Do you actually need a separate ETL tool? Fivetran, Python scripts, and warehouse-native ingestion compared

TL;DR

• Data teams are increasingly questioning standalone ETL vendors like Fivetran due to usage-based pricing shocks, connector limitations, and rigid pipelines.
• Three main data ingestion architectures exist today: managed ETL vendors (Buy), homegrown Python and open-source ELT (Build), and warehouse-native ingestion (Consolidate).
• Warehouse-native ingestion runs custom Python logic directly inside the data warehouse, combining Python flexibility with no infrastructure to manage.
• MotherDuck Flights offers a Python-first, agent-native ingestion approach natively integrated into MotherDuck's modern cloud data warehouse, eliminating the need to manage separate servers, schedulers, or secrets.

For teams that write Python and already use a data warehouse, the case for a standalone ETL vendor has weakened considerably. Fivetran is the dominant player in this space, and for many teams it was the default first choice. This article compares three architectural patterns: managed ETL vendors (like Fivetran), homegrown Python scripts, and warehouse-native ingestion platforms (like MotherDuck Flights), helping you determine which option is the best fit for your team.

Why teams start questioning their ETL vendors like Fivetran

Data teams rarely switch tools without a compelling reason. Fivetran has long been the default choice for managed ETL, but predictable inflection points emerge as a company's data stack matures, and teams start evaluating alternatives.

Pricing surprises: Cost is the most common catalyst. Usage-based pricing models, like Fivetran's Monthly Active Rows (MAR), can lead to unpredictable bills as data volume scales.

Teams also encounter a double vendor premium problem: paying high margins for managed extraction compute in an external ETL platform while simultaneously paying for transformation compute in the cloud data warehouse. Recent structural changes, such as a per-connector minimum charge, have led some customers to reduce connector counts or switch vendors.

Connector gaps: Managed vendors offer hundreds of connectors, yet a business may depend on a niche SaaS tool or internal API lacking a pre-built integration. This forces the team to build a custom Python script anyway. The resulting hybrid system defeats the purpose of a single managed solution.

Ownership and complexity: Adding another vendor means another contract, security review, bill, and system to monitor. Consolidating the data stack reduces this administrative and operational overhead.

Limited flexibility: Pre-built connectors are powerful but often rigid. They cannot always handle:

• Non-standard APIs
• Bespoke pagination
• Custom authentication schemes
• Proprietary internal databases

These limitations force teams to write scripts to fill gaps their ETL vendor can't cover.

The agent-native shift: As teams adopt AI agents to interact with their data, agents require access to the freshest data. When the ingestion pipeline lives in a separate, closed system, it creates an operational barrier that prevents agents from rapidly drafting, testing, and proposing ingestion logic in sandboxed environments.

When teams hit these breaking points, they face three architectural choices: pay another vendor (Buy), build it from scratch (Build), or run ingestion on their existing compute layer (Consolidate).

The three real approaches to data ingestion

Approach 1: Managed ETL/ELT vendors. Third-party SaaS platforms like Fivetran provide 700+ pre-built connectors, and Stitch offers a broad catalog of pre-built source connectors. These vendors maintain the connectors and sync data on a schedule.

Approach 2: Homegrown Python & Open-Source ELT. The classic build-it-yourself model where teams write custom extraction code or deploy open-source ELT frameworks (like Meltano). While powerful, these approaches require teams to manage substantial self-hosted infrastructure, including EKS/ECS clusters, cloud orchestrators, schedulers (such as cron or Airflow), and secrets managers.

Approach 3: Warehouse-native ingestion. This modern architectural pattern uses the data warehouse's own compute to run ingestion logic directly. Unlike traditional managed file loading or WAL replication tools, MotherDuck Flights runs arbitrary custom Python for ingestion directly inside the warehouse.

MotherDuck Flights is an agent-native data pipelines feature in MotherDuck that allows users to build and deploy data pipelines using a flexible, general-purpose Python runtime. The warehouse natively manages scheduling, secrets, and run history.

Side-by-side comparison: Managed ETL vs. Python vs. warehouse-native

Which approach is right for your team?

There is no single optimal ingestion method. The right architectural choice depends entirely on your team's engineering capacity, source heterogeneity, and operational budget.

1. Managed ETL vendor

A managed vendor like Fivetran is the ideal fit for teams needing to connect to hundreds of different SaaS applications quickly without writing code. This is the standard approach for organizations lacking dedicated Python engineering capacity or those with complex compliance requirements demanding vendor-owned pipeline reliability.

This option is particularly well-suited for teams that need robust, managed micro-batch Change Data Capture (CDC) using Write-Ahead Logging (WAL) for databases like PostgreSQL, without having to build and maintain heavy streaming infrastructure.

The primary trade-offs include:

• Pricing at scale, particularly with models charging per Monthly Active Rows (MAR)
• Limited flexibility for custom logic

2. Homegrown Python & Open-Source ELT

Homegrown scripts are the right choice for teams with:

• A small number of stable sources
• Highly custom internal APIs with no commercial connectors
• A strong Python engineering culture requiring absolute control

This approach works well for early-stage startups avoiding vendor lock-in. The trade-off is significant operational overhead. Someone on the team must own the infrastructure: compute, schedulers, alerting, and secrets. As the number of scripts grows, technical debt and maintenance hours scale with them.

Even if teams use open-source data engineering frameworks like Airbyte to avoid writing from scratch, they still take on the burden of self-hosting heavy infrastructure like EKS/ECS.

Using declarative Python libraries like dlt simplifies pipeline logic, but in a homegrown setup, teams must still manage the compute and scheduling layers themselves. Near real-time CDC replication from databases requires heavy streaming infrastructure (like Debezium and Kafka), which is a massive operational burden.

3. Warehouse-native ingestion (e.g., MotherDuck Flights)

This approach is ideal for teams that are comfortable writing Python, want to consolidate their data stack around MotherDuck, and prefer predictable, budget-friendly MotherDuck compute costs over the expense of a separate ETL vendor.

Warehouse-native ingestion is well-suited for:

• Custom extraction using declarative Python libraries like dlt for ready-made connectors
• Use cases where high-watermark polling is sufficient for database syncs
• MotherDuck Flights acting as the compute and scheduling engine

Flights is the only option here with native agent support. Teams wanting an AI agent to rapidly draft, test, and propose ingestion logic in sandboxed environments can do so natively via the MotherDuck MCP server.

This approach is not for:

• Teams needing 200+ pre-built, zero-code connectors out-of-the-box
• Teams with zero Python capacity
• Teams requiring massive-scale, WAL-based CDC for near real-time replica streams

Conclusion

Data ingestion has evolved beyond a simple build-versus-buy decision. For teams evaluating Fivetran alternatives, three distinct paths now exist:

• Managed ETL for zero-code breadth and managed micro-batch CDC
• Homegrown Python and open-source ELT for total control
• Warehouse-native ingestion for flexible logic without the infrastructure overhead

Ready to see warehouse-native ingestion in action? Try MotherDuck Flights, available on all plans including free trials.