Schema#

The schema describes the structure of normalized data (e.g., tables, columns, data types, etc.) and
provides instructions on how the data should be processed and loaded. dlt generates schemas from
the data during the normalization process. Users can affect this standard behavior by providing
hints that change how tables, columns, and other metadata are generated and how the data is
loaded. Such hints can be passed in the code, i.e., to the dlt.resource decorator or pipeline.run
method. Schemas can also be exported and imported as files, which can be directly modified.

💡 dlt associates a schema with a source and a table schema with a
resource.

Schema content hash and version#

Each schema file contains a content-based hash version_hash that is used to:

1. Detect manual changes to the schema (i.e., user edits content).
2. Detect if the destination database schema is synchronized with the file schema.

Each time the schema is saved, the version hash is updated.

Each schema contains a numeric version which increases automatically whenever the schema is updated and
saved. The numeric version is meant to be human-readable. There are cases (parallel processing) where
the order is lost.

💡 The schema in the destination is migrated if its hash is not stored in the _dlt_versions table. In
principle, many pipelines may send data to a single dataset. If table names clash, then a single
table with the union of the columns will be created. If columns clash, and they have different
types, etc., then the load may fail if the data cannot be coerced.

Naming convention#

dlt creates tables, nested tables, and column schemas from the data. The data being loaded,
typically JSON documents, contains identifiers (i.e., key names in a dictionary) with any Unicode
characters, any lengths, and naming styles. On the other hand, the destinations accept very strict
namespaces for their identifiers. Like Redshift, that accepts case-insensitive alphanumeric
identifiers with a maximum of 127 characters.

Each schema contains a naming convention that tells dlt how to translate identifiers to the
namespace that the destination understands. This convention can be configured, changed in code, or enforced via
destination.

The default naming convention:

1. Converts identifiers to snake_case, small caps. Removes all ASCII characters except ASCII
   alphanumerics and underscores.
2. Adds _ if the name starts with a number.
3. Multiples of _ are converted into a single _.
4. Nesting is expressed as double _ in names.
5. It shortens the identifier if it exceeds the length at the destination.

💡 The standard behavior of dlt is to use the same naming convention for all destinations so
users always see the same tables and columns in their databases.

💡 If you provide any schema elements that contain identifiers via decorators or arguments (i.e.,
table_name or columns), all the names used will be converted via the naming convention when
adding to the schema. For example, if you execute dlt.run(... table_name="CamelCase") the data
will be loaded into camel_case.

💡 Use simple, short, small caps identifiers for everything!

To retain the original naming convention (like keeping "createdAt" as it is instead of converting it to "created_at"), you can use the direct naming convention in "config.toml" as follows:

[schema]
naming="direct"

The naming convention is configurable, and users can easily create their own
conventions that, i.e., pass all the identifiers unchanged if the destination accepts that (i.e.,
DuckDB).

Data normalizer#

The data normalizer changes the structure of the input data so it can be loaded into the destination. The standard dlt normalizer creates a relational structure from Python dictionaries and lists. Elements of that structure, such as table and column definitions, are added to the schema.

The data normalizer is configurable, and users can plug in their own normalizers, for example, to handle nested table linking differently or generate parquet-like data structures instead of nested tables.

dlt ships with two JSON normalizer variants:

To switch the normalizer, set it in config.toml using the shorthand name:

[schema]
json_normalizer = {module = "relational_no_coercion"}

Or per source:

[sources.my_source.schema]
json_normalizer = {module = "relational_no_coercion"}

Shorthand names (without dots) resolve to the dlt.common.normalizers.json namespace. You can also use full module paths for custom normalizers.

Tables and columns#

The key components of a schema are tables and columns. You can find a dictionary of tables in the tables key or via the tables property of the Schema object.

A table schema has the following properties:

1. name and description.
2. columns with a dictionary of table schemas.
3. write_disposition hint telling dlt how new data coming to the table is loaded.
4. schema_contract - describes a contract on the table.
5. parent is a part of the nested reference, defined on a nested table and points to the parent table.

The table schema is extended by the data normalizer. The standard data normalizer adds propagated columns to it.

A column schema contains the following properties:

1. name and description of a column in a table.

Data type information:

1. data_type with a column data type.
2. precision is a precision for text, timestamp, time, bigint, binary, and decimal types.
3. scale is a scale for the decimal type.
4. timezone is a flag indicating TZ aware or NTZ timestamp and time. The default value is true.
5. nullable tells if the column is nullable or not.
6. is_variant indicates that the column was generated as a variant of another column.

A column schema contains the following basic hints:

1. primary_key marks a column as part of the primary key.
2. unique indicates that the column is unique. On some destinations, this generates a unique index.
3. merge_key marks a column as part of the merge key used by merge load.

Hints below are used to create nested references:

1. row_key is a special form of primary key created by dlt to uniquely identify rows of data.
2. parent_key is a special form of foreign key used by nested tables to refer to parent tables.
3. root_key marks a column as part of the root key, which is a type of foreign key always referring to the root table.
4. _dlt_list_idx is an index on a nested list from which a nested table is created.

dlt lets you define additional performance hints:

1. partition marks a column to be used to partition data.
2. cluster marks a column to be used to cluster data.
3. sort marks a column as sortable/having order. On some destinations, this non-unique generates an index.

Variant columns#

Variant columns are generated by a normalizer when it encounters a data item with a type that cannot be coerced into an existing column.

How coercion works (default normalizer)#

When a value's Python type does not match the column's data type, the default relational normalizer attempts to coerce the value. For example, if a column is text and you send an int, the value is converted to a string. If a column is bigint and you send "200", it is parsed as an integer. This works for many common type pairs.

A variant column is created only when coercion fails — i.e., the types are truly incompatible. Let's consider our getting started example with a slightly different approach, where id is an integer type at the beginning:

data = [
  {"id": 1, "human_name": "Alice"}
]

Once the pipeline runs, we will have the following schema:

Now imagine the data has changed and the id field also contains strings:

data = [
  {"id": 1, "human_name": "Alice"},
  {"id": "idx-nr-456", "human_name": "Bob"}
]

The value "idx-nr-456" cannot be parsed as an integer, so dlt creates a variant column id__v_text:

On the other hand, if the id field was already a string, then introducing new data with id containing other types will not change the schema because they can be coerced to string.

Now go ahead and try to add a new record where id is a float number; you should see a new field id__v_double in the schema.

No-coercion normalizer#

If you use relational_no_coercion (see Data normalizer), every type mismatch produces a variant column. No cross-type conversion is attempted. For example, sending "200" to a bigint column creates col__v_text instead of parsing it as an integer. This gives you stricter schema control: the original values are always preserved exactly as received, and mismatches are immediately visible as separate columns.

Compound hints#

Compound hints are column-level properties that you can apply to multiple columns to define a composite (compound) value. Supported compound hints are:

• primary_key
• merge_key
• cluster
• partition

When passing these hints to a resource, keep in mind that:

1. Direct primary_key and merge_key hints override column-level hints when both are set#

In the example below, the primary_key="col_1" argument takes precedence over any primary_key hints defined in columns. As a result, only col_1 will be treated as the primary key.

Note that direct primary_key and merge_key hints are always authoritative within a single resource definition, even if they are set to an empty value (e.g. "" or []). In that case, the empty direct hint forces any column-level key hints to be ignored. In the following example, col_2 will not receive a primary key hint.

The same precedence rule applies to merge_key. It also applies when direct key and column-level hints are both provided via apply_hints. In the example below, only col_1 will be treated as the merge key.

2. Redefining hints on an already extracted resource replaces previous configuration#

If you redefine a compound hint for a resource that has already been extracted, the new hint configuration overwrites the previous one. This means that old and new compound hints are not merged.

In the example below, the resource is first defined with partition on col_2. After the first run, we update the resource hints and set partition on col_1 instead. On the next run, only col_1 will remain partitioned, and col_2 will no longer have the partition property.

Be aware that redefining primary_key or merge_key to an empty value on an extracted resource does not clear any key properties in the schema. In the example below, col_1 will remain the primary key.

3. Column-level compound hints via apply_hints are merged#

When you call apply_hints with column-level compound hints on a resource that has already been extracted, the new column hints are merged into the existing schema. In the example below, the first run defines col_2 as a primary key. After the run, we add a primary_key hint for col_1 via the columns argument of apply_hints. On the next run, both col_1 and col_2 are treated as primary keys.

4. Direct key hints via apply_hints replace existing key properties#

Unlike column-level hints, direct key hints provided through apply_hints are treated as authoritative. As a result, they replace any existing key configuration instead of being merged. In the example below, setting primary_key="col_1" via apply_hints replaces the previously defined primary key on col_2.

Data types#

wei is a datatype that tries to best represent native Ethereum 256-bit integers and fixed-point decimals. It works correctly on Postgres and BigQuery. All other destinations have insufficient precision.

json data type tells dlt to load that element as JSON or string and not attempt to flatten or create a nested table out of it. Note that structured types like arrays or maps are not supported by dlt at this point.

time data type is saved in the destination without timezone info; if timezone is included, time is converted to UTC and then to naive.

Handling of timestamp and time zones#

By default, dlt normalizes timestamps (tz-aware and naive) into time zone aware types in UTC timezone. Since 1.16.0, it fully honors the timezone boolean hint if set
explicitly on a column or by a source/resource. Normalizers do not infer this hint from data. The same rules apply for tabular data (arrow/pandas/polars) and Python objects:

Ultimately, the destination will interpret the timestamp values. Some destinations:

• do not support naive timestamps (i.e. BigQuery) and will interpret them as naive UTC by attaching UTC timezone
• do not support tz-aware timestamps (i.e. Dremio, Athena) and will strip timezones from timestamps being loaded
• do not store timezone at all and all timestamps are converted to UTC
• store timezone as column level property and internally convert timestamps to UTC (i.e. postgres)
• store timezone and offset (i.e. MSSQL). However, we could not find any destination that can read back the original timezones

dlt sets sessions to UTC timezone to minimize chances of erroneous conversion.

Handling precision#

The precision and scale are interpreted by the particular destination and are validated when a column is created. Destinations that do not support precision for a given data type will ignore it.

The precision for bigint is mapped to available integer types, i.e., TINYINT, INT, BIGINT. The default is 64 bits (8 bytes) precision (BIGINT).

Selected destinations honor precision hint on timestamp. Precision is a numeric value in range of 0 (seconds) to 9 (nanoseconds) and sets the fractional
number of seconds stored in a column. The default value is 6 (microseconds) which is Python datetime precision. postgres, duckdb, snowflake, synapse and mssql allow setting precision. Additionally, duckdb and filesystem (via parquet) allow for nanosecond precision if:

• you configure parquet version to 2.6
• you yield tabular data (arrow tables/pandas/polars). dlt coerces all Python datetime objects into pendulum with microsecond precision.

Handling nulls#

In general, destinations are responsible for NULL enforcement. dlt does not verify nullability of data in arrow tables and Python objects. Note that:

• there's an exception to that rule if a Python object (dict) contains explicit None for a non-nullable key. This check will be eliminated. Note that if a value
  for a key is not present at all, nullability check is not done
• nullability is checked by Arrow when saving parquet files. This is a new behavior and dlt normalizes it for older arrow versions.

Structured types#

dlt has experimental support for structured types that currently piggyback on json data type and may be set only by yielding arrow tables. dlt does not
evolve nested types and will not migrate destination schemas to match. Nested types are enabled for filesystem, iceberg, delta and lancedb destinations.

Table references#

dlt tables refer to other tables. It supports two types of such references:

1. Nested reference created automatically when nested data (i.e., a json document containing a nested list) is converted into relational form. These references use specialized column and table hints and are used, for example, when merging data.
2. Table references are optional, user-defined annotations that are not verified and enforced but may be used by downstream tools, for example, to generate automatic tests or models for the loaded data.

Nested references: root and nested tables#

When dlt normalizes nested data into a relational schema, it automatically creates root and nested tables and links them using nested references.

1. All tables receive a column with the row_key hint (named _dlt_id by default) to uniquely identify each row of data.
2. Nested tables receive a parent table hint with the name of the parent table. The root table does not have a parent hint defined.
3. Nested tables receive a column with the parent_key hint (named _dlt_parent_id by default) that refers to the row_key of the parent table.

parent + row_key + parent_key form a nested reference: from the nested table to the parent table and are extensively used when loading data. Both replace and merge write dispositions.

row_key is created as follows:

1. A random string on root tables, except for upsert, insert-only, and
   scd2 merge strategies, where it is a deterministic hash of the primary_key (or whole row, so-called content_hash, if PK is not defined).
2. A deterministic hash of parent_key, parent table name, and position in the list (_dlt_list_idx)
   for nested tables.

You are able to bring your own row_key by adding a _dlt_id column/field to your data (both root and nested). All data types with an equal operator are supported.

merge write disposition requires an additional nested reference that goes from nested to root table, skipping all parent tables in between. This reference is created by adding a column with a hint root_key (named _dlt_root_id by default) to nested tables.

Generate custom linking for nested tables#

Using nested_hints in @dlt.resource you can model your own relations between root and nested tables. You do that by specifying primary_key or merge_key on
a nested table.

In the above example we effectively convert customers__purchases table into a top level table that is linked to customers table id column with customer_id foreign key.

1. we declare compound primary key on purchases on (customer_id, id) columns
2. we add a mapping function that will push the customer id to purchases as customer_id
3. we declare table reference from purchases to customers (this is optional)
4. we set merge write disposition on purchases.

Here's resulting schema. Note that regular linking for nested tables was not generated. Instead customer__purchases table has compound
primary key, write disposition, load id but still receives data from purchases nested list.

tables:
  customers:
    columns:
      id:
        nullable: false
        primary_key: true
        data_type: bigint
      name:
        data_type: text
        nullable: true
      city:
        data_type: text
        nullable: true
      _dlt_id:
        data_type: text
        nullable: false
        unique: true
        row_key: true
      _dlt_load_id:
        data_type: text
        nullable: false
    write_disposition: merge
    resource: customers
  customers__purchases:
    columns:
      customer_id:
        data_type: bigint
        primary_key: true
        nullable: false
      id:
        nullable: false
        primary_key: true
        data_type: bigint
      name:
        data_type: text
        nullable: true
      price:
        data_type: decimal
        nullable: true
      _dlt_root_id:
        data_type: text
        nullable: false
        root_key: true
      _dlt_id:
        data_type: text
        nullable: false
        unique: true
        row_key: true
      _dlt_load_id:
        data_type: text
        nullable: false
    references:
    - referenced_table: customers
      columns:
      - customer_id
      referenced_columns:
      - id
    write_disposition: merge
    resource: customers

Table references#

You can annotate tables with table references. @dlt.resource implements references argument that declares table references. Those references
are not enforced by dlt. See example above.

Schema settings#

The settings section of the schema file lets you define various global rules that impact how tables
and columns are inferred from data. For example, you can assign a primary_key hint to all columns named id or force a timestamp data type on all columns containing timestamp with the use of a regex pattern.

Data type autodetectors#

You can define a set of functions that will be used to infer the data type of a column from a
value. The functions are run from top to bottom on the lists. Look in detections.py to see what is
available. The iso_timestamp detector that looks for ISO 8601 strings and converts them to timestamp
is enabled by default.

settings:
  detections:
    - timestamp
    - iso_timestamp
    - iso_date
    - large_integer
    - hexbytes_to_text
    - wei_to_double

Alternatively, you can add and remove detections from code:

  source = data_source()
  # remove iso time detector
  source.schema.remove_type_detection("iso_timestamp")
  # convert UNIX timestamp (float, within a year from NOW) into timestamp
  source.schema.add_type_detection("timestamp")

Above, we modify a schema that comes with a source to detect UNIX timestamps with the timestamp detector.

Column hint rules#

You can define global rules that will apply hints to newly inferred columns. These rules apply to normalized column names. You can use column names directly or with regular expressions. dlt matches the column names after they have been normalized with naming conventions.

By default, the schema adopts hint rules from the json(relational) normalizer to support correct hinting of columns added by the normalizer:

settings:
  default_hints:
    row_key:
      - _dlt_id
    parent_key:
      - _dlt_parent_id
    not_null:
      - _dlt_id
      - _dlt_root_id
      - _dlt_parent_id
      - _dlt_list_idx
      - _dlt_load_id
    unique:
      - _dlt_id
    root_key:
      - _dlt_root_id

Above, we require an exact column name match for a hint to apply. You can also use a regular expression (which we call SimpleRegex) as follows:

settings:
    partition:
      - re:_timestamp$

Above, we add a partition hint to all columns ending with _timestamp. You can do the same thing in the code:

  from dlt.common.schema.typing import TSimpleRegex

  source = data_source()
  # this will update existing hints with the hints passed
  source.schema.merge_hints({"partition": [TSimpleRegex("re:_timestamp$")]})

Preferred data types#

You can define rules that will set the data type for newly created columns. Put the rules under the preferred_types key of settings. On the left side, there's a rule on a column name; on the right side is the data type. You can use column names directly or with regular expressions. dlt matches the column names after they have been normalized with naming conventions.

Example:

settings:
  preferred_types:
    re:timestamp: timestamp
    inserted_at: timestamp
    created_at: timestamp
    updated_at: timestamp

Above, we prefer the timestamp data type for all columns containing the timestamp substring and define a few exact matches, i.e., created_at.
Here's the same thing in code:

  source = data_source()
  source.schema.update_preferred_types(
    {
      TSimpleRegex("re:timestamp"): "timestamp",
      TSimpleRegex("inserted_at"): "timestamp",
      TSimpleRegex("created_at"): "timestamp",
      TSimpleRegex("updated_at"): "timestamp",
    }
  )

Applying data types directly with @dlt.resource and apply_hints#

dlt offers the flexibility to directly apply data types and hints in your code, bypassing the need for importing and adjusting schemas. This approach is ideal for rapid prototyping and handling data sources with dynamic schema requirements.

Direct specification in @dlt.resource#

Directly define data types and their properties, such as nullability, within the @dlt.resource decorator. This eliminates the dependency on external schema files. For example:

@dlt.resource(name='my_table', columns={"my_column": {"data_type": "bool", "nullable": True}})
def my_resource():
    for i in range(10):
        yield {'my_column': i % 2 == 0}

This code snippet sets up a nullable boolean column named my_column directly in the decorator.

Using apply_hints#

When dealing with dynamically generated resources or needing to programmatically set hints, apply_hints is your tool. It's especially useful for applying hints across various collections or tables at once.

For example, to apply a json data type across all collections from a MongoDB source:

all_collections = ["collection1", "collection2", "collection3"] # replace with your actual collection names
source_data = mongodb().with_resources(*all_collections)

for col in all_collections:
    source_data.resources[col].apply_hints(columns={"column_name": {"data_type": "json"}})

pipeline = dlt.pipeline(
    pipeline_name="mongodb_pipeline",
    destination="duckdb",
    dataset_name="mongodb_data"
)
load_info = pipeline.run(source_data)

This example iterates through MongoDB collections, applying the json data type to a specified column, and then processes the data with pipeline.run.

View and print the schema#

To view and print the default schema in a clear YAML format, use the command:

pipeline.default_schema.to_pretty_yaml()

This can be used in a pipeline as:

# Create a pipeline
pipeline = dlt.pipeline(
               pipeline_name="chess_pipeline",
               destination='duckdb',
               dataset_name="games_data")

# Run the pipeline
load_info = pipeline.run(source)

# Print the default schema in a pretty YAML format
print(pipeline.default_schema.to_pretty_yaml())

This will display a structured YAML representation of your schema, showing details like tables, columns, data types, and metadata, including version, version_hash, and engine_version.

Export and import schema files#

Please follow the guide on how to adjust a schema to export and import YAML
schema files in your pipeline.

Attaching schemas to sources#

We recommend not creating schemas explicitly. Instead, users should provide a few global schema
settings and then let the table and column schemas be generated from the resource hints and the
data itself.

The dlt.source decorator accepts a schema instance that you can create yourself and modify in
whatever way you wish. The decorator also supports a few typical use cases:

Schema created implicitly by decorator#

If no schema instance is passed, the decorator creates a schema with the name set to the source name and
all the settings to default.

Automatically load schema file stored with source python module#

If no schema instance is passed, and a file with a name {source name}.schema.yaml exists in the
same folder as the module with the decorated function, it will be automatically loaded and used as
the schema.

This should make it easier to bundle a fully specified (or pre-configured) schema with a source.

Schema is modified in the source function body#

What if you can configure your schema or add some tables only inside your schema function, when, for example,
you have the source credentials and user settings available? You could, for example, add detailed
schemas of all the database tables when someone requests table data to be loaded. This information
is available only at the moment the source function is called.

Similarly to the source_state() and resource_state(), the source and resource function has the current
schema available via dlt.current.source_schema().

Example:

@dlt.source
def textual(nesting_level: int):
    # get the source schema from the `current` context
    schema = dlt.current.source_schema()
    # remove date detector
    schema.remove_type_detection("iso_timestamp")
    # convert UNIX timestamp (float, within a year from NOW) into timestamp
    schema.add_type_detection("timestamp")

    return dlt.resource([])