your own pipelines, even if you have the skill set. This is commodity work.
Where teams get into trouble is at scale.
Fivetran prices based on Monthly Active Rows. As data volumes grow or churn increases, the bill tends to move in one direction. Up. Replicating entire databases, especially ones with analytics tables that are frequently recreated or have many frequent changes, can cause row counts to spike quickly. What started as a convenient managed service can turn into a four-figure monthly cost.
At some point, you hit an inflection where it's cheaper to build and operate the pipeline yourself. That point is different for every company. The important thing is knowing it exists.
"What started as a convenient managed service can turn into a four-figure monthly cost."
Cost Over Time
The Inflection Point
Managed ELT costs scale with data volume. Self-hosted costs stay relatively flat after initial setup.
Use the Trial, Then Think Past It
One thing Fivetran does well is cost visibility early on. You can run a trial, see how many rows you're ingesting, and get a reasonable estimate of ongoing cost.
That's the moment to ask not just "can we afford this now?", but "how big can this get before it becomes a problem?"
If the data set is small and stable, the answer might be never. If it's large or growing quickly, you should already be thinking about alternatives. Turning off unused connectors, scoping syncs tightly, and avoiding full database replication unless necessary can help. But these tools shouldn't be assumed to be permanent infrastructure by default.
Compliance Changes the Math Quickly
Regulated data is another common surprise.
Many managed vendors technically meet security requirements. Encryption, access controls, audit logs. But formal compliance documents, things like BAAs, are often gated behind enterprise tiers. Those tiers can start at five figures per month.
At that point, you're no longer comparing software to software. You're comparing software to hiring an engineer. And that changes the decision entirely.
This isn't unique to Fivetran. It's a general SaaS pattern. But it's something teams should factor in early if compliance requirements are likely to grow.
5 figures/mo
where enterprise tiers with compliance docs (like BAAs) often start
When Building It Yourself Makes Sense
There are plenty of cases where custom pipelines are the right answer.
Database replication is a common one. Using change data capture, you can stream only what's changed using the database's native logs. From there, data can be written to object storage in formats like Parquet and loaded into your warehouse or lake.
Services like AWS DMS can handle much of this plumbing.
This approach takes more upfront work and ongoing maintenance, but for large or predictable datasets, it's often far more cost-effective.
These Tools Are Accelerators, Not Forever Solutions
The real value of managed ELT tools is how fast they get you started. They collapse weeks of setup into hours and let you focus on higher-value work early on.
What they don't remove is the need to think ahead. Pricing changes. Vendors get acquired. Enterprise requirements creep in. These tools are part of a system, not the system itself.
Used intentionally, they're excellent. Used without an exit plan, they can become expensive surprises.
"Used intentionally, they're excellent. Used without an exit plan, they can become expensive surprises."
Decision Framework
When to use what
A quick heuristic based on real-world patterns.
Small, stable dataset
HubSpot sync, analytics events
Keep managed. Not worth the engineering time.
Growing fast or high-churn data
Full DB replication, event streams
Run the trial. Model the cost curve. Set a ceiling.
Compliance requirements (BAA, HIPAA)
Healthcare, finance, regulated SaaS
Enterprise tiers often cost more than an engineer.
Large, predictable volumes
CDC from production DBs, data lake ingestion
Self-host with CDC + object storage. Tooling is mature.
The right mindset is to treat them as accelerators. Get moving quickly. Learn your data. Know your costs. And understand in advance where you'll switch approaches if the tradeoffs stop working.
