Incremental loading is the act of loading only new or changed data and not old records that have already been loaded.
For example, a bank loads only the latest transactions, or a company updates its database with new or modified user
information. In this article, we’ll discuss a few incremental loading strategies.

Incremental loading strategies#

In this guide, we will discuss various incremental loading methods using dlt, specifically:

Code examples#

1. Full load (replace)#

A full load strategy completely overwrites the existing data with the new dataset. This is useful when you want to refresh the entire table with the latest data.

Here’s a walkthrough:

1. The initial table, named "contact," in the SQL source looks like this:

   id	name	created_at
   1	Alice	2024-07-01
   2	Bob	2024-07-02

2. The Python code illustrates the process of loading data from an SQL source into BigQuery using the dlt pipeline. Please note the write_disposition = "replace" used below.

   ⧉

   def load_full_table_resource() -> None:
       """Load a full table, replacing existing data."""
       pipeline = dlt.pipeline(
           pipeline_name="mysql_database",
           destination='bigquery',
           dataset_name="dlt_contacts"
       )
   
       # Load the full table "contact"
       source = sql_database().with_resources("contact")
   
       # Run the pipeline
       info = pipeline.run(source, write_disposition="replace")
   
       # Print the info
       print(info)
   
   load_full_table_resource()
   

3. After running the dlt pipeline, the data loaded into the BigQuery "contact" table looks like:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01	1721878309.021546	tgyMM73iMz0cQg
   2	2	Bob	2024-07-02	1721878309.021546	88P0bD796pXo/Q

4. Next, the "contact" table in the SQL source is updated—two new rows are added, and the row with id = 2 is removed. The updated data source ("contact" table) now presents itself as follows:

   id	name	created_at
   1	Alice	2024-07-01
   3	Charlie	2024-07-03
   4	Dave	2024-07-04

5. The "contact" table created in BigQuery after running the pipeline again:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01	1721878309.021546	S5ye6fMhYECZA
   2	3	Charlie	2024-07-03	1721878309.021546	eT0zheRx9ONWuQ
   3	4	Dave	2024-07-04	1721878309.021546	gtflF8BdL2NO/Q

What happened?

After running the pipeline, the original data in the "contact" table (Alice and Bob) is completely replaced with the new updated table with data “Charlie” and “Dave” added and “Bob” removed. This strategy is useful for scenarios where the entire dataset needs to be refreshed or replaced with the latest information.

2. Append new records based on incremental ID#

This strategy appends only new records to the table based on an incremental ID. It is useful for scenarios where each new record has a unique, incrementing identifier.

Here’s a walkthrough:

1. The initial table, named "contact," in the SQL source looks like this:

   id	name	created_at
   1	Alice	2024-07-01
   2	Bob	2024-07-02

2. The Python code demonstrates loading data from an SQL source into BigQuery using an incremental variable, id. This variable tracks new or updated records in the dlt pipeline. Please note the write_disposition = "append" used below.

   ⧉

   def load_incremental_id_table_resource() -> None:
       """Load a table incrementally based on an ID."""
       pipeline = dlt.pipeline(
           pipeline_name="mysql_database",
           destination='bigquery',
           dataset_name="dlt_contacts",
       )
   
       # Load table "contact" incrementally based on ID
       source = sql_database().with_resources("contact")
       source.contact.apply_hints(incremental=dlt.sources.incremental("id"))
   
       # Run the pipeline with append write disposition
       info = pipeline.run(source, write_disposition="append")
   
       # Print the info
       print(info)
   

3. After running the dlt pipeline, the data loaded into the BigQuery "contact" table looks like:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01	1721878309.021546	YQfmAu8xysqWmA
   2	2	Bob	2024-07-02	1721878309.021546	Vcb5KKah/RpmQw

4. Next, the "contact" table in the SQL source is updated—two new rows are added, and the row with id = 2 is removed. The updated data source now presents itself as follows:

   id	name	created_at
   1	Alice	2024-07-01
   3	Charlie	2024-07-03
   4	Dave	2024-07-04

5. The "contact" table created in BigQuery after running the pipeline again:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01	1721878309.021546	OW9ZyAzkXg4D4w
   2	2	Bob	2024-07-02	1721878309.021546	skVYZ/ppQuztUg
   3	3	Charlie	2024-07-03	1721878309.021546	y+T4Q2JDnR33jg
   4	4	Dave	2024-07-04	1721878309.021546	MAXrGhNNADXAiQ

What happened?

In this scenario, the pipeline appends new records (Charlie and Dave) to the existing data (Alice and Bob) without affecting the pre-existing entries. This strategy is ideal when only new data needs to be added, preserving the historical data.

Append new records based on timestamp ("created_at")#

This strategy appends only new records to the table based on a date/timestamp field. It is useful for scenarios where records are created with a timestamp, and you want to load only those records created after a certain date.

Here’s a walkthrough:

1. The initial dataset, named "contact," in the SQL source looks like this:

   id	name	created_at
   1	Alice	2024-07-01 00:00:00
   2	Bob	2024-07-02 00:00:00

2. The Python code illustrates the process of loading data from an SQL source into BigQuery using the dlt pipeline. Please note the write_disposition = "append", with created_at being used as the incremental parameter.

   ⧉

   def load_incremental_timestamp_table_resource() -> None:
       """Load a table incrementally based on created_at timestamp."""
       pipeline = dlt.pipeline(
           pipeline_name="mysql_databasecdc",
           destination='bigquery',
           dataset_name="dlt_contacts",
       )
   
       # Load table "contact", incrementally starting at a given timestamp
       source = sql_database().with_resources("contact")
       source.contact.apply_hints(incremental=dlt.sources.incremental(
           "created_at", initial_value=datetime.datetime(2024, 4, 1, 0, 0, 0)))
   
       # Run the pipeline
       info = pipeline.run(source, write_disposition="append")
   
       # Print the info
       print(info)
   
   load_incremental_timestamp_table_resource()
   

3. After running the dlt pipeline, the data loaded into the BigQuery "contact" table looks like:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01 00:00:00 UTC	1721878309.021546	5H8ca6C89umxHA
   2	2	Bob	2024-07-02 00:00:00 UTC	1721878309.021546	M61j4aOSqs4k2w

4. Next, the "contact" table in the SQL source is updated—two new rows are added, and the row with id = 2 is removed. The updated data source now presents itself as follows:

   id	name	created_at
   1	Alice	2024-07-01 00:00:00
   3	Charlie	2024-07-03 00:00:00
   4	Dave	2024-07-04 00:00:00

5. The "contact" table created in BigQuery after running the pipeline again:

   Row	id	name	created_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01 00:00:00 UTC	1721878309.021546	Petj6R+B/63sWA
   2	2	Bob	2024-07-02 00:00:00 UTC	1721878309.021546	3Rr3VmY+av+Amw
   3	3	Charlie	2024-07-03 00:00:00 UTC	1721878309.021546	L/MnhG19xeMrvQ
   4	4	Dave	2024-07-04 00:00:00 UTC	1721878309.021546	W6ZdfvTzfRXlsA

What happened?

The pipeline adds new records (Charlie and Dave) that have a created_at timestamp after the specified initial value while retaining the existing data (Alice and Bob). This approach is useful for loading data incrementally based on when it was created.

4. Merge (update/insert) records based on timestamp ("last_modified_at") and ID#

This strategy merges records based on a composite key of ID and a timestamp field. It updates existing records and inserts new ones as necessary.

Here’s a walkthrough:

1. The initial dataset, named ‘contact’, in the SQL source looks like this:

   id	name	last_modified_at
   1	Alice	2024-07-01 00:00:00
   2	Bob	2024-07-02 00:00:00

2. The Python code illustrates the process of loading data from an SQL source into BigQuery using the dlt pipeline. Please note the write_disposition = "merge", with last_modified_at being used as the incremental parameter.

   ⧉

   def load_merge_table_resource() -> None:
       """Merge (update/insert) records based on last_modified_at timestamp and ID."""
       pipeline = dlt.pipeline(
           pipeline_name="mysql_database",
           destination='bigquery',
           dataset_name="dlt_contacts",
       )
   
       # Merge records, 'contact' table, based on ID and last_modified_at timestamp
       source = sql_database().with_resources("contact")
       source.contact.apply_hints(incremental=dlt.sources.incremental(
           "last_modified_at", initial_value=datetime.datetime(2024, 4, 1, 0, 0, 0)),
           primary_key="id")
   
       # Run the pipeline
       info = pipeline.run(source, write_disposition="merge")
   
       # Print the info
       print(info)
   
   load_merge_table_resource()
   

3. After running the dlt pipeline, the data loaded into BigQuery ‘contact’ table looks like:

   Row	id	name	last_modified_at	_dlt_load_id	_dlt_id
   1	1	Alice	2024-07-01 00:00:00 UTC	1721878309.021546	ObbVlxcly3VknQ
   2	2	Bob	2024-07-02 00:00:00 UTC	1721878309.021546	Vrlkus/haaKlEg

4. Next, the "contact" table in the SQL source is updated— “Alice” is updated to “Alice Updated”, and a new row “Hank” is added:

   id	name	last_modified_at
   1	Alice Updated	2024-07-08 00:00:00
   3	Hank	2024-07-08 00:00:00

5. The "contact" table created in BigQuery after running the pipeline again:

   Row	id	name	last_modified_at	_dlt_load_id	_dlt_id
   1	2	Bob	2024-07-02 00:00:00 UTC	1721878309.021546	Cm+AcDZLqXSDHQ
   2	1	Alice Updated	2024-07-08 00:00:00 UTC	1721878309.021546	OeMLIPw7rwFG7g
   3	3	Hank	2024-07-08 00:00:00 UTC	1721878309.021546	Ttp6AI2JxqffpA

What happened?

The pipeline updates the record for Alice with the new data, including the updated last_modified_at timestamp, and adds a new record for Hank. This method is beneficial when you need to ensure that records are both updated and inserted based on a specific timestamp and ID.

The examples provided explain how to use dlt to achieve different incremental loading scenarios, highlighting the changes before and after running each pipeline.