diff --git a/dlt/__init__.py b/dlt/__init__.py
index e8a1b7bf92..328817efd2 100644
--- a/dlt/__init__.py
+++ b/dlt/__init__.py
@@ -42,7 +42,6 @@
 )
 from dlt.pipeline import progress
 from dlt import destinations
-from dlt.destinations.dataset import dataset as _dataset
 
 pipeline = _pipeline
 current = _current
@@ -80,7 +79,6 @@
     "TCredentials",
     "sources",
     "destinations",
-    "_dataset",
 ]
 
 # verify that no injection context was created
diff --git a/dlt/common/destination/reference.py b/dlt/common/destination/reference.py
index 048fe2186f..827034ddca 100644
--- a/dlt/common/destination/reference.py
+++ b/dlt/common/destination/reference.py
@@ -592,6 +592,10 @@ def __getattr__(self, table: str) -> SupportsReadableRelation: ...
 
     def ibis(self) -> IbisBackend: ...
 
+    def row_counts(
+        self, *, data_tables: bool = True, dlt_tables: bool = False, table_names: List[str] = None
+    ) -> SupportsReadableRelation: ...
+
 
 class JobClientBase(ABC):
     def __init__(
diff --git a/dlt/destinations/dataset/dataset.py b/dlt/destinations/dataset/dataset.py
index e443045e49..fc55393a60 100644
--- a/dlt/destinations/dataset/dataset.py
+++ b/dlt/destinations/dataset/dataset.py
@@ -1,4 +1,4 @@
-from typing import Any, Union, TYPE_CHECKING
+from typing import Any, Union, TYPE_CHECKING, List
 
 from dlt.common.json import json
 
@@ -133,6 +133,32 @@ def table(self, table_name: str) -> SupportsReadableRelation:
             table_name=table_name,
         )  # type: ignore[abstract]
 
+    def row_counts(
+        self, *, data_tables: bool = True, dlt_tables: bool = False, table_names: List[str] = None
+    ) -> SupportsReadableRelation:
+        """Returns a dictionary of table names and their row counts, returns counts of all data tables by default"""
+        """If table_names is provided, only the tables in the list are returned regardless of the data_tables and dlt_tables flags"""
+
+        selected_tables = table_names or []
+        if not selected_tables:
+            if data_tables:
+                selected_tables += self.schema.data_table_names(seen_data_only=True)
+            if dlt_tables:
+                selected_tables += self.schema.dlt_table_names()
+
+        # Build UNION ALL query to get row counts for all selected tables
+        queries = []
+        for table in selected_tables:
+            queries.append(
+                f"SELECT '{table}' as table_name, COUNT(*) as row_count FROM"
+                f" {self.sql_client.make_qualified_table_name(table)}"
+            )
+
+        query = " UNION ALL ".join(queries)
+
+        # Execute query and build result dict
+        return self(query)
+
     def __getitem__(self, table_name: str) -> SupportsReadableRelation:
         """access of table via dict notation"""
         return self.table(table_name)
diff --git a/dlt/pipeline/pipeline.py b/dlt/pipeline/pipeline.py
index 9bd2d6911f..74466a09e4 100644
--- a/dlt/pipeline/pipeline.py
+++ b/dlt/pipeline/pipeline.py
@@ -1750,7 +1750,7 @@ def __getstate__(self) -> Any:
         # pickle only the SupportsPipeline protocol fields
         return {"pipeline_name": self.pipeline_name}
 
-    def _dataset(
+    def dataset(
         self, schema: Union[Schema, str, None] = None, dataset_type: TDatasetType = "auto"
     ) -> SupportsReadableDataset:
         """Returns a dataset object for querying the destination data.
diff --git a/docs/website/docs/build-a-pipeline-tutorial.md b/docs/website/docs/build-a-pipeline-tutorial.md
index f85d2e19ea..36d30a184f 100644
--- a/docs/website/docs/build-a-pipeline-tutorial.md
+++ b/docs/website/docs/build-a-pipeline-tutorial.md
@@ -262,20 +262,30 @@ In this example, the first pipeline loads the data using `pipedrive_source()`. T
 
 #### [Using the `dlt` SQL client](dlt-ecosystem/transformations/sql.md)
 
-Another option is to leverage the `dlt` SQL client to query the loaded data and perform transformations using SQL statements. You can execute SQL statements that change the database schema or manipulate data within tables. Here's an example of inserting a row into the `customers` table using the `dlt` SQL client:
+Another option is to leverage the `dlt` SQL client to query the loaded data and perform transformations using SQL statements. You can execute SQL statements that change the database schema or manipulate data within tables. Here's an example of creating a new table with aggregated sales data in duckdb:
 
 ```py
-pipeline = dlt.pipeline(destination="bigquery", dataset_name="crm")
+pipeline = dlt.pipeline(destination="duckdb", dataset_name="crm")
 
 with pipeline.sql_client() as client:
     client.execute_sql(
-        "INSERT INTO customers VALUES (%s, %s, %s)", 10, "Fred", "fred@fred.com"
-    )
+        """ CREATE TABLE aggregated_sales AS
+            SELECT 
+                category,
+                region,
+                SUM(amount) AS total_sales,
+                AVG(amount) AS average_sales
+            FROM 
+                sales
+            GROUP BY 
+                category, 
+                region;
+    """)
 ```
 
 In this example, the `execute_sql` method of the SQL client allows you to execute SQL statements. The statement inserts a row with values into the `customers` table.
 
-#### [Using Pandas](dlt-ecosystem/transformations/pandas.md)
+#### [Using Pandas](dlt-ecosystem/transformations/python.md)
 
 You can fetch query results as Pandas data frames and perform transformations using Pandas functionalities. Here's an example of reading data from the `issues` table in DuckDB and counting reaction types using Pandas:
 
@@ -287,11 +297,8 @@ pipeline = dlt.pipeline(
     dev_mode=True
 )
 
-with pipeline.sql_client() as client:
-    with client.execute_query(
-        'SELECT "reactions__+1", "reactions__-1", reactions__laugh, reactions__hooray, reactions__rocket FROM issues'
-    ) as cursor:
-        reactions = cursor.df()
+# get a dataframe of all reactions from the dataset
+reactions = pipeline.dataset().issues.select("reactions__+1", "reactions__-1", "reactions__laugh", "reactions__hooray", "reactions__rocket").df()
 
 counts = reactions.sum(0).sort_values(0, ascending=False)
 ```
diff --git a/docs/website/docs/dlt-ecosystem/destinations/duckdb.md b/docs/website/docs/dlt-ecosystem/destinations/duckdb.md
index 2b284e991a..a4537195ff 100644
--- a/docs/website/docs/dlt-ecosystem/destinations/duckdb.md
+++ b/docs/website/docs/dlt-ecosystem/destinations/duckdb.md
@@ -118,7 +118,7 @@ to disable tz adjustments.
 
 ## Destination configuration
 
-By default, a DuckDB database will be created in the current working directory with a name `<pipeline_name>.duckdb` (`chess.duckdb` in the example above). After loading, it is available in `read/write` mode via `with pipeline.sql_client() as con:`, which is a wrapper over `DuckDBPyConnection`. See [duckdb docs](https://duckdb.org/docs/api/python/overview#persistent-storage) for details.
+By default, a DuckDB database will be created in the current working directory with a name `<pipeline_name>.duckdb` (`chess.duckdb` in the example above). After loading, it is available in **read/write** mode via `with pipeline.sql_client() as con:`, which is a wrapper over `DuckDBPyConnection`. See [duckdb docs](https://duckdb.org/docs/api/python/overview#persistent-storage) for details. If you want to **read** data, use [pipeline.dataset()](../../general-usage/dataset-access/dataset) instead of `sql_client`.
 
 The `duckdb` credentials do not require any secret values. [You are free to pass the credentials and configuration explicitly](../../general-usage/destination.md#pass-explicit-credentials). For example:
 ```py
diff --git a/docs/website/docs/dlt-ecosystem/transformations/dbt/dbt.md b/docs/website/docs/dlt-ecosystem/transformations/dbt/dbt.md
index 449f8b8bde..59eb340ef2 100644
--- a/docs/website/docs/dlt-ecosystem/transformations/dbt/dbt.md
+++ b/docs/website/docs/dlt-ecosystem/transformations/dbt/dbt.md
@@ -1,10 +1,10 @@
 ---
-title: Transform the data with dbt
+title: Transforming data with dbt
 description: Transforming the data loaded by a dlt pipeline with dbt
 keywords: [transform, dbt, runner]
 ---
 
-# Transform the data with dbt
+# Transforming data with dbt
 
 [dbt](https://github.com/dbt-labs/dbt-core) is a framework that allows for the simple structuring of your transformations into DAGs. The benefits of using dbt include:
 
@@ -105,8 +105,8 @@ You can run the example with dbt debug log: `RUNTIME__LOG_LEVEL=DEBUG python dbt
 
 ## Other transforming tools
 
-If you want to transform the data before loading, you can use Python. If you want to transform the data after loading, you can use dbt or one of the following:
+If you want to transform your data before loading, you can use Python. If you want to transform your data after loading, you can use dbt or one of the following:
 
 1. [`dlt` SQL client.](../sql.md)
-2. [Pandas.](../pandas.md)
+2. [Python with dataframes or arrow tables.](../python.md)
 
diff --git a/docs/website/docs/dlt-ecosystem/transformations/index.md b/docs/website/docs/dlt-ecosystem/transformations/index.md
new file mode 100644
index 0000000000..6c51e8cd8d
--- /dev/null
+++ b/docs/website/docs/dlt-ecosystem/transformations/index.md
@@ -0,0 +1,27 @@
+---
+title: Transforming your data
+description: How to transform your data
+keywords: [datasets, data, access, transformations]
+---
+import DocCardList from '@theme/DocCardList';
+
+# Transforming data
+
+If you'd like to transform your data after a pipeline load, you have 3 options available to you:
+
+* [Using dbt](./dbt/dbt.md) - dlt provides a convenient dbt wrapper to make integration easier.
+* [Using the `dlt` SQL client](./sql.md) - dlt exposes an SQL client to transform data on your destination directly using SQL.
+* [Using Python with DataFrames or Arrow tables](./python.md) - you can also transform your data using Arrow tables and DataFrames in Python.
+
+If you need to preprocess some of your data before it is loaded, you can learn about strategies to:
+
+* [Rename columns.](../../general-usage/customising-pipelines/renaming_columns)
+* [Pseudonymize columns.](../../general-usage/customising-pipelines/pseudonymizing_columns)
+* [Remove columns.](../../general-usage/customising-pipelines/removing_columns)
+
+This is particularly useful if you are trying to remove data related to PII or other sensitive data, you want to remove columns that are not needed for your use case or you are using a destination that does not support certain data types in your source data.
+
+
+# Learn more
+<DocCardList />
+
diff --git a/docs/website/docs/dlt-ecosystem/transformations/pandas.md b/docs/website/docs/dlt-ecosystem/transformations/pandas.md
deleted file mode 100644
index e431313d1c..0000000000
--- a/docs/website/docs/dlt-ecosystem/transformations/pandas.md
+++ /dev/null
@@ -1,42 +0,0 @@
----
-title: Transform the data with Pandas
-description: Transform the data loaded by a dlt pipeline with Pandas
-keywords: [transform, pandas]
----
-
-# Transform the data with Pandas
-
-You can fetch the results of any SQL query as a dataframe. If the destination supports that
-natively (i.e., BigQuery and DuckDB), `dlt` uses the native method. Thanks to this, reading
-dataframes can be really fast! The example below reads GitHub reactions data from the `issues` table and
-counts the reaction types.
-
-```py
-pipeline = dlt.pipeline(
-    pipeline_name="github_pipeline",
-    destination="duckdb",
-    dataset_name="github_reactions",
-    dev_mode=True
-)
-with pipeline.sql_client() as client:
-    with client.execute_query(
-        'SELECT "reactions__+1", "reactions__-1", reactions__laugh, reactions__hooray, reactions__rocket FROM issues'
-    ) as cursor:
-        # calling `df` on a cursor, returns the data as a pandas data frame
-        reactions = cursor.df()
-counts = reactions.sum(0).sort_values(0, ascending=False)
-```
-
-The `df` method above returns all the data in the cursor as a data frame. You can also fetch data in
-chunks by passing the `chunk_size` argument to the `df` method.
-
-Once your data is in a Pandas dataframe, you can transform it as needed.
-
-## Other transforming tools
-
-If you want to transform the data before loading, you can use Python. If you want to transform the
-data after loading, you can use Pandas or one of the following:
-
-1. [dbt.](dbt/dbt.md) (recommended)
-2. [`dlt` SQL client.](sql.md)
-
diff --git a/docs/website/docs/dlt-ecosystem/transformations/python.md b/docs/website/docs/dlt-ecosystem/transformations/python.md
new file mode 100644
index 0000000000..d43f8caaca
--- /dev/null
+++ b/docs/website/docs/dlt-ecosystem/transformations/python.md
@@ -0,0 +1,109 @@
+---
+title: Transforming data in Python with Arrow tables or DataFrames
+description: Transforming data loaded by a dlt pipeline with pandas dataframes or arrow tables
+keywords: [transform, pandas]
+---
+
+# Transforming data in Python with Arrow tables or DataFrames
+
+You can transform your data in Python using Pandas DataFrames or Arrow tables. To get started, please read the [dataset docs](../../general-usage/dataset-access/dataset).
+
+
+## Interactively transforming your data in Python
+
+Using the methods explained in the [dataset docs](../../general-usage/dataset-access/dataset), you can fetch data from your destination into a DataFrame or Arrow table in your local Python process and work with it interactively. This even works for filesystem destinations:
+
+
+The example below reads GitHub reactions data from the `issues` table and
+counts the reaction types.
+
+```py
+pipeline = dlt.pipeline(
+    pipeline_name="github_pipeline",
+    destination="duckdb",
+    dataset_name="github_reactions",
+    dev_mode=True
+)
+
+# get a dataframe of all reactions from the dataset
+reactions = pipeline.dataset().issues.select("reactions__+1", "reactions__-1", "reactions__laugh", "reactions__hooray", "reactions__rocket").df()
+
+# calculate and print out the sum of all reactions
+counts = reactions.sum(0).sort_values(0, ascending=False)
+print(counts)
+
+# alternatively, you can fetch the data as an arrow table
+reactions = pipeline.dataset().issues.select("reactions__+1", "reactions__-1", "reactions__laugh", "reactions__hooray", "reactions__rocket").arrow()
+# ... do transformations on the arrow table
+```
+
+## Persisting your transformed data
+
+Since dlt supports DataFrames and Arrow tables from resources directly, you can use the same pipeline to load the transformed data back into the destination.
+
+
+### A simple example
+
+A simple example that creates a new table from an existing user table but only with columns that do not contain private information. Note that we use the `iter_arrow()` method on the relation to iterate over the arrow table instead of fetching it all at once.
+
+```py
+pipeline = dlt.pipeline(
+    pipeline_name="users_pipeline",
+    destination="duckdb",
+    dataset_name="users_raw",
+    dev_mode=True
+)
+
+# get user relation with only a few columns selected, but omitting email and name
+users = pipeline.dataset().users.select("age", "amount_spent", "country")
+
+# load the data into a new table called users_clean in the same dataset
+pipeline.run(users.iter_arrow(chunk_size=1000), table_name="users_clean")
+```
+
+### A more complex example
+
+The example above could easily be done in SQL. Let's assume you'd like to actually do in Python some Arrow transformations. For this will create a resources from which we can yield the modified Arrow tables. The same is possibly with DataFrames.
+
+```py
+import pyarrow.compute as pc
+
+pipeline = dlt.pipeline(
+    pipeline_name="users_pipeline",
+    destination="duckdb",
+    dataset_name="users_raw",
+    dev_mode=True
+)
+
+# NOTE: this resource will work like a regular resource and support write_disposition, primary_key, etc.
+# NOTE: For selecting only users above 18, we could also use the filter method on the relation with ibis expressions
+@dlt.resource(table_name="users_clean")
+def users_clean():
+    users = pipeline.dataset().users
+    for arrow_table in users.iter_arrow(chunk_size=1000):
+
+        # we want to filter out users under 18
+        age_filter = pc.greater_equal(arrow_table["age"], 18)
+        arrow_table = arrow_table.filter(age_filter)
+
+        # we want to hash the email column
+        arrow_table = arrow_table.append_column("email_hash", pc.sha256(arrow_table["email"]))
+
+        # we want to remove the email column and name column
+        arrow_table = arrow_table.drop(["email", "name"])
+
+        # yield the transformed arrow table
+        yield arrow_table
+
+
+pipeline.run(users_clean())
+```
+
+## Other transforming tools
+
+If you want to transform your data before loading, you can use Python. If you want to transform the
+data after loading, you can use Pandas or one of the following:
+
+1. [dbt.](dbt/dbt.md) (recommended)
+2. [`dlt` SQL client.](sql.md)
+
diff --git a/docs/website/docs/dlt-ecosystem/transformations/sql.md b/docs/website/docs/dlt-ecosystem/transformations/sql.md
index ffd348d1a0..60f3e7f7a5 100644
--- a/docs/website/docs/dlt-ecosystem/transformations/sql.md
+++ b/docs/website/docs/dlt-ecosystem/transformations/sql.md
@@ -1,33 +1,52 @@
 ---
-title: Transform the data with SQL
+title: Transforming data with SQL
 description: Transforming the data loaded by a dlt pipeline with the dlt SQL client
 keywords: [transform, sql]
 ---
 
-# Transform the data using the `dlt` SQL client
+# Transforming data using the `dlt` SQL client
 
 A simple alternative to dbt is to query the data using the `dlt` SQL client and then perform the
-transformations using Python. The `execute_sql` method allows you to execute any SQL statement,
+transformations using SQL statements in Python. The `execute_sql` method allows you to execute any SQL statement,
 including statements that change the database schema or data in the tables. In the example below, we
 insert a row into the `customers` table. Note that the syntax is the same as for any standard `dbapi`
 connection.
 
+:::info
+* This method will work for all SQL destinations supported by `dlt`, but not for the filesystem destination.
+* Read the [SQL client docs](../../ general-usage/dataset-access/dataset) for more information on how to access data with the SQL client.
+* If you are simply trying to read data, you should use the powerful [dataset interface](../../general-usage/dataset-access/dataset) instead.
+:::
+
+
+Typically you will use this type of transformation if you can create or update tables directly from existing tables
+without any need to insert data from your Python environment. 
+
+The example below creates a new table `aggregated_sales` that contains the total and average sales for each category and region
+
+
 ```py
-pipeline = dlt.pipeline(destination="bigquery", dataset_name="crm")
-try:
-    with pipeline.sql_client() as client:
-        client.execute_sql(
-            "INSERT INTO customers VALUES (%s, %s, %s)",
-            10,
-            "Fred",
-            "fred@fred.com"
-        )
-except Exception:
-    ...
+pipeline = dlt.pipeline(destination="duckdb", dataset_name="crm")
+
+# NOTE: this is the duckdb sql dialect, other destinations may use different expressions
+with pipeline.sql_client() as client:
+    client.execute_sql(
+        """ CREATE OR REPLACE TABLE aggregated_sales AS
+            SELECT 
+                category,
+                region,
+                SUM(amount) AS total_sales,
+                AVG(amount) AS average_sales
+            FROM 
+                sales
+            GROUP BY 
+                category, 
+                region;
+    """)
 ```
 
-In the case of SELECT queries, the data is returned as a list of rows, with the elements of a row
-corresponding to selected columns.
+You can also use the `execute_sql` method to run select queries. The data is returned as a list of rows, with the elements of a row
+corresponding to selected columns. A more convenient way to extract data is to use dlt datasets. 
 
 ```py
 try:
@@ -44,9 +63,9 @@ except Exception:
 
 ## Other transforming tools
 
-If you want to transform the data before loading, you can use Python. If you want to transform the
+If you want to transform your data before loading, you can use Python. If you want to transform the
 data after loading, you can use SQL or one of the following:
 
 1. [dbt](dbt/dbt.md) (recommended).
-2. [Pandas](pandas.md).
+2. [Python with DataFrames or Arrow tables](python.md).
 
diff --git a/docs/website/docs/dlt-ecosystem/verified-sources/rest_api/basic.md b/docs/website/docs/dlt-ecosystem/verified-sources/rest_api/basic.md
index 14d9ecb04b..ea3c9c768b 100644
--- a/docs/website/docs/dlt-ecosystem/verified-sources/rest_api/basic.md
+++ b/docs/website/docs/dlt-ecosystem/verified-sources/rest_api/basic.md
@@ -306,7 +306,7 @@ A resource configuration is used to define a [dlt resource](../../../general-usa
 - `write_disposition`: The write disposition for the resource.
 - `primary_key`: The primary key for the resource.
 - `include_from_parent`: A list of fields from the parent resource to be included in the resource output. See the [resource relationships](#include-fields-from-the-parent-resource) section for more details.
-- `processing_steps`: A list of [processing steps](#processing-steps-filter-and-transform-data) to filter and transform the data.
+- `processing_steps`: A list of [processing steps](#processing-steps-filter-and-transform-data) to filter and transform your data.
 - `selected`: A flag to indicate if the resource is selected for loading. This could be useful when you want to load data only from child resources and not from the parent resource.
 - `auth`: An optional `AuthConfig` instance. If passed, is used over the one defined in the [client](#client) definition. Example:
 ```py
diff --git a/docs/website/docs/general-usage/dataset-access/dataset.md b/docs/website/docs/general-usage/dataset-access/dataset.md
index b2e3f03d4d..f9c01603f6 100644
--- a/docs/website/docs/general-usage/dataset-access/dataset.md
+++ b/docs/website/docs/general-usage/dataset-access/dataset.md
@@ -19,7 +19,7 @@ Here's a full example of how to retrieve data from a pipeline and load it into a
 # and you have loaded data to a table named 'items' in the destination
 
 # Step 1: Get the readable dataset from the pipeline
-dataset = pipeline._dataset()
+dataset = pipeline.dataset()
 
 # Step 2: Access a table as a ReadableRelation
 items_relation = dataset.items  # Or dataset["items"]
@@ -39,7 +39,10 @@ Assuming you have a `Pipeline` object (let's call it `pipeline`), you can obtain
 
 ```py
 # Get the readable dataset from the pipeline
-dataset = pipeline._dataset()
+dataset = pipeline.dataset()
+
+# print the row counts of all tables in the destination as dataframe
+print(dataset.row_counts().df())
 ```
 
 ### Access tables as `ReadableRelation`
@@ -116,6 +119,18 @@ for items_chunk in items_relation.iter_fetch(chunk_size=500):
 
 The methods available on the ReadableRelation correspond to the methods available on the cursor returned by the SQL client. Please refer to the [SQL client](./sql-client.md#supported-methods-on-the-cursor) guide for more information.
 
+## Special queries
+
+You can use the `row_counts` method to get the row counts of all tables in the destination as a DataFrame.
+
+```py
+# print the row counts of all tables in the destination as dataframe
+print(dataset.row_counts().df())
+
+# or as tuples
+print(dataset.row_counts().fetchall())
+```
+
 ## Modifying queries
 
 You can refine your data retrieval by limiting the number of records, selecting specific columns, or chaining these operations.
@@ -168,7 +183,7 @@ dlt will then wrap an `ibis.UnboundTable` with a `ReadableIbisRelation` object u
 
 ```py
 # now that ibis is installed, we can get a dataset with ibis relations
-dataset = pipeline._dataset()
+dataset = pipeline.dataset()
 
 # get two relations
 items_relation = dataset["items"]
@@ -284,7 +299,9 @@ other_pipeline = dlt.pipeline(pipeline_name="other_pipeline", destination="duckd
 other_pipeline.run(limited_items_relation.iter_arrow(chunk_size=10_000), table_name="limited_items")
 ```
 
-### Using `ibis` to query the data
+Learn more about [transforming data in Python with Arrow tables or DataFrames](../../dlt-ecosystem/transformations/python).
+
+### Using `ibis` to query data
 
 Visit the [Native Ibis integration](./ibis-backend.md) guide to learn more.
 
diff --git a/docs/website/docs/general-usage/dataset-access/ibis-backend.md b/docs/website/docs/general-usage/dataset-access/ibis-backend.md
index 9f9b65e9c0..bc8487940e 100644
--- a/docs/website/docs/general-usage/dataset-access/ibis-backend.md
+++ b/docs/website/docs/general-usage/dataset-access/ibis-backend.md
@@ -28,7 +28,7 @@ pip install ibis-framework[duckdb]
 
 ```py
 # get the dataset from the pipeline
-dataset = pipeline._dataset()
+dataset = pipeline.dataset()
 dataset_name = pipeline.dataset_name
 
 # get the native ibis connection from the dataset
diff --git a/docs/website/docs/general-usage/destination.md b/docs/website/docs/general-usage/destination.md
index fa133b6257..ba42869957 100644
--- a/docs/website/docs/general-usage/destination.md
+++ b/docs/website/docs/general-usage/destination.md
@@ -128,7 +128,7 @@ When loading data, `dlt` will access the destination in two cases:
 1. At the beginning of the `run` method to sync the pipeline state with the destination (or if you call `pipeline.sync_destination` explicitly).
 2. In the `pipeline.load` method - to migrate the schema and load the load package.
 
-Obviously, `dlt` will access the destination when you instantiate [sql_client](../dlt-ecosystem/transformations/sql.md).
+`dlt` will also access the destination when you instantiate [sql_client](../dlt-ecosystem/transformations/sql.md).
 
 :::note
 `dlt` will not import the destination dependencies or access destination configuration if access is not needed. You can build multi-stage pipelines where steps are executed in separate processes or containers - the `extract` and `normalize` step do not need destination dependencies, configuration, and actual connection.
diff --git a/docs/website/docs/general-usage/state.md b/docs/website/docs/general-usage/state.md
index 46aa1d63ce..d1fb426452 100644
--- a/docs/website/docs/general-usage/state.md
+++ b/docs/website/docs/general-usage/state.md
@@ -123,14 +123,13 @@ def comments(user_id: str):
     # on the first pipeline run, the user_comments table does not yet exist so do not check at all
     # alternatively, catch DatabaseUndefinedRelation which is raised when an unknown table is selected
     if not current_pipeline.first_run:
-        with current_pipeline.sql_client() as client:
-            # we may get the last user comment or None which we replace with 0
-            max_id = (
-                client.execute_sql(
-                    "SELECT MAX(_id) FROM user_comments WHERE user_id=?", user_id
-                )[0][0]
-                or 0
-            )
+        # get user comments table from pipeline dataset
+        user_comments = current_pipeline.dataset().user_comments
+        # get last user comment id with ibis expression, ibis-extras need to be installed
+        max_id_df = user_comments.filter(user_comments.user_id == user_id).select(user_comments["_id"].max()).df()
+        # if there are no comments for the user, max_id will be None, so we replace it with 0
+        max_id = max_id_df[0][0] if len(max_id_df.index) else 0
+
     # use max_id to filter our results (we simulate an API query)
     yield from [
         {"_id": i, "value": letter, "user_id": user_id}
diff --git a/docs/website/docs/intro.md b/docs/website/docs/intro.md
index b20d41c494..bc227b85ad 100644
--- a/docs/website/docs/intro.md
+++ b/docs/website/docs/intro.md
@@ -70,6 +70,10 @@ pipeline = dlt.pipeline(
 )
 
 load_info = pipeline.run(source)
+
+# print load info and posts table as dataframe
+print(load_info)
+print(pipeline.dataset().posts.df())
 ```
 
 Follow the [REST API source tutorial](./tutorial/rest-api) to learn more about the source configuration and pagination methods.
@@ -92,6 +96,10 @@ pipeline = dlt.pipeline(
 )
 
 load_info = pipeline.run(source)
+
+# print load info and the "family" table as dataframe
+print(load_info)
+print(pipeline.dataset().family.df())
 ```
 
 Follow the [SQL source tutorial](./tutorial/sql-database) to learn more about the source configuration and supported databases.
@@ -116,6 +124,10 @@ pipeline = dlt.pipeline(
 )
 
 load_info = pipeline.run(resource)
+
+# print load info and the "example" table as dataframe
+print(load_info)
+print(pipeline.dataset().example.df())
 ```
 
 Follow the [filesystem source tutorial](./tutorial/filesystem) to learn more about the source configuration and supported storage services.
@@ -128,7 +140,7 @@ dlt is able to load data from Python generators or directly from Python data str
 ```py
 import dlt
 
-@dlt.resource
+@dlt.resource(table_name="foo_data")
 def foo():
     for i in range(10):
         yield {"id": i, "name": f"This is item {i}"}
@@ -139,6 +151,10 @@ pipeline = dlt.pipeline(
 )
 
 load_info = pipeline.run(foo)
+
+# print load info and the "foo_data" table as dataframe
+print(load_info)
+print(pipeline.dataset().foo_data.df())
 ```
 
 Check out the [Python data structures tutorial](./tutorial/load-data-from-an-api) to learn about dlt fundamentals and advanced usage scenarios.
diff --git a/docs/website/docs/tutorial/load-data-from-an-api.md b/docs/website/docs/tutorial/load-data-from-an-api.md
index ddfef2cbe8..73f780ba7a 100644
--- a/docs/website/docs/tutorial/load-data-from-an-api.md
+++ b/docs/website/docs/tutorial/load-data-from-an-api.md
@@ -72,7 +72,25 @@ Load package 1692364844.460054 is LOADED and contains no failed jobs
 
 `dlt` just created a database schema called **mydata** (the `dataset_name`) with a table **users** in it.
 
-### Explore the data
+### Explore data in Python
+
+You can use dlt [datasets](../general-usage/dataset-access/dataset) to easily query the data in pure Python.
+
+```py
+# get the dataset
+dataset = pipeline.dataset("mydata")
+
+# get the user relation
+table = dataset.users
+
+# query the full table as dataframe
+print(table.df())
+
+# query the first 10 rows as arrow table
+print(table.limit(10).arrow())
+```
+
+### Explore data in Streamlit
 
 To allow a sneak peek and basic discovery, you can take advantage of [built-in integration with Streamlit](../reference/command-line-interface#show-tables-and-data-in-the-destination):
 
diff --git a/docs/website/sidebars.js b/docs/website/sidebars.js
index 8e8c11fc09..ca75c29392 100644
--- a/docs/website/sidebars.js
+++ b/docs/website/sidebars.js
@@ -211,13 +211,10 @@ const sidebars = {
         },
         {
           type: 'category',
-          label: 'Transform the data',
+          label: 'Transforming data',
           link: {
-            type: 'generated-index',
-            title: 'Transform the data',
-            description: 'If you want to transform the data after loading, you can use one of the following methods: dbt, SQL, Pandas.',
-            slug: 'dlt-ecosystem/transformations',
-            keywords: ['transformations'],
+            type: 'doc',
+            id: 'dlt-ecosystem/transformations/index',
           },
           items: [
             {
@@ -228,8 +225,8 @@ const sidebars = {
                 'dlt-ecosystem/transformations/dbt/dbt_cloud',
               ]
             },
+            'dlt-ecosystem/transformations/python',
             'dlt-ecosystem/transformations/sql',
-            'dlt-ecosystem/transformations/pandas',
             'general-usage/customising-pipelines/renaming_columns',
             'general-usage/customising-pipelines/pseudonymizing_columns',
             'general-usage/customising-pipelines/removing_columns'
diff --git a/tests/destinations/test_readable_dbapi_dataset.py b/tests/destinations/test_readable_dbapi_dataset.py
index bc58a18fa0..e3b318e8d4 100644
--- a/tests/destinations/test_readable_dbapi_dataset.py
+++ b/tests/destinations/test_readable_dbapi_dataset.py
@@ -9,7 +9,7 @@
 
 
 def test_query_builder() -> None:
-    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline")._dataset()
+    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline").dataset()
 
     # default query for a table
     assert dataset.my_table.query.strip() == 'SELECT  * FROM "pipeline_dataset"."my_table"'
@@ -55,7 +55,7 @@ def test_query_builder() -> None:
 
 
 def test_copy_and_chaining() -> None:
-    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline")._dataset()
+    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline").dataset()
 
     # create releation and set some stuff on it
     relation = dataset.items
@@ -80,7 +80,7 @@ def test_copy_and_chaining() -> None:
 
 
 def test_computed_schema_columns() -> None:
-    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline")._dataset()
+    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline").dataset()
     relation = dataset.items
 
     # no schema present
@@ -107,7 +107,7 @@ def test_computed_schema_columns() -> None:
 
 
 def test_prevent_changing_relation_with_query() -> None:
-    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline")._dataset()
+    dataset = dlt.pipeline(destination="duckdb", pipeline_name="pipeline").dataset()
     relation = dataset("SELECT * FROM something")
 
     with pytest.raises(ReadableRelationHasQueryException):
diff --git a/tests/extract/test_incremental.py b/tests/extract/test_incremental.py
index 3ebc9d1201..d63dac93f2 100644
--- a/tests/extract/test_incremental.py
+++ b/tests/extract/test_incremental.py
@@ -228,7 +228,7 @@ def test_pandas_index_as_dedup_key() -> None:
     no_index_r = some_data.with_name(new_name="no_index")
     p.run(no_index_r)
     p.run(no_index_r)
-    data_ = p._dataset().no_index.arrow()
+    data_ = p.dataset().no_index.arrow()
     assert data_.schema.names == ["created_at", "id"]
     assert data_["id"].to_pylist() == ["a", "b", "c", "d", "e", "f", "g"]
 
@@ -240,7 +240,7 @@ def test_pandas_index_as_dedup_key() -> None:
     unnamed_index_r.incremental.primary_key = "__index_level_0__"
     p.run(unnamed_index_r)
     p.run(unnamed_index_r)
-    data_ = p._dataset().unnamed_index.arrow()
+    data_ = p.dataset().unnamed_index.arrow()
     assert data_.schema.names == ["created_at", "id", "index_level_0"]
     # indexes 2 and 3 are removed from second batch because they were in the previous batch
     # and the created_at overlapped so they got deduplicated
@@ -258,7 +258,7 @@ def _make_named_index(df_: pd.DataFrame) -> pd.DataFrame:
     named_index_r.incremental.primary_key = "order_id"
     p.run(named_index_r)
     p.run(named_index_r)
-    data_ = p._dataset().named_index.arrow()
+    data_ = p.dataset().named_index.arrow()
     assert data_.schema.names == ["created_at", "id", "order_id"]
     assert data_["order_id"].to_pylist() == [0, 1, 2, 3, 4, 0, 1, 4]
 
@@ -268,7 +268,7 @@ def _make_named_index(df_: pd.DataFrame) -> pd.DataFrame:
     )
     p.run(named_index_impl_r)
     p.run(named_index_impl_r)
-    data_ = p._dataset().named_index_impl.arrow()
+    data_ = p.dataset().named_index_impl.arrow()
     assert data_.schema.names == ["created_at", "id"]
     assert data_["id"].to_pylist() == ["a", "b", "c", "d", "e", "f", "g"]
 
diff --git a/tests/load/duckdb/test_duckdb_client.py b/tests/load/duckdb/test_duckdb_client.py
index 49475ce43f..652f75772a 100644
--- a/tests/load/duckdb/test_duckdb_client.py
+++ b/tests/load/duckdb/test_duckdb_client.py
@@ -282,14 +282,14 @@ def test_drops_pipeline_changes_bound() -> None:
     p = dlt.pipeline(pipeline_name="quack_pipeline", destination="duckdb")
     p.run([1, 2, 3], table_name="p_table")
     p = p.drop()
-    assert len(p._dataset().p_table.fetchall()) == 3
+    assert len(p.dataset().p_table.fetchall()) == 3
 
     # drops internal duckdb
     p = dlt.pipeline(pipeline_name="quack_pipeline", destination=duckdb(":pipeline:"))
     p.run([1, 2, 3], table_name="p_table")
     p = p.drop()
     with pytest.raises(DatabaseUndefinedRelation):
-        p._dataset().p_table.fetchall()
+        p.dataset().p_table.fetchall()
 
 
 def test_duckdb_database_delete() -> None:
diff --git a/tests/load/filesystem/test_sql_client.py b/tests/load/filesystem/test_sql_client.py
index 4f537d129c..cf4bbfb286 100644
--- a/tests/load/filesystem/test_sql_client.py
+++ b/tests/load/filesystem/test_sql_client.py
@@ -377,7 +377,7 @@ def items():
 
     pipeline.run([items()], loader_file_format=destination_config.file_format)
 
-    df = pipeline._dataset().items.df()
+    df = pipeline.dataset().items.df()
     assert len(df.index) == 20
 
     @dlt.resource(table_name="items")
@@ -387,5 +387,5 @@ def items2():
     pipeline.run([items2()], loader_file_format=destination_config.file_format)
 
     # check df and arrow access
-    assert len(pipeline._dataset().items.df().index) == 50
-    assert pipeline._dataset().items.arrow().num_rows == 50
+    assert len(pipeline.dataset().items.df().index) == 50
+    assert pipeline.dataset().items.arrow().num_rows == 50
diff --git a/tests/load/pipeline/test_bigquery.py b/tests/load/pipeline/test_bigquery.py
index cb65c6bcf1..83982bb998 100644
--- a/tests/load/pipeline/test_bigquery.py
+++ b/tests/load/pipeline/test_bigquery.py
@@ -384,8 +384,8 @@ def resource():
     bigquery_adapter(resource, autodetect_schema=True)
     pipeline.run(resource)
 
-    assert len(pipeline._dataset().items.df()) == 5
-    assert len(pipeline._dataset().items__nested.df()) == 5
+    assert len(pipeline.dataset().items.df()) == 5
+    assert len(pipeline.dataset().items__nested.df()) == 5
 
     @dlt.resource(primary_key="id", table_name="items", write_disposition="merge")
     def resource2():
@@ -395,5 +395,5 @@ def resource2():
     bigquery_adapter(resource2, autodetect_schema=True)
     pipeline.run(resource2)
 
-    assert len(pipeline._dataset().items.df()) == 7
-    assert len(pipeline._dataset().items__nested.df()) == 7
+    assert len(pipeline.dataset().items.df()) == 7
+    assert len(pipeline.dataset().items__nested.df()) == 7
diff --git a/tests/load/pipeline/test_duckdb.py b/tests/load/pipeline/test_duckdb.py
index a7aa4d36e4..2d1138a51d 100644
--- a/tests/load/pipeline/test_duckdb.py
+++ b/tests/load/pipeline/test_duckdb.py
@@ -273,10 +273,10 @@ def test_duckdb_credentials_separation(
     p2 = dlt.pipeline("p2", destination=duckdb(credentials=":pipeline:"))
 
     p1.run([1, 2, 3], table_name="p1_data")
-    p1_dataset = p1._dataset()
+    p1_dataset = p1.dataset()
 
     p2.run([1, 2, 3], table_name="p2_data")
-    p2_dataset = p2._dataset()
+    p2_dataset = p2.dataset()
 
     # both dataset should have independent duckdb databases
     # destinations should be bounded to pipelines still
diff --git a/tests/load/test_read_interfaces.py b/tests/load/test_read_interfaces.py
index d2f5f7951e..bca844d2c8 100644
--- a/tests/load/test_read_interfaces.py
+++ b/tests/load/test_read_interfaces.py
@@ -25,6 +25,7 @@
     ReadableRelationUnknownColumnException,
 )
 from tests.load.utils import drop_pipeline_data
+from dlt.destinations.dataset import dataset as _dataset
 
 EXPECTED_COLUMNS = ["id", "decimal", "other_decimal", "_dlt_load_id", "_dlt_id"]
 
@@ -169,9 +170,9 @@ def test_explicit_dataset_type_selection(populated_pipeline: Pipeline):
     from dlt.destinations.dataset.ibis_relation import ReadableIbisRelation
 
     assert isinstance(
-        populated_pipeline._dataset(dataset_type="default").items, ReadableDBAPIRelation
+        populated_pipeline.dataset(dataset_type="default").items, ReadableDBAPIRelation
     )
-    assert isinstance(populated_pipeline._dataset(dataset_type="ibis").items, ReadableIbisRelation)
+    assert isinstance(populated_pipeline.dataset(dataset_type="ibis").items, ReadableIbisRelation)
 
 
 @pytest.mark.no_load
@@ -183,7 +184,7 @@ def test_explicit_dataset_type_selection(populated_pipeline: Pipeline):
     ids=lambda x: x.name,
 )
 def test_arrow_access(populated_pipeline: Pipeline) -> None:
-    table_relationship = populated_pipeline._dataset().items
+    table_relationship = populated_pipeline.dataset().items
     total_records = _total_records(populated_pipeline)
     chunk_size = _chunk_size(populated_pipeline)
     expected_chunk_counts = _expected_chunk_count(populated_pipeline)
@@ -216,7 +217,7 @@ def test_arrow_access(populated_pipeline: Pipeline) -> None:
 )
 def test_dataframe_access(populated_pipeline: Pipeline) -> None:
     # access via key
-    table_relationship = populated_pipeline._dataset()["items"]
+    table_relationship = populated_pipeline.dataset()["items"]
     total_records = _total_records(populated_pipeline)
     chunk_size = _chunk_size(populated_pipeline)
     expected_chunk_counts = _expected_chunk_count(populated_pipeline)
@@ -233,7 +234,6 @@ def test_dataframe_access(populated_pipeline: Pipeline) -> None:
     if not skip_df_chunk_size_check:
         assert len(df.index) == chunk_size
 
-    # lowercase results for the snowflake case
     assert set(df.columns.values) == set(EXPECTED_COLUMNS)
 
     # iterate all dataframes
@@ -256,7 +256,7 @@ def test_dataframe_access(populated_pipeline: Pipeline) -> None:
 )
 def test_db_cursor_access(populated_pipeline: Pipeline) -> None:
     # check fetch accessors
-    table_relationship = populated_pipeline._dataset().items
+    table_relationship = populated_pipeline.dataset().items
     total_records = _total_records(populated_pipeline)
     chunk_size = _chunk_size(populated_pipeline)
     expected_chunk_counts = _expected_chunk_count(populated_pipeline)
@@ -290,8 +290,7 @@ def test_db_cursor_access(populated_pipeline: Pipeline) -> None:
     ids=lambda x: x.name,
 )
 def test_hint_preservation(populated_pipeline: Pipeline) -> None:
-    # NOTE: for now hints are only preserved for the default dataset
-    table_relationship = populated_pipeline._dataset(dataset_type="default").items
+    table_relationship = populated_pipeline.dataset(dataset_type="default").items
     # check that hints are carried over to arrow table
     expected_decimal_precision = 10
     expected_decimal_precision_2 = 12
@@ -319,10 +318,94 @@ def test_hint_preservation(populated_pipeline: Pipeline) -> None:
 )
 def test_loads_table_access(populated_pipeline: Pipeline) -> None:
     # check loads table access, we should have one entry
-    loads_table = populated_pipeline._dataset()[populated_pipeline.default_schema.loads_table_name]
+    loads_table = populated_pipeline.dataset()[populated_pipeline.default_schema.loads_table_name]
     assert len(loads_table.fetchall()) == 1
 
 
+@pytest.mark.no_load
+@pytest.mark.essential
+@pytest.mark.parametrize(
+    "populated_pipeline",
+    configs,
+    indirect=True,
+    ids=lambda x: x.name,
+)
+def test_row_counts(populated_pipeline: Pipeline) -> None:
+    total_records = _total_records(populated_pipeline)
+
+    dataset = populated_pipeline.dataset()
+    # default is all data tables
+    assert set(dataset.row_counts().df().itertuples(index=False, name=None)) == {
+        (
+            "items",
+            total_records,
+        ),
+        (
+            "double_items",
+            total_records,
+        ),
+        (
+            "items__children",
+            total_records * 2,
+        ),
+    }
+    # get only one data table
+    assert set(
+        dataset.row_counts(table_names=["items"]).df().itertuples(index=False, name=None)
+    ) == {
+        (
+            "items",
+            total_records,
+        ),
+    }
+    # get all dlt tables
+    assert set(
+        dataset.row_counts(dlt_tables=True, data_tables=False)
+        .df()
+        .itertuples(index=False, name=None)
+    ) == {
+        (
+            "_dlt_version",
+            1,
+        ),
+        (
+            "_dlt_loads",
+            1,
+        ),
+        (
+            "_dlt_pipeline_state",
+            1,
+        ),
+    }
+    # get them all
+    assert set(dataset.row_counts(dlt_tables=True).df().itertuples(index=False, name=None)) == {
+        (
+            "_dlt_version",
+            1,
+        ),
+        (
+            "_dlt_loads",
+            1,
+        ),
+        (
+            "_dlt_pipeline_state",
+            1,
+        ),
+        (
+            "items",
+            total_records,
+        ),
+        (
+            "double_items",
+            total_records,
+        ),
+        (
+            "items__children",
+            total_records * 2,
+        ),
+    }
+
+
 @pytest.mark.no_load
 @pytest.mark.essential
 @pytest.mark.parametrize(
@@ -334,7 +417,7 @@ def test_loads_table_access(populated_pipeline: Pipeline) -> None:
 def test_sql_queries(populated_pipeline: Pipeline) -> None:
     # simple check that query also works
     tname = populated_pipeline.sql_client().make_qualified_table_name("items")
-    query_relationship = populated_pipeline._dataset()(f"select * from {tname} where id < 20")
+    query_relationship = populated_pipeline.dataset()(f"select * from {tname} where id < 20")
 
     # we selected the first 20
     table = query_relationship.arrow()
@@ -346,7 +429,7 @@ def test_sql_queries(populated_pipeline: Pipeline) -> None:
         f"SELECT i.id, di.double_id FROM {tname} as i JOIN {tdname} as di ON (i.id = di.id) WHERE"
         " i.id < 20 ORDER BY i.id ASC"
     )
-    join_relationship = populated_pipeline._dataset()(query)
+    join_relationship = populated_pipeline.dataset()(query)
     table = join_relationship.fetchall()
     assert len(table) == 20
     assert list(table[0]) == [0, 0]
@@ -363,7 +446,7 @@ def test_sql_queries(populated_pipeline: Pipeline) -> None:
     ids=lambda x: x.name,
 )
 def test_limit_and_head(populated_pipeline: Pipeline) -> None:
-    table_relationship = populated_pipeline._dataset().items
+    table_relationship = populated_pipeline.dataset().items
 
     assert len(table_relationship.head().fetchall()) == 5
     assert len(table_relationship.limit(24).fetchall()) == 24
@@ -384,7 +467,7 @@ def test_limit_and_head(populated_pipeline: Pipeline) -> None:
     ids=lambda x: x.name,
 )
 def test_column_selection(populated_pipeline: Pipeline) -> None:
-    table_relationship = populated_pipeline._dataset(dataset_type="default").items
+    table_relationship = populated_pipeline.dataset(dataset_type="default").items
     columns = ["_dlt_load_id", "other_decimal"]
     data_frame = table_relationship.select(*columns).head().df()
     assert [v.lower() for v in data_frame.columns.values] == columns
@@ -421,18 +504,18 @@ def test_schema_arg(populated_pipeline: Pipeline) -> None:
     """Simple test to ensure schemas may be selected via schema arg"""
 
     # if there is no arg, the defautl schema is used
-    dataset = populated_pipeline._dataset()
+    dataset = populated_pipeline.dataset()
     assert dataset.schema.name == populated_pipeline.default_schema_name
     assert "items" in dataset.schema.tables
 
     # setting a different schema name will try to load that schema,
     # not find one and create an empty schema with that name
-    dataset = populated_pipeline._dataset(schema="unknown_schema")
+    dataset = populated_pipeline.dataset(schema="unknown_schema")
     assert dataset.schema.name == "unknown_schema"
     assert "items" not in dataset.schema.tables
 
     # providing the schema name of the right schema will load it
-    dataset = populated_pipeline._dataset(schema=populated_pipeline.default_schema_name)
+    dataset = populated_pipeline.dataset(schema=populated_pipeline.default_schema_name)
     assert dataset.schema.name == populated_pipeline.default_schema_name
     assert "items" in dataset.schema.tables
 
@@ -450,7 +533,7 @@ def test_ibis_expression_relation(populated_pipeline: Pipeline) -> None:
     import ibis  # type: ignore
 
     # now we should get the more powerful ibis relation
-    dataset = populated_pipeline._dataset()
+    dataset = populated_pipeline.dataset()
     total_records = _total_records(populated_pipeline)
 
     items_table = dataset["items"]
@@ -653,11 +736,11 @@ def test_ibis_dataset_access(populated_pipeline: Pipeline) -> None:
     # check correct error if not supported
     if populated_pipeline.destination.destination_type not in SUPPORTED_DESTINATIONS:
         with pytest.raises(NotImplementedError):
-            populated_pipeline._dataset().ibis()
+            populated_pipeline.dataset().ibis()
         return
 
     total_records = _total_records(populated_pipeline)
-    ibis_connection = populated_pipeline._dataset().ibis()
+    ibis_connection = populated_pipeline.dataset().ibis()
 
     map_i = lambda x: x
     if populated_pipeline.destination.destination_type == "dlt.destinations.snowflake":
@@ -709,7 +792,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
     total_records = _total_records(populated_pipeline)
 
     # check dataset factory
-    dataset = dlt._dataset(
+    dataset = _dataset(
         destination=populated_pipeline.destination, dataset_name=populated_pipeline.dataset_name
     )
     # verfiy that sql client and schema are lazy loaded
@@ -722,7 +805,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
     # check that schema is loaded by name
     dataset = cast(
         ReadableDBAPIDataset,
-        dlt._dataset(
+        _dataset(
             destination=populated_pipeline.destination,
             dataset_name=populated_pipeline.dataset_name,
             schema=populated_pipeline.default_schema_name,
@@ -733,7 +816,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
     # check that schema is not loaded when wrong name given
     dataset = cast(
         ReadableDBAPIDataset,
-        dlt._dataset(
+        _dataset(
             destination=populated_pipeline.destination,
             dataset_name=populated_pipeline.dataset_name,
             schema="wrong_schema_name",
@@ -745,7 +828,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
     # check that schema is loaded if no schema name given
     dataset = cast(
         ReadableDBAPIDataset,
-        dlt._dataset(
+        _dataset(
             destination=populated_pipeline.destination,
             dataset_name=populated_pipeline.dataset_name,
         ),
@@ -756,7 +839,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
     # check that there is no error when creating dataset without schema table
     dataset = cast(
         ReadableDBAPIDataset,
-        dlt._dataset(
+        _dataset(
             destination=populated_pipeline.destination,
             dataset_name="unknown_dataset",
         ),
@@ -779,7 +862,7 @@ def test_standalone_dataset(populated_pipeline: Pipeline) -> None:
 
     dataset = cast(
         ReadableDBAPIDataset,
-        dlt._dataset(
+        _dataset(
             destination=populated_pipeline.destination,
             dataset_name=populated_pipeline.dataset_name,
         ),
diff --git a/tests/pipeline/test_dlt_versions.py b/tests/pipeline/test_dlt_versions.py
index fbd4d412b3..51de3e0f76 100644
--- a/tests/pipeline/test_dlt_versions.py
+++ b/tests/pipeline/test_dlt_versions.py
@@ -538,5 +538,5 @@ def test_normalize_path_separator_legacy_behavior(test_storage: FileStorage) ->
                     "_dlt_load_id",
                 }
                 # datasets must be the same
-                data_ = pipeline._dataset().issues_2.select("issue_id", "id").fetchall()
+                data_ = pipeline.dataset().issues_2.select("issue_id", "id").fetchall()
                 print(data_)
diff --git a/tests/pipeline/test_pipeline.py b/tests/pipeline/test_pipeline.py
index b32854b110..2d72e23462 100644
--- a/tests/pipeline/test_pipeline.py
+++ b/tests/pipeline/test_pipeline.py
@@ -1754,7 +1754,7 @@ def test_column_name_with_break_path() -> None:
     # get data
     assert_data_table_counts(pipeline, {"custom__path": 1})
     # get data via dataset with dbapi
-    data_ = pipeline._dataset().custom__path[["example_custom_field__c", "reg_c"]].fetchall()
+    data_ = pipeline.dataset().custom__path[["example_custom_field__c", "reg_c"]].fetchall()
     assert data_ == [("custom", "c")]
 
 
@@ -1778,7 +1778,7 @@ def test_column_name_with_break_path_legacy() -> None:
     # get data
     assert_data_table_counts(pipeline, {"custom_path": 1})
     # get data via dataset with dbapi
-    data_ = pipeline._dataset().custom_path[["example_custom_field_c", "reg_c"]].fetchall()
+    data_ = pipeline.dataset().custom_path[["example_custom_field_c", "reg_c"]].fetchall()
     assert data_ == [("custom", "c")]
 
 
@@ -1806,7 +1806,7 @@ def flattened_dict():
     assert table["columns"]["value__timestamp"]["data_type"] == "timestamp"
 
     # make sure data is there
-    data_ = pipeline._dataset().flattened__dict[["delta", "value__timestamp"]].limit(1).fetchall()
+    data_ = pipeline.dataset().flattened__dict[["delta", "value__timestamp"]].limit(1).fetchall()
     assert data_ == [(0, now)]
 
 
@@ -1836,7 +1836,7 @@ def flattened_dict():
     assert set(table["columns"]) == {"delta", "value__timestamp", "_dlt_id", "_dlt_load_id"}
     assert table["columns"]["value__timestamp"]["data_type"] == "timestamp"
     # make sure data is there
-    data_ = pipeline._dataset().flattened_dict[["delta", "value__timestamp"]].limit(1).fetchall()
+    data_ = pipeline.dataset().flattened_dict[["delta", "value__timestamp"]].limit(1).fetchall()
     assert data_ == [(0, now)]
 
 
diff --git a/tests/pipeline/test_pipeline_extra.py b/tests/pipeline/test_pipeline_extra.py
index a51052d247..32b16c234f 100644
--- a/tests/pipeline/test_pipeline_extra.py
+++ b/tests/pipeline/test_pipeline_extra.py
@@ -521,7 +521,7 @@ def test_parquet_with_flattened_columns() -> None:
     assert "issue__reactions__url" in pipeline.default_schema.tables["events"]["columns"]
     assert "issue_reactions_url" not in pipeline.default_schema.tables["events"]["columns"]
 
-    events_table = pipeline._dataset().events.arrow()
+    events_table = pipeline.dataset().events.arrow()
     assert "issue__reactions__url" in events_table.schema.names
     assert "issue_reactions_url" not in events_table.schema.names
 
@@ -536,7 +536,7 @@ def test_parquet_with_flattened_columns() -> None:
     info = pipeline.run(events_table, table_name="events", loader_file_format="parquet")
     assert_load_info(info)
 
-    events_table_new = pipeline._dataset().events.arrow()
+    events_table_new = pipeline.dataset().events.arrow()
     assert events_table.schema == events_table_new.schema
     # double row count
     assert events_table.num_rows * 2 == events_table_new.num_rows