Moving off Airbyte: Regaining control over your data movement.
Airbyte promises a self serve UI, a catalog of connectors, no code. Sounds great on paper, but in practice this vision is plagued by skill barriers, reliability issues, and high cost of ownership.
Adrian Brudaru,
Co-Founder & CDO
This is why teams leave:
- Self-hosted: you pay in devops time plus infrastructure cost.
- Managed: you pay metered pricing, partly for operations you do not control.
- Both: connector maintenance happens too often, on someone else's schedule. It causes cost, delays, and team frustration.
What moving gets you
Let analysts self-serve, keep your SLAs, stop waiting on the catalog
dlt encodes the senior engineer's decisions into the product, so an analyst can build like a senior data engineer. By leveraging an agent conversationally, the control stays with you, at any level of detail, from "build me a Salesforce pipeline" to "build this endpoint, use this incremental strategy, test fields a, b, c, d for uniqueness and report the result". It works for the entire team.
Who builds with it:
- A CTO who is not a data engineer migrated his whole Airbyte stack in 5 days.
- Any Python developer: at dentolo, ingestion moved from a small group of tool experts to something anyone on the team could author, review, and ship.
- Analysts: at Hiveapp, five senior analysts author and maintain the pipelines themselves - data engineering is no longer the bottleneck for their analytics work.
Claude can write pipelines, so everyone can. dlt pipelines show 99.83% success in the wild, measured on recent 50M open-source production runs. This is not luck. It is what happens when best practices are built into the product. From simple user interfaces, explicit error reporting to schema evolution, hardware resource management, retries on transient errors - dlt pushes the quality standard for data ingestion.
Ingestion moved from being owned by a small group with deep knowledge of specific tools to something any Python developer on the team could author, review, and ship. The question changed from 'who knows the tool?' to 'what data do we need next?'
— Euan Johnston, Senior Data Engineer, dentolo
The catalog is not a limit anymore: dltHub Context covers 9,700+ sources, and the agent can build any source, including internal APIs.
As for reliability, dlt observes 99.83% run success across 50m recent open source production runs.
Stop operating a platform
When a source changes on dlthub, the agent reads the logs, writes the fix, tests it in dev, and waits for your OK, usually in under ten minutes. The fix does not wait in a community PR queue or vendor backlog; it happens in your workspace, on your schedule. Watch the 3-minute maintenance video.
This matches what we observe in the wild: setups with 30–50 pipelines maintained by one professional. There is no VM or Kubernetes to keep running, and no part-time devops engineer feeding the deployment. You are billed for the minutes a pipeline actually runs. Nothing idles between syncs.
Detach the billing meter from volume
If you are on the managed offering, you want control over your bill. The fix is simple: meter the cost of operation, not the outcome. Pay for infrastructure, not rows or gigabytes — the same way you already pay Snowflake, Databricks, or your cloud.
A simple example: we benchmarked what one hour of dltHub can move. An hour of dltHub costs around $1. Here is the cost of moving the same data on Airbyte:
| Source | $1 of dltHub moves | On Airbyte* |
|---|---|---|
| Parquet | ~170 GB · 1.1B rows | ~$1,700+ |
| Postgres | ~65 GB · 350M rows | ~$650 |
| JSON | ~4.6 GB · 47M rows | ~$46 |
| REST APIs | rate-limit bound | ~$3.40 |
*estimates based on public airbyte pricing of $10/gb or $15/1m rows for apis.
The takeaway: on rate-limited APIs the difference is small (~3x). On bulk data, paying for compute instead of volume cuts the running cost by 99%+.
$100/connector/year ongoing TCO
dlt pipelines in the wild are observed to be 99.83% successful. This translates into a maintenance event every 2.2 years. With agentic maintenance, this cost drops further. Overall, it averages out to around $100 worth of human labor and a few dollars in LLM tokens ($2–5).
Read more about the dltHub total cost of ownership here.
The cost of moving
First, an offer for assistance: bring us the stack you want to move. When we have capacity, we can move 30 pipelines over a two-week timeline.
If you have the talent in-house, senior engineers report moving at around the rate of 1 connector per day.
Why dltHub, not dlt plus Claude skills?
Many consultancies already run migrations with open-source dlt and their own Claude skills. That works, and it takes a good platform engineer to operationalize well. dltHub makes the outcome a property of the platform instead of the person: schemas, lineage, traces, transformations, quality checks, and notebooks in one place — the record the agent works from.
- Lower TCO — agentic maintenance reads the logs, fixes the code, tests in dev, and reports back before asking for deployment confirmation.
- Anyone can operate it — the context travels with the pipeline, not with the engineer who built it.
- Best practice by default — we maintain the skills, so pipelines run as intended by the manufacturer.
- Safe by design — the agent reads prod, tests in local dev, and ships only with your approval.
With DIY skills, each of these is a maybe. Skills can be rebuilt in an afternoon; the context only exists if something has been collecting it all along.
Ready to move?
Are you a practitioner? Try dltHub free: Ask your agent to run uvx dlthub-start@latest and rebuild the connector that alerts you most.
Moving over 5-10 pipelines and need a hand? Book a call with our team.