ClickHouse#

Install dlt with ClickHouse#

To install the DLT library with ClickHouse dependencies:

pip install "dlt[clickhouse]"

Setup guide#

1. Initialize the dlt project#

Let's start by initializing a new dlt project as follows:

dlt init chess clickhouse

dlt init command will initialize your pipeline with chess as the source and ClickHouse as the destination.

The above command generates several files and directories, including .dlt/secrets.toml and a requirements file for ClickHouse. You can install the necessary dependencies specified in the requirements file by executing it as follows:

pip install -r requirements.txt

or with pip install "dlt[clickhouse]", which installs the dlt library and the necessary dependencies for working with ClickHouse as a destination.

2. Setup ClickHouse database#

To load data into ClickHouse, you need to create a ClickHouse database. While we recommend asking our GPT-4 assistant for details, we've provided a general outline of the process below:

1. You can use an existing ClickHouse database or create a new one.

2. To create a new database, connect to your ClickHouse server using the clickhouse-client command line tool or a SQL client of your choice.

3. Run the following SQL commands to create a new database, user, and grant the necessary permissions:

   ⧉

   CREATE DATABASE IF NOT EXISTS dlt;
   CREATE USER dlt IDENTIFIED WITH sha256_password BY 'Dlt*12345789234567';
   GRANT CREATE, ALTER, SELECT, DELETE, DROP, TRUNCATE, OPTIMIZE, SHOW, INSERT, dictGet ON dlt.* TO dlt;
   GRANT SELECT ON INFORMATION_SCHEMA.COLUMNS TO dlt;
   GRANT CREATE TEMPORARY TABLE, S3 ON *.* TO dlt;
   

3. Add credentials#

1. Next, set up the ClickHouse credentials in the .dlt/secrets.toml file as shown below:

   ⧉

   [destination.clickhouse.credentials]
   database = "dlt" # The database name you created.
   username = "dlt" # ClickHouse username, default is usually "default".
   password = "Dlt*12345789234567" # ClickHouse password, if any.
   host = "localhost" # ClickHouse server host.
   port = 9000 # ClickHouse native TCP protocol port, default is 9000.
   http_port = 8443 # ClickHouse HTTP port, default is 9000.
   secure = 1 # Set to 1 if using HTTPS, else 0.
   

   :::info Network Ports
   The http_port parameter specifies the port number to use when connecting to the ClickHouse server's HTTP interface.
   The default non-secure HTTP port for ClickHouse is 8123.
   This is different from the default port 9000, which is used for the native TCP protocol.

   You must additionally set http_port if you are not using external staging (i.e., you don't set the staging destination parameter in your pipeline). This is because dlt's built-in ClickHouse local storage staging uses the clickhouse-connect library, which communicates with ClickHouse over HTTP.

   Make sure your ClickHouse server is configured to accept HTTP connections on the port specified by http_port. For example:

   • If you set http_port = 8123 (default non-secure HTTP port), then ClickHouse should be listening for HTTP requests on port 8123.
   • If you set http_port = 8443, then ClickHouse should be listening for secure HTTPS requests on port 8443.

   If you're using external staging, you can omit the http_port parameter, since clickhouse-connect will not be used in this case.

   For local development and testing with ClickHouse running locally, it is recommended to use the default non-secure HTTP port 8123 by setting http_port=8123 or omitting the parameter.

   Please see the ClickHouse network port documentation for further reference.
   :::

2. You can pass a database connection string similar to the one used by the clickhouse-driver library. The credentials above will look like this:

   ⧉

   # Keep it at the top of your TOML file, before any section starts
   destination.clickhouse.credentials="clickhouse://dlt:Dlt*12345789234567@localhost:9000/dlt?secure=1"
   

3. Add configuration options#

You can set the following configuration options in the .dlt/secrets.toml file:

[destination.clickhouse]
dataset_table_separator = "___" # The default separator for dataset table names from the dataset.
table_engine_type = "merge_tree" # The default table engine to use.
dataset_sentinel_table_name = "dlt_sentinel_table" # The default name for sentinel tables.
staging_use_https = true # Whether to connect to the staging bucket via https (defaults to True)
select_sequential_consistency = 1 # Ensures read-after-write consistency on ClickHouse Cloud (defaults to 1)

Write disposition#

All write dispositions are supported.

Data loading#

Data is loaded into ClickHouse using the most efficient method depending on the data source:

• For local files, the clickhouse-connect library is used to directly load files into ClickHouse tables using the INSERT command.
• For files in remote storage like S3, Google Cloud Storage, or Azure Blob Storage, ClickHouse table functions like s3, gcs, and azureBlobStorage are used to read the files and insert the data into tables.

Read-after-write consistency#

By default, dlt sets select_sequential_consistency=1 on all ClickHouse connections. This ensures
that SELECT queries always see the results of preceding writes, even on ClickHouse Cloud
(SharedMergeTree) or self-hosted clusters where queries may be routed to different nodes. If you are
running read-only workloads and want to avoid the small metadata check overhead, set it to 0 in
your configuration:

[destination.clickhouse]
select_sequential_consistency = 0

Merge on replicated clusters#

When merging tables that have nested children, dlt creates short-lived temp tables to stage delete
keys and insert keys. The temp tables follow the configured table_engine_type so the staged rows replicate the same way the destination tables do. Single-table merges use lightweight DELETE directly
and don't need temp tables.

Datasets#

ClickHouse does not support multiple datasets in one database; dlt relies on datasets to exist for multiple reasons.
To make ClickHouse work with dlt, tables generated by dlt in your ClickHouse database will have their names prefixed with the dataset name, separated by
the configurable dataset_table_separator.
Additionally, a special sentinel table that doesn't contain any data will be created, so dlt knows which virtual datasets already exist in a
clickhouse
destination.

Supported file formats#

• JSONL is the preferred format for both direct loading and staging.
• Parquet is supported for both direct loading and staging.

The clickhouse destination has a few specific deviations from the default SQL destinations:

1. ClickHouse has an experimental object datatype, but we've found it to be a bit unpredictable, so the dlt clickhouse destination will load the json datatype to a text column.
   If you need
   this feature, get in touch with our Slack community, and we will consider adding it.
2. ClickHouse does not support the time datatype. Time will be loaded to a text column.
3. ClickHouse does not support the binary datatype. Binary will be loaded to a text column. When loading from JSONL, this will be a base64 string; when loading from parquet, this will be
   the binary object converted to text.
4. ClickHouse accepts adding columns to a populated table that aren’t null.
5. ClickHouse can produce rounding errors under certain conditions when using the float/double datatype. Make sure to use decimal if you can’t afford to have rounding errors. Loading the value 12.7001 to a double column with the loader file format jsonl set will predictably produce a rounding error, for example.

Supported column hints#

ClickHouse supports the following column hints:

• primary_key - marks the column as part of the primary key. Multiple columns can have this hint to create a composite primary key.

Choosing a table engine#

dlt defaults to the MergeTree table engine. You can specify an alternate table engine in two ways:

Setting a default table engine in the configuration#

You can set a default table engine for all resources and dlt tables by adding the table_engine_type parameter to your ClickHouse credentials in the .dlt/secrets.toml file:

[destination.clickhouse]
# ... (other configuration options)
table_engine_type = "merge_tree" # The default table engine to use.

Setting the table engine for specific resources#

You can also set the table engine for specific resources using the clickhouse_adapter, which will override the default engine set in .dlt/secrets.toml for that resource:

from dlt.destinations.adapters import clickhouse_adapter

@dlt.resource()
def my_resource():
    ...

clickhouse_adapter(my_resource, table_engine_type="merge_tree")

Supported values for table_engine_type are:

• merge_tree (default) - creates tables using the MergeTree engine, suitable for most use cases. Learn more about MergeTree.
• shared_merge_tree - creates tables using the SharedMergeTree engine, optimized for cloud-native environments with shared storage. This table is only available on ClickHouse Cloud, and it is the default selection if merge_tree is selected. Learn more about SharedMergeTree.
• replicated_merge_tree - creates tables using the ReplicatedMergeTree engine, which supports data replication across multiple nodes for high availability. Learn more about ReplicatedMergeTree. This defaults to shared_merge_tree on ClickHouse Cloud.
• replacing_merge_tree - creates tables using the ReplacingMergeTree engine, which deduplicates rows with the same sorting key during background merges. See native deduplication with ReplacingMergeTree below. Learn more about ReplacingMergeTree.
• Experimental support for the Log engine family with stripe_log and tiny_log.

For local development and testing with ClickHouse running locally, the MergeTree engine is recommended.

Native deduplication with ReplacingMergeTree#

When loading data with append write disposition, you can use ClickHouse's ReplacingMergeTree engine to handle deduplication and deletes at the storage engine level — without using dlt's merge write disposition. This is useful when you want ClickHouse to manage upserts natively, for example when appending CDC events or syncing data from an external source where rows may repeat.

To enable this, combine replacing_merge_tree with dlt column hints:

• primary_key — defines the dedup key. Rows with the same key are considered duplicates.
• dedup_sort (optional) — specifies the version column. ClickHouse keeps the row with the highest value in this column. Must be a non-nullable UInt*, Date, or DateTime type. If omitted, ClickHouse uses last-write-wins semantics (the most recently inserted row survives).
• hard_delete (optional) — specifies a UInt8 deletion marker column. Rows where this column is 1 are filtered out when querying with FINAL. To physically remove deleted rows during background merges, set the clean_deleted_rows table setting (see example below). Requires dedup_sort to be set (ClickHouse constraint: is_deleted requires a version column).

dlt automatically wires these hints into the ReplacingMergeTree engine parameters.

:::note Dedup key and sorting key
In ClickHouse, ReplacingMergeTree deduplicates based on the ORDER BY columns, not PRIMARY KEY. When no explicit sorting key is set via clickhouse_adapter(sort=...), ClickHouse defaults ORDER BY to PRIMARY KEY, so primary_key effectively controls deduplication. However, if you set a custom sorting key with clickhouse_adapter(sort=...), that sorting key becomes the dedup key and may override your primary_key intent. Avoid combining sort with replacing_merge_tree unless you want the sorting key to define which rows are duplicates.
:::

import dlt
from dlt.destinations.adapters import clickhouse_adapter

@dlt.resource(
    write_disposition="append",
    primary_key="id",
    columns={
        "version": {"dedup_sort": "desc", "nullable": False},
        "is_deleted": {"hard_delete": True, "nullable": False},
    },
)
def events(data):
    yield from data

clickhouse_adapter(
    events,
    table_engine_type="replacing_merge_tree",
    settings={"clean_deleted_rows": "Always"},
)

pipeline = dlt.pipeline("my_pipeline", destination="clickhouse")

# first load
pipeline.run(events([
    {"id": 1, "name": "Alice", "version": 1, "is_deleted": False},
    {"id": 2, "name": "Bob", "version": 1, "is_deleted": False},
]))

# second load: update Alice, delete Bob
pipeline.run(events([
    {"id": 1, "name": "Alice Updated", "version": 2, "is_deleted": False},
    {"id": 2, "name": "Bob", "version": 2, "is_deleted": True},
]))

This generates ENGINE = ReplacingMergeTree(version, is_deleted), so ClickHouse keeps only the row with the highest version per id. The clean_deleted_rows setting tells ClickHouse to physically remove rows marked as deleted during background merges.

:::caution Data consistency
ReplacingMergeTree deduplicates during background merges, which happen asynchronously. Until a merge runs, duplicate rows may be visible in queries. To get consistent (deduplicated) results, use the FINAL modifier:

SELECT * FROM my_table FINAL

Without clean_deleted_rows, rows with is_deleted=1 are only filtered out by FINAL queries but remain physically stored. With clean_deleted_rows='Always', ClickHouse removes them during merges.

This is different from dlt's merge write disposition, which deduplicates immediately during loading via explicit SQL statements. Choose ReplacingMergeTree when eventual consistency is acceptable and you want ClickHouse to manage deduplication natively.
:::

Sorting and partitioning#

You can use the clickhouse_adapter to specify a sorting and/or partition key:

from dlt.destinations.adapters import clickhouse_adapter

@dlt.resource
def my_resource():
    ...

clickhouse_adapter(
   my_resource,
   sort = sorting_key,
   partition = partition_key
)

sort and partition are used to generate the ORDER BY and PARTITION BY clauses of the table creation statement, and they accept either a sequence of column names or a SQL expression:

1. sequence of column names: recommended if column transformations are not required
2. SQL expression: use if column transformations are required

The two examples below show both approaches to adapt my_resource:

@dlt.resource(
   columns={
      "timestamp": {"nullable": False},
      "town": {"nullable": False},
      "street": {"nullable": False}
   }
)
def my_resource():
    ...

Example 1: sequence of column names#

clickhouse_adapter(
   my_resource,
   sort=["timestamp", "street"],
   partition=["town"]
)

Corresponding SQL:

CREATE TABLE ...
...
ORDER BY (timestamp, street)
PARTITION by town

Example 2: SQL expression#

clickhouse_adapter(
   my_resource,
   sort="(upper(town), street)",
   partition="toYYYYMMDD(timestamp)"
)

Corresponding SQL:

CREATE TABLE ...
...
ORDER BY (upper(town), street)
PARTITION BY toYYYYMMDD(timestamp)

Usage notes#

SQL expressions are used as is to generate the SQL clauses. Hence, when providing a SQL expression, use normalized column names:

@dlt.resource(columns={"TIMESTAMP": {"nullable": False}}) # non-normalized column name (upper case)
def my_resource():
    ...

clickhouse_adapter(my_resource, partition="toYYYYMMDD(timestamp)") # RIGHT: normalized column name (lower case)

clickhouse_adapter(my_resource, partition="toYYYYMMDD(TIMESTAMP)") # WRONG: non-normalized column name (lower case)

sort and partition column hints#

dlt automatically creates sort/partition column hints for columns present in the sort/partition value provided to clickhouse_adapter (when this value is a SQL expression, we parse it to extract the column names).

Although it's possible to set sort/partition column hints directly, we recommend using clickhouse_adapter instead.

If you still choose to set sort/partition column hints yourself, know that:

• columns are added to the ORDER BY/PARTITION BY clause in order of appearance in the schema
• they may be overridden/removed if you also use clickhouse_adapter: the adapter takes precedence, and it will set column hints in accordance with the values provided to its sort/partition parameters

MergeTree table settings#

Use the settings parameter of the clickhouse_adapter to specify MergeTree settings for the table:

from dlt.destinations.adapters import clickhouse_adapter

@dlt.resource
def my_resource():
    ...

clickhouse_adapter(
   my_resource,
   settings = {
      "allow_nullable_key": True,
      "max_suspicious_broken_parts": 500,
      "deduplicate_merge_projection_mode": "ignore"
   }
)

The settings parameter is used to generate the SETTINGS clause of the table creation statement:

CREATE TABLE ...
...
SETTINGS allow_nullable_key = true, max_suspicious_broken_parts = 500, deduplicate_merge_projection_mode = 'ignore'

Column codecs#

Use the codecs parameter of the clickhouse_adapter to specify codecs for the table's columns:

from dlt.destinations.adapters import clickhouse_adapter

@dlt.resource
def my_resource():
    ...

clickhouse_adapter(
   my_resource,
   codecs = {
      "town": "ZSTD(3)",
      "number": "Delta, ZSTD(2)"
   }
)

The codecs are used in the table creation statement as follows:

CREATE TABLE my_resource
(
   `town` String CODEC(ZSTD(3)),
   `street` String, -- no codec specified
   `number` Int64 CODEC(Delta, ZSTD(2))
)
...

Staging support#

ClickHouse supports Amazon S3, Google Cloud Storage, and Azure Blob Storage as file staging destinations.

dlt will upload Parquet or JSONL files to the staging location and use ClickHouse table functions to load the data directly from the staged files.

Please refer to the filesystem documentation to learn how to configure credentials for the staging destinations:

• Amazon S3
• Google Cloud Storage
• Azure Blob Storage

To run a pipeline with staging enabled:

pipeline = dlt.pipeline(
  pipeline_name='chess_pipeline',
  destination='clickhouse',
  staging='filesystem', # add this to activate staging
  dataset_name='chess_data'
)

Using Google Cloud or S3-compatible storage as a staging area#

dlt supports using S3-compatible storage services, including Google Cloud Storage (GCS), as a staging area when loading data into ClickHouse.
This is handled automatically by
ClickHouse's GCS table function, which dlt uses under the hood.

The ClickHouse GCS table function only supports authentication using Hash-based Message Authentication Code (HMAC) keys, which is compatible with the Amazon S3 API.
To enable this, GCS provides an S3
compatibility mode that emulates the S3 API, allowing ClickHouse to access GCS buckets via its S3 integration.

For detailed instructions on setting up S3-compatible storage with dlt, including AWS S3, MinIO, and Cloudflare R2, refer to
the dlt documentation on filesystem destinations.

To set up GCS staging with HMAC authentication in dlt:

1. Create HMAC keys for your GCS service account by following the Google Cloud guide.

2. Configure the HMAC keys (aws_access_key_id and aws_secret_access_key) as well as endpoint_url in your dlt project's ClickHouse destination settings in config.toml, similar to how you would configure AWS S3 credentials:

[destination.filesystem]
bucket_url = "s3://my_awesome_bucket"

[destination.filesystem.credentials]
aws_access_key_id = "JFJ$$*f2058024835jFffsadf"
aws_secret_access_key = "DFJdwslf2hf57)%$02jaflsedjfasoi"
project_id = "my-awesome-project"
endpoint_url = "https://storage.googleapis.com"

When using S3 for a staging area you can alternatively have ClickHouse authenticate using Role-based access with the
supported extra_credentials argument by setting this with the destination credentials:

import dlt
from dlt.destinations import clickhouse

destination = clickhouse(
  credentials={
    # Other credentials you need
    "s3_extra_credentials": {
      'role_arn': 'arn:your:role' # The AWS Role assumed by ClickHouse
    }
  }
)

pipeline = dlt.pipeline(
  pipeline_name='chess_pipeline',
  destination=destination,
  staging='filesystem', # add this to activate staging
  dataset_name='chess_data'
)

dbt support#

Integration with dbt is generally supported via dbt-clickhouse but not tested by us. Note how
we support datasets by prefixing the table names. You should take it into account when writing models (or use empty dataset to avoid prefixing).

Syncing of dlt state#

This destination fully supports dlt state sync.