---
title: Naming convention
description: Control how dlt creates table, column and other identifiers
keywords: [identifiers, snake case, case sensitive, case insensitive, naming]
---

# Naming convention
dlt creates table and column identifiers from the data. The data source, i.e., a stream of JSON documents, may have identifiers (i.e., key names in a dictionary) with any Unicode characters, of any length, and naming style. On the other hand, destinations require that you follow strict rules when you name tables, columns, or collections.
A good example is [Redshift](../dlt-ecosystem/destinations/redshift.md) that accepts case-insensitive alphanumeric identifiers with a maximum of 127 characters.

`dlt` groups tables from a single [source](source.md) in a [schema](schema.md). Each schema defines a **naming convention** that tells `dlt` how to translate identifiers to the
namespace that the destination understands. Naming conventions are, in essence, functions that map strings from the source identifier format into the destination identifier format. For example, our **snake_case** (default) naming convention will translate the `DealFlow` source identifier into the `deal_flow` destination identifier.

You can pick which naming convention to use. `dlt` provides a few to [choose from](#available-naming-conventions). You can [easily add your own](#write-your-own-naming-convention) as well.

:::tip
The standard behavior of `dlt` is to **use the same naming convention for all destinations** so users always see the same table and column names in their databases.
:::

### Use default naming convention (snake_case)
**snake_case** is a case-insensitive naming convention, converting source identifiers into lower-case snake case identifiers with a reduced alphabet.

- Spaces around identifiers are trimmed.
- Keeps ASCII alphanumerics and underscores, replaces all other characters with underscores (with the exceptions below).
- Replaces `+` and `*` with `x`, `-` with `_`, `@` with `a`, and `|` with `l`.
- Prepends `_` if the name starts with a number.
- Multiples of `_` are converted into a single `_`.
- Replaces all trailing `_` with `x`.

Uses `__` as a nesting separator for tables and flattened column names.

:::tip
If you do not like **snake_case**, your next safe option is **sql_ci**, which generates SQL-safe, lowercase, case-insensitive identifiers without any other transformations. To permanently change the default naming convention on a given machine:
1. Set an environment variable `SCHEMA__NAMING` to `sql_ci_v1` OR
2. Add the following line to your global `config.toml` (the one in your home dir, i.e., `~/.dlt/config.toml`):
```toml
[schema]
naming="sql_ci_v1"
```
:::

## Source identifiers vs destination identifiers
### Pick the right identifier form when defining resources
`dlt` keeps source (not normalized) identifiers during data [extraction](../reference/explainers/how-dlt-works.md#extract) and translates them during [normalization](../reference/explainers/how-dlt-works.md#normalize). For you, it means:
1. If you write a [transformer](resource.md#process-resources-with-dlttransformer) or a [mapping/filtering function](resource.md#filter-transform-and-pivot-data), you will see the original data, without any normalization. Use the source identifiers to access the dicts!
2. If you define a `primary_key` or `cursor` that participates in [cursor field incremental loading](incremental/cursor.md), use the source identifiers (`dlt` uses them to inspect source data, `Incremental` class is just a filtering function).
3. When defining any other hints, i.e., `columns` or `merge_key`, you can pick source or destination identifiers. `dlt` normalizes all hints together with your data.
4. The `Schema` object (i.e., obtained from the pipeline or from `dlt` source via `discover_schema`) **always contains destination (normalized) identifiers**.

### Understand the identifier normalization
Identifiers are translated from source to destination form in the **normalize** step. Here's how `dlt` picks the naming convention:

* The default naming convention is **snake_case**.
* Each destination may define a preferred naming convention in [destination capabilities](destination.md#pass-additional-parameters-and-change-destination-capabilities). Some destinations (i.e., Weaviate) need a specialized naming convention and will override the default.
* You can [configure a naming convention explicitly](#configure-naming-convention). Such configuration overrides the destination settings.
* This naming convention is used when new schemas are created. It happens when the pipeline is run for the first time.
* Schemas preserve the naming convention when saved. Your running pipelines will maintain existing naming conventions if not requested otherwise.
* `dlt` applies the final naming convention in the `normalize` step. Jobs (files) in the load package now have destination identifiers. The pipeline schema is duplicated, locked, and saved in the load package and will be used by the destination.

:::warning
If you change the naming convention and `dlt` detects a change in the destination identifiers for tables/collections/files that already exist and store data, the normalize process will fail. This prevents an unwanted schema migration. New columns and tables will be created for identifiers that changed.
:::

### Case-sensitive and insensitive destinations
Naming conventions declare if the destination identifiers they produce are case-sensitive or insensitive. This helps `dlt` to [generate case-sensitive / insensitive identifiers for the destinations that support both](destination.md#control-how-dlt-creates-table-column-and-other-identifiers). For example, if you pick a case-insensitive naming like **snake_case** or **sql_ci_v1**, with Snowflake, `dlt` will generate all uppercase identifiers that Snowflake sees as case-insensitive. If you pick a case-sensitive naming like **sql_cs_v1**, `dlt` will generate quoted case-sensitive identifiers that preserve identifier capitalization.

Note that many destinations are exclusively case-insensitive, of which some preserve the casing of identifiers (i.e., **duckdb**) and some will case-fold identifiers when creating tables (i.e., **Redshift**, **Athena** do lowercase on the names). `dlt` is able to detect resulting identifier [collisions](#avoid-identifier-collisions) and stop the load process before data is mangled.

### Identifier shortening
Identifier shortening happens during normalization. `dlt` takes the maximum length of the identifier from the destination capabilities and will trim the identifiers that are too long. The default shortening behavior generates short deterministic hashes of the source identifiers and places them in the middle of the destination identifier. This (with a high probability) avoids shortened identifier collisions.

### Compound (flattened) identifiers
`dlt` combines several identifiers in order to name nested tables and flattened columns. For example:
```json
{
  "column":
    {
      "value": 1
    }
}
```
generates flattened column name `column__value`. Where `__` is a path separator (in **snake case**). Each component in the combined identifier is normalized
separately and shortened as a whole.

:::note
Combined identifier is also a valid single identifier. Starting from
`dlt` version above 1.4.0 normalization is fully idempotent and normalized
`column__value` will be still `column__value`.
:::

:::warning
Previously double underscores were contracted into single underscore. That
prevented using data loaded by `dlt` as a data source without identifier modifications. `dlt` maintains backward compatibility for version >1.4.0 as follows:

* All schemas stored locally or at destination will be migrated to backward compatible mode by setting a flag `use_break_path_on_normalize` i.e.:
```yaml
normalizers:
  names: dlt.common.normalizers.names.snake_case
  use_break_path_on_normalize: true
  json:
    module: dlt.common.normalizers.json.relational
```
* Backward compatible behavior may be explicitly enabled by setting
`SCHEMA__USE_BREAK_PATH_ON_NORMALIZE` to `TRUE` or via `config.toml`:
```toml
[schema]
use_break_path_on_normalize=true
```
:::

### 🚧 [WIP] Name convention changes are lossy
`dlt` does not store the source identifiers in the schema so when the naming convention changes (or we increase the maximum identifier length), it is not able to generate a fully correct set of new identifiers. Instead, it will re-normalize already normalized identifiers. We are currently working to store the full identifier lineage - source identifiers will be stored and mapped to the destination in the schema.

## Pick your own naming convention

### Configure naming convention
You can use `config.toml`, environment variables, or any other configuration provider to set the naming convention name. The configured naming convention **overrides all other settings**:
- Changes the naming convention stored in the already created schema.
- Overrides the destination capabilities preference.
```toml
[schema]
naming="sql_ci_v1"
```
The configuration above will request **sql_ci_v1** for all pipelines (schemas). An environment variable `SCHEMA__NAMING` set to `sql_ci_v1` has the same effect.

You have the option to set the naming convention per source:
```toml
[sources.zendesk]
config="prop"
[sources.zendesk.schema]
naming="sql_cs_v1"
[sources.zendesk.credentials]
password="pass"
```
The snippet above demonstrates how to apply a certain naming for an example `zendesk` source.

You can set the naming convention in your code via destination factory. This will overwrite destination's preferred convention and make it
a default one for the whole pipeline:

```py
import dlt

dest_ = dlt.destinations.postgres(naming_convention="sql_cs_v1")
```

You can use naming conventions that you created yourself or got from other users. In that case, you should pass a full Python import path to the [module that contains the naming convention](#write-your-own-naming-convention):

### Available naming conventions
You can pick from a few built-in naming conventions.

* `snake_case` - the default.
* `duck_case` - case-sensitive, allows all Unicode characters like emoji 💥.
* `direct` - case-sensitive, allows all Unicode characters, does not contract underscores.
* `sql_cs_v1` - case-sensitive, generates SQL-safe identifiers.
* `sql_ci_v1` - case-insensitive, generates SQL-safe lowercase identifiers.
* `s3_tables` - extends `snake_case` to comply with S3 Tables [naming rules](https://docs.aws.amazon.com/AmazonS3/latest/userguide/s3-tables-buckets-naming.html#naming-rules-table).

:::note
When using the `s3_tables` naming convention, any leading underscores are removed from dataset and table names. Hence, `dlt` tables will be prefixed with `dlt_` instead of the usual `_dlt_`. For example, `_dlt_loads` will be `dlt_loads`.
:::

### Ignore naming convention for `dataset_name`
You control the dataset naming normalization separately. Set `enable_dataset_name_normalization` to `false` to ignore the naming convention for `dataset_name`:

```toml
[destination.snowflake]
enable_dataset_name_normalization=false
```

In that case, the `dataset_name` would be preserved the same as it was set in the pipeline:
```py
import dlt

pipeline = dlt.pipeline(dataset_name="MyCamelCaseName")
```

The default value for the `enable_dataset_name_normalization` configuration option is `true`.
:::note
The same setting would be applied to [staging dataset](../dlt-ecosystem/staging#staging-dataset). Thus, if you set `enable_dataset_name_normalization` to `false`, the staging dataset name would also **not** be normalized.
:::

:::warning
Depending on the destination, certain names may not be allowed. To ensure your dataset can be successfully created, use the default normalization option.
:::

## Avoid identifier collisions
`dlt` detects various types of identifier collisions and ignores the others.
1. dlt detects collisions if a case-sensitive naming convention is used on a case-insensitive destination.
2. dlt detects collisions if a change of naming convention changes the identifiers of tables already created in the destination.
3. dlt detects collisions when the naming convention is applied to column names of arrow tables.

`dlt` will not detect a collision when normalizing source data. If you have a dictionary, keys will be merged if they collide after being normalized.
You can create a custom naming convention that does not generate collisions on data, see examples below.

## Write your own naming convention

Custom naming conventions are classes that derive from `NamingConvention`, which you can import from `dlt.common.normalizers.naming`. We recommend the following module layout:
1. Each naming convention resides in a separate Python module (file).
2. The class is always named `NamingConvention`.

In that case, you can use a fully qualified module name in [schema configuration](#configure-naming-convention) or pass the module fully qualified name [explicitly](#configure-naming-convention).

```toml
[schema]
naming="tests.common.cases.normalizers.sql_upper"
```
`dlt` will import `tests.common.cases.normalizers.sql_upper` and use the `NamingConvention` class found in it as the naming convention.

:::tip
Do not pass custom naming convention as modules if you do it explicitly. We recommend pattern below:
```py
import dlt

dest_ = dlt.destinations.postgres(naming_convention="my_package.sql_cs_latin2")
```

⛔ avoid this or you may get pickle errors i.e. when using parallel normalization:
```py
import dlt

import my_package.sql_cs_latin2  # type: ignore[import-not-found]

dest_ = dlt.destinations.postgres(naming_convention=my_package.sql_cs_latin2)
```
:::


We include [two examples](../examples/custom_naming) of naming conventions that you may find useful:

1. A variant of `sql_ci` that generates identifier collisions with a low (user-defined) probability by appending a deterministic tag to each name.
2. A variant of `sql_cs` that allows for LATIN (i.e., umlaut) characters.

:::note
Note that the fully qualified name of your custom naming convention will be stored in the schema, and dlt will attempt to import it when the schema is loaded from storage.
You should distribute your custom naming conventions with your pipeline code or via a pip package from which it can be imported.
:::