Access loaded data in Python#

This guide explains how to access and manipulate data that has been loaded into your destination using the dlt Python library. After running your pipelines and loading data, you can use the pipeline.dataset() and data frame expressions, Ibis or SQL to query the data and read it as records, Pandas frames or Arrow tables.

Quick start example#

Here's a full example of how to retrieve data from a pipeline and load it into a Pandas DataFrame or a PyArrow Table.

Getting started#

Assuming you have a Pipeline object (let's call it pipeline), you can obtain a Dataset which contains the credentials and schema to your destination dataset. You can construct a query and execute it on the dataset to retrieve a Relation which you may use to retrieve data from the Dataset.

Note: The Dataset and Relation objects are lazy-loading. They will only query and retrieve data when you perform an action that requires it, such as fetching data into a DataFrame or iterating over the data. This means that simply creating these objects does not load data into memory, making your code more efficient.

Access the dataset#

Access tables as dataset#

The simplest way of getting a Relation from a Dataset is to get a full table relation:

Creating relations with sql query strings#

Reading data#

Once you have a Relation, you can read data in various formats and sizes.

Fetch the entire table#

As a Pandas DataFrame#

As a PyArrow Table#

As a list of Python tuples#

Lazy loading behavior#

The Dataset and Relation objects are lazy-loading. This means that they do not immediately fetch data when you create them. Data is only retrieved when you perform an action that requires it, such as calling .df(), .arrow(), or iterating over the data. This approach optimizes performance and reduces unnecessary data loading.

Iterating over data in chunks#

To handle large datasets efficiently, you can process data in smaller chunks.

Iterate as Pandas DataFrames#

Iterate as PyArrow Tables#

Iterate as lists of tuples#

The methods available on the Relation correspond to the methods available on the cursor returned by the SQL client. Please refer to the SQL client guide for more information.

Connection Handling#

For every call that actually fetches data from the destination, such as df(), arrow(), fetchall() etc., the dataset will open a connection and close it after it has been retrieved or the iterator is completed. You can keep the connection open for multiple requests with the dataset context manager:

Special queries#

You can use the row_counts method to get the row counts of all tables in the destination as a DataFrame.

Modifying queries#

You can refine your data retrieval by limiting the number of records, selecting specific columns, sorting the results, filtering rows, aggregating minimum and maximum values on a specific column, or chaining these operations.

Limit the number of records#

Using head() to get the first 5 records#

Select specific columns#

Sort results#

Filter rows#

Aggregate data#

Filter to an incremental cursor#

Relation.incremental(incremental) adds a WHERE clause derived from a dlt.sources.incremental cursor so a relation only sees rows in the cursor window.

import dlt
from dlt.common.pendulum import pendulum

dataset = pipeline.dataset()

# bounded read: all rows in [2026-01-01, 2026-02-01)
cursor = dlt.sources.incremental(
    "created_at",
    initial_value=pendulum.datetime(2026, 1, 1, tz="UTC"),
    end_value=pendulum.datetime(2026, 2, 1, tz="UTC"),
)
rows = dataset.table("events").incremental(cursor).fetchall()

Or pass it directly on dataset.table(..., incremental=...):

rows = dataset.table("events", incremental=cursor).fetchall()

Relation.incremental() accepts cursor paths in two forms:

• column — filters on a column of the relation's base table.
• table.column — automatically joins table via the dataset schema and filters on the joined column. The joined table's columns are not added to the projection. If the same table is already joined, the existing join is reused.

Cursor on an auto-joined column#

A dotted cursor_path of the form table.column auto-joins table and filters on the joined column. This uses the same schema-reference resolution as Relation.join() — table must be reachable from the current relation's base table via dlt's parent/child references. The joined columns are not added to the projection, and an existing JOIN to the same table is reused.

A common case is filtering any user table by dlt load time via _dlt_loads:

# only rows from loads that happened after 2026-01-01
cursor = dlt.sources.incremental(
    "_dlt_loads.inserted_at",
    initial_value=pendulum.datetime(2026, 1, 1, tz="UTC"),
)
events = dataset.table("events", incremental=cursor)

The translation from Incremental to SQL follows these rules:

• last_value_func must be max or min. Custom callables can't be pushed down to SQL.
• range_start / range_end decide endpoint inclusivity ("closed" -> >=/<=, "open" -> >/<); operator direction follows last_value_func.
• on_cursor_value_missing="include" translates to ... OR cursor IS NULL; "exclude" to ... AND cursor IS NOT NULL. "raise" cannot raise mid-query in SQL pushdown, so it falls back to IS NOT NULL and emits a warning when the cursor column is nullable.
• lag is applied to the lower bound exactly as it would be during a resource extraction.

See Incremental transformations for using this in @dlt.hub.transformation, including stateful cursors, scheduler-owned windows, and _dlt_loads.inserted_at load-time cursors.

Join related tables#

The join() method follows relationships already defined in the dlt schema. It can resolve direct schema references between tables as well as multi-hop parent/child paths when one table is an ancestor or descendant of the other. This makes join() well suited for navigating nested tables created by dlt and tables connected by explicit references. Joined columns are appended from the target table only and are prefixed with the target table name, or with the alias you provide.

By default, join() creates an inner join. Use kind="left", "right", or "full" to choose another SQL join type.

When you do not specify an alias, joined columns use the joined table name as their prefix. For example, dataset["users"].join("users__orders") adds columns such as users__orders__order_id. When you pass alias="orders", the same column is projected as orders__order_id instead. Use alias to make result columns easier to read or to avoid output name conflicts.

Limits: join() only works when dlt can resolve a supported schema-defined path between the current relation's base table and the target table. Both sides must be base-table relations, for example dataset["users"].join("users__orders"). You cannot call join() after transforming a relation with methods such as select() or where().

join() does not support:

• arbitrary join conditions
• joins on columns that are not defined as schema references
• self-joins
• joins across different datasets
• joins between tables that are only related indirectly through a shared ancestor or another non-linear schema path

In practice, this means join() supports ancestor/descendant navigation, but not general graph traversal across the schema.

For example:

• dataset["users__orders__items"].join("users") works because users is an ancestor in the nested table hierarchy
• joining two sibling tables just because both descend from users does not work
• joining two tables on a custom predicate such as orders.customer_email = customers.email does not work unless that relationship is defined in the schema

When join() needs intermediate tables to reach the target, those tables are used only to build the join path. Their columns are not added to the result automatically. Only columns from the explicitly joined target table are appended.

For arbitrary join logic, use Ibis.

Chain operations#

You can combine select, limit, and other methods.

Modifying queries with ibis expressions#

If you install the amazing ibis library, you can use ibis expressions to modify your queries.

pip install ibis-framework

dlt will then allow you to get an ibis.Table for each table which you can use to build a query with ibis expressions, which you can then execute on your dataset.

You can learn more about the available expressions on the ibis for sql users page.

Migrating from the previous dlt / ibis implementation#

As describe above, the new way to use ibis expressions is to first get one or many Table objects and construct your expression. Then, you can pass it Dataset to get a Relation to execute the full query and retrieve data.

An example from our previous docs for joining a customers and a purchase table was this:

# get two relations
customers_relation = dataset["customers"]
purchases_relation = dataset["purchases"]

# join them using an ibis expression
joined_relation = customers_relation.join(
    purchases_relation, customers_relation.id == purchases_relation.customer_id
)

# ... do other ibis operations

# directly fetch the data on the expression we have built
df = joined_relation.df()

The migrated version looks like this:

# we convert the dlt.Relation an Ibis Table object
customers_expression = dataset.table("customers").to_ibis()
purchases_expression = dataset.table("purchases").to_ibis()

# join them using an ibis expression, same code as above
joined_epxression = customers_expression.join(
    purchases_expression, customers_expression.id == purchases_expression.customer_id
)

# ... do other ibis operations, would be same as before

# now convert the expression to a relation
joined_relation = dataset(joined_epxression)

# execute as before
df = joined_relation.df()

Supported destinations#

All SQL and filesystem destinations supported by dlt can utilize this data access interface.

Reading data from filesystem#

For filesystem destinations, dlt uses DuckDB under the hood to create views on iceberg and delta tables or from Parquet, JSONL and csv files. This allows you to query data stored in files using the same interface as you would with SQL databases. If you plan on accessing data in buckets or the filesystem a lot this way, it is advised to load data into delta or iceberg tables, as DuckDB is able to only load the parts of the data actually needed for the query to work.

Examples#

Fetch one record as a tuple#

Fetch many records as tuples#

Iterate over data with limit and column selection#

Note: When iterating over filesystem tables, the underlying DuckDB may give you a different chunk size depending on the size of the parquet files the table is based on.

Advanced usage#

Loading a Relation into a pipeline table#

Since the iter_arrow and iter_df methods are generators that iterate over the full Relation in chunks, you can use them as a resource for another (or even the same) dlt pipeline:

Learn more about transforming data in Python with Arrow tables or DataFrames.

Datasets with multiple schemas#

When a pipeline loads data from several sources, each source produces its own schema. By default, all schemas share one physical dataset and pipeline.dataset() includes every schema automatically, so tables from all sources are queryable together. If two schemas define a table with the same name, dlt merges their columns and combines rows from both — missing columns are filled with NULL.

Breaking changes#

:::caution Breaking changes introduced in dlt 1.25.0
The following changes affect existing code that uses pipeline.dataset():

pipeline.dataset() now includes all schemas by default. Previously, calling pipeline.dataset() without a schema argument returned only the default schema's tables. Now, when use_single_dataset is enabled (the default) and the pipeline has multiple schemas, all schemas are included automatically. Code that assumed only one schema's tables are visible may now see additional tables or extra rows in shared table names. To restore the previous single-schema behavior, pass the schema explicitly:

# Before (implicit single schema):
ds = pipeline.dataset()

# After (explicit single schema, equivalent to the old behavior):
ds = pipeline.dataset(schema=pipeline.default_schema_name)

:::

Staging dataset#

So far, we've been using the append write disposition in our example pipeline. This means that each time we run the pipeline, the data is appended to the existing tables. When you use the merge write disposition, dlt creates a staging database schema for staging data. This schema is named <dataset_name>_staging by default and contains the same tables as the destination schema. When you run the pipeline, the data from the staging tables is loaded into the destination tables in a single atomic transaction.

Let's illustrate this with an example. We change our pipeline to use the merge write disposition:

import dlt

@dlt.resource(primary_key="id", write_disposition="merge")
def users():
    yield [
        {'id': 1, 'name': 'Alice 2'},
        {'id': 2, 'name': 'Bob 2'}
    ]

pipeline = dlt.pipeline(
    pipeline_name='quick_start',
    destination='duckdb',
    dataset_name='mydata'
)

load_info = pipeline.run(users)

Running this pipeline will create a schema in the destination database with the name mydata_staging.
If you inspect the tables in this schema, you will find the mydata_staging.users table identical to the mydata.users table in the previous example.

Here is what the tables may look like after running the pipeline:

mydata_staging.users

mydata.users

Notice that the mydata.users table now contains the data from both the previous pipeline run and the current one.

dev_mode (versioned datasets)#

When you set the dev_mode argument to True in the dlt.pipeline call, dlt creates a versioned dataset.
This means that each time you run the pipeline, the data is loaded into a new dataset (a new database schema).
The dataset name is the same as the dataset_name you provided in the pipeline definition with a datetime-based suffix.

We modify our pipeline to use the dev_mode option to see how this works:

import dlt

data = [
    {'id': 1, 'name': 'Alice'},
    {'id': 2, 'name': 'Bob'}
]

pipeline = dlt.pipeline(
    pipeline_name='quick_start',
    destination='duckdb',
    dataset_name='mydata',
    dev_mode=True # <-- add this line
)
load_info = pipeline.run(data, table_name="users")

Every time you run this pipeline, a new schema will be created in the destination database with a datetime-based suffix. The data will be loaded into tables in this schema.
For example, the first time you run the pipeline, the schema will be named mydata_20230912064403, the second time it will be named mydata_20230912064407, and so on.

Internal dlt tables#

dlt automatically creates internal tables in the destination schema to track pipeline runs, support incremental loading, and manage schema versions. These tables use the _dlt_ prefix.

_dlt_loads#

This table records each pipeline run. Every time you execute a pipeline, a new row is added to this table with a unique load_id. This table tracks which loads have been completed and supports chaining of transformations.

Only rows with status = 0 are considered complete. Other values represent incomplete or interrupted loads. The status column can also be used to coordinate multi-step transformations.

_dlt_pipeline_state#

This table stores the internal state of the pipeline for each run. This state enables incremental loading and allows the pipeline to resume from where it left off if a previous run was interrupted.

The state column contains a serialized Python dictionary that includes:

- Incremental progress (e.g. last item or timestamp processed).
- Checkpoints for transformations.
- Source-specific metadata and settings.

This allows dlt to resume interrupted pipelines, avoid reloading already processed data, and ensure pipelines are idempotent and efficient.

The version_hash is recalculated on each update. dlt uses this table to implement last-value incremental loading. If a run fails or stops, this table ensures the next run picks up from the correct checkpoint.

_dlt_version#

This table tracks the history of all schema versions used by the pipeline. Every time dlt updates the schema. For example, when new columns or tables are added, a new entry is written to this table.

By keeping previous schema definitions, _dlt_version ensures that:

• Older data remains readable
• New data uses updated schema rules
• Backward compatibility is maintained

This table also supports troubleshooting and compatibility checks. It lets you track which schema and engine version were used for any load. This helps with debugging and ensures safe evolution of your data model.

Ibis#

Ibis is a powerful portable Python dataframe library. Learn more about what it is and how to use it in the official documentation.

dlt provides an easy way to hand over your loaded dataset to an Ibis backend connection.

Prerequisites#

To use the Ibis backend, you will need to have the ibis-framework package with the correct Ibis extra installed. The following example will install the DuckDB backend:

pip install ibis-framework[duckdb]

Get an Ibis connection from your dataset#

dlt datasets have a helper method to return an Ibis connection to the destination they live on. The returned object is a native Ibis connection to the destination, which you can use to read and even transform data. Please consult the Ibis documentation to learn more about what you can do with Ibis.

:::caution Breaking change in dlt 1.25.0
dataset.ibis() now passes all schemas from the dataset to the Ibis backend. On filesystem destinations, this means Ibis will see tables from every schema in the dataset and not just the default one. If two schemas define the same table name, the Ibis table will contain rows from both schemas combined. To get the previous single-schema behavior, create the dataset with an explicit schema: pipeline.dataset(schema="my_schema").ibis().
:::

# get the dataset from the pipeline
dataset = pipeline.dataset()
dataset_name = pipeline.dataset_name

# get the native ibis connection from the dataset
ibis_connection = dataset.ibis()

# list all tables in the dataset
# NOTE: You need to provide the dataset name to ibis, in ibis datasets are named databases
print(ibis_connection.list_tables(database=dataset_name))

# get the items table
table = ibis_connection.table("items", database=dataset_name)

# print the first 10 rows
print(table.limit(10).execute())

# Visit the ibis docs to learn more about the available methods

Marimo#

marimo is a reactive Python notebook. It completely revamps the Jupyter notebook experience. Whenever code is executed or you interact with a UI element, dependent cells are re-executed ensuring consistency between code and displayed outputs.

This page shows how dlt + marimo + ibis provide a rich environment to explore loaded data, write data transformations, and create data applications.

Prerequisites#

To install marimo and ibis with the duckdb extras, run the following command:

pip install marimo "ibis-framework[duckdb]"

Launch marimo#

Use this command to launch marimo (replace my_notebook.py with desired name). It will print a link to access the notebook web app.

marimo edit my_notebook.py

> Edit my_notebook.py in your browser 📝
> ➜ URL: http://localhost:2718?access_token=Qfo_Hj2RbXqiqM4VT3XOwA

Here's a screenshot of the interface you should see:

Features#

Use custom dlt widgets#

Inside your marimo notebook, you can use composable widgets built and maintained by the dlt team. This requires the mowidgets package (Python 3.11+).

Import them from dlt.helpers.marimo and pass them to the render() function:

#%% cell 1
from dlt.helpers.marimo import render, load_package_viewer, pipeline_selector

#%% cell 2
render(pipeline_selector)

#%% cell 3
render(load_package_viewer, pipeline_path="/path/to/pipeline")

Available widgets: pipeline_selector, load_package_viewer, schema_viewer.

View dataset tables and columns#

After loading data with dlt, you can access it via the dataset interface, including a native ibis connection.

In marimo, the Datasources panel provides a GUI to explore data tables and columns. When a cell contains a variable that's an ibis connection, it is automatically registered.

Accessing data with SQL#

Clicking on the Add table to notebook button will create a new SQL cell that you can use to query data. The output cell provides a rich and interactive results dataframe.

Accessing data with Python#

You can also retrieve Ibis tables (lazy expressions) using Python. The Datasources panel will show under Python the output schema of your Ibis query, and the cell output will display detailed query planning.

Use .execute(), .to_pandas(), .to_polars(), or .to_pyarrow() to execute the Ibis expression and retrieve data that can displayed in a rich and interactive dataframe.

Create a dashboard and data apps#

marimo notebooks can be deployed as web applications with interactive UI and charts and the code hidden. Try adding marimo UI input elements, rich markdown, and charts (matplotlib, plotly, altair, etc.). Combined, dlt + marimo + ibis make it easy to build a simple dashboard on top of fresh data.

Further reading#

• Learn about marimo dataframe and SQL features
• Explore databases using the marimo GUI
• Learn about marimo if you're coming from Streamlit

Important considerations#

• Memory usage: Loading full tables into memory without iterating or limiting can consume significant memory, potentially leading to crashes if the dataset is large. Always consider using limits or chunked iteration.

• Lazy evaluation: Dataset and Relation objects delay data retrieval until necessary. This design improves performance and resource utilization.

• Custom SQL queries: When executing custom SQL queries, remember that additional methods like limit() or select() won't modify the query. Include all necessary clauses directly in your SQL statement.