dlt

Overview

dlt is the odd one out in this category, and deliberately so. Its own README says it plainly: a library, not a platform. There is no server, no control plane and no interface to deploy.

Pipelines are Python. You define a source, name a destination and dataset, and call run. Resources are generators decorated in code, declaring write disposition, primary keys and schema contracts alongside the extraction logic.

That means it runs wherever Python runs, inside Airflow, Dagster, a Lambda, a notebook or a laptop.

The licence is genuinely permissive. The library is Apache 2.0 with commercial use permitted, the only core in this category under an approved open-source licence rather than a source-available one.

Schema handling is the technical draw. New columns appear automatically, type changes create variant columns, and schema contracts let you set evolve, freeze, discard rows or discard columns at table, column and type level.

Connector strategy is the opposite of Fivetran's. Rather than a large catalogue there are about 30 verified sources plus three generic toolkits: a REST API source that needs no custom connector, a SQL source covering any SQLAlchemy database, and a filesystem source.

Merge strategies are unusually complete for a library, including a real type 2 slowly changing dimension mode that uses a row hash as a surrogate key and works on partial extracts.

Destinations lean towards AI as well as analytics, with LanceDB, Qdrant, Weaviate and Hugging Face alongside the usual warehouses, lakes and table formats.

The commercial layer is dltHub, renamed from dlt+, adding a managed runtime, orchestration, transformations in preview, a context catalogue and an MCP server so coding agents can inspect pipelines directly.

Two things to weigh. Governance disclosure is thin, with no named certifications published at all, and the commercial floor is steep, with nothing published between free and $12,000 a month.

Key features and capabilities

The same headings are used for every elt & ingestion entry, so two tools can be read side by side.

Connectors and coverage

• About 30 verified sources including Salesforce, Stripe, HubSpot, Notion, Zendesk, Shopify and Kafka
• A REST API source connects to any REST API without a custom connector being written
• A SQL source covers any SQLAlchemy-supported database, more than 30 of them, through one implementation
• It also ingests Python data structures, DataFrames and Arrow tables directly, which its rivals do not
• No certified or community support tiers are published

How pipelines are built

• Python only, using decorators and generators; there is no interface, YAML or SQL authoring
• Orchestration is not included in the library; you bring Airflow, Dagster or cron
• dltHub adds a managed runtime with code-defined schedules, dependencies and freshness gates
• Transformations in Python and SQL are a dltHub addition, in public preview

Loading and sync behaviour

• Write dispositions of append, replace and merge, with four merge strategies
• Delete-insert is the default, staging and deduplicating within a single atomic transaction
• A true type 2 slowly changing dimension mode, using a row hash as a surrogate key, working on partial extracts
• Upsert and insert-only strategies for speed and for immutable event data
• A Postgres replication source gives logical-replication change capture; there is no general managed CDC engine

Destinations

• Warehouses including BigQuery, Snowflake, Redshift, Azure Synapse and Microsoft Fabric
• Databases including Postgres, SQL Server, ClickHouse, DuckDB and MotherDuck
• Lakes and table formats including Delta Lake, Iceberg, Databricks, Dremio, DuckLake and Athena
• AI destinations including LanceDB, Qdrant, Weaviate and Hugging Face datasets
• A custom destination function covers reverse ETL

Security, monitoring and lineage

• No named compliance certifications are published on any first-party page
• A 2026 post argues for auditable ingestion suitable for regulated life sciences, which is readiness not certification
• Schema contracts are the real governance primitive, preventing unexpected structural change
• The enterprise tier lists role-based access control, audit logs and service levels without documenting them
• Encryption, PII handling and data residency are not published

Extending it yourself

• Writing a source is the normal path, a Python generator plus a decorator, not an escape hatch
• Custom destinations are supported, including for reverse ETL
• An MCP server lets coding agents analyse pipelines and datasets directly
• More than 40 guided agent skills, optimised for Claude, Codex and Cursor
• Free fundamentals and advanced courses, delivered as notebooks rather than video

Where it runs and what it costs

• The library runs anywhere Python does, with no server, control plane or licence key
• dltHub adds a managed runtime deployed with a single command, and a managed Iceberg lakehouse
• Bring your own storage is the published data sovereignty lever
• Regions for the managed runtime are not published, and self-hosted dltHub is not offered

Pricing

Open sourceFree library; dltHub from $12,000 a month

The library is Apache 2.0, free, with commercial use permitted and no volume limits. dltHub, the commercial platform, is $12,000 a month on a twelve-month minimum, including 5,000 credits, with overage from $1.00 a credit falling to $0.80 and a further 5% off for annual commitment. Purchased credits roll over; included ones do not. A 14-day trial gives $30 of credit with no card. That gap is the sharpest in this category: there is no published middle tier between free and $144,000 a year.

Vendor pricing page →

Demos and videos

• Build faster in the dltHub workspace Product demo · YouTube
• dlt fundamentals course Tutorial · dlthub.com

About dltHub

dltHub is based in Berlin with offices in San Francisco, Montreal and Amsterdam, and was founded by Matthaus Krzykowski, Marcin Rudolf, Anna Hoffmann and Adrian Brudaru; no founding year is published. It is private and backed by Bessemer Venture Partners, Foundation Capital and Dig Ventures with angels from MotherDuck, Hugging Face and Miro, though no round amount or date appears on its own pages. It publishes adoption instead: more than 3,000 open-source users in production as of April 2025 and over 50,000 developers. The library is Apache 2.0; the commercial platform is under a licence named only as the dltHub License.

Berlin, Germany · dlthub.com