Skip to content

Latest commit

 

History

History
358 lines (272 loc) · 24.6 KB

README.md

File metadata and controls

358 lines (272 loc) · 24.6 KB
Raw

clickhouse dbt logo

dbt-clickhouse

This plugin ports dbt functionality to Clickhouse.

The plugin uses syntax that requires ClickHouse version 22.1 or newer. We do not test older versions of Clickhouse. We also do not currently test Replicated tables or the related ON CLUSTER functionality.

Installation

Use your favorite Python package manager to install the app from PyPI, e.g.

pip install dbt-core dbt-clickhouse

NOTE: Beginning in v1.8, dbt-core and adapters are decoupled. Therefore, the installation mentioned above explicitly includes both dbt-core and the desired adapter.If you use a version prior to 1.8.0 the pip installation command should look like this:

pip install dbt-clickhouse

Supported features

  • Table materialization
  • View materialization
  • Incremental materialization
  • Materialized View materializations (uses the TO form of MATERIALIZED VIEW, experimental)
  • Seeds
  • Sources
  • Docs generate
  • Tests
  • Snapshots
  • Most dbt-utils macros (now included in dbt-core)
  • Ephemeral materialization
  • Distributed table materialization (experimental)
  • Distributed incremental materialization (experimental)
  • Contracts

Usage Notes

SET Statement Warning

In many environments, using the SET statement to persist a ClickHouse setting across all DBT queries is not reliable and can cause unexpected failures. This is particularly true when using HTTP connections through a load balancer that distributes queries across multiple nodes (such as ClickHouse cloud), although in some circumstances this can also happen with native ClickHouse connections. Accordingly, we recommend configuring any required ClickHouse settings in the "custom_settings" property of the DBT profile as a best practice, instead of relying on a prehook "SET" statement as has been occasionally suggested.

Database

The dbt model relation identifier database.schema.table is not compatible with Clickhouse because Clickhouse does not support a schema. So we use a simplified approach schema.table, where schema is the Clickhouse database. Using the default database is not recommended.

Example Profile

Default values are in brackets:

your_profile_name:
  target: dev
  outputs:
    dev:
      type: clickhouse
      schema: [default] # ClickHouse database for dbt models

      # optional
      driver: [http] # http or native.  If not set this will be autodetermined based on port setting
      host: [localhost] 
      port: [8123]  # If not set, defaults to 8123, 8443, 9000, 9440 depending on the secure and driver settings 
      user: [default] # User for all database operations
      password: [<empty string>] # Password for the user
      cluster: [<empty string>] # If set, certain DDL/table operations will be executed with the `ON CLUSTER` clause using this cluster. Distributed materializations require this setting to work. See the following ClickHouse Cluster section for more details.
      verify: [True] # Validate TLS certificate if using TLS/SSL
      secure: [False] # Use TLS (native protocol) or HTTPS (http protocol)
      retries: [1] # Number of times to retry a "retriable" database exception (such as a 503 'Service Unavailable' error)
      compression: [<empty string>] # Use gzip compression if truthy (http), or compression type for a native connection
      connect_timeout: [10] # Timeout in seconds to establish a connection to ClickHouse
      send_receive_timeout: [300] # Timeout in seconds to receive data from the ClickHouse server
      cluster_mode: [False] # Use specific settings designed to improve operation on Replicated databases (recommended for ClickHouse Cloud)
      use_lw_deletes: [False] # Use the strategy `delete+insert` as the default incremental strategy.
      check_exchange: [True] # Validate that clickhouse support the atomic EXCHANGE TABLES command.  (Not needed for most ClickHouse versions)
      local_suffix: [_local] # Table suffix of local tables on shards for distributed materializations.
      local_db_prefix: [<empty string>] # Database prefix of local tables on shards for distributed materializations. If empty, it uses the same database as the distributed table.
      allow_automatic_deduplication: [False] # Enable ClickHouse automatic deduplication for Replicated tables
      tcp_keepalive: [False] # Native client only, specify TCP keepalive configuration. Specify custom keepalive settings as [idle_time_sec, interval_sec, probes].
      custom_settings: [{}] # A dictionary/mapping of custom ClickHouse settings for the connection - default is empty.
      
      # Native (clickhouse-driver) connection settings
      sync_request_timeout: [5] # Timeout for server ping
      compress_block_size: [1048576] # Compression block size if compression is enabled

Model Configuration

Option Description Default if any
engine The table engine (type of table) to use when creating tables MergeTree()
order_by A tuple of column names or arbitrary expressions. This allows you to create a small sparse index that helps find data faster. tuple()
partition_by A partition is a logical combination of records in a table by a specified criterion. The partition key can be any expression from the table columns.
sharding_key Sharding key determines the destination server when inserting into distributed engine table. The sharding key can be random or as an output of a hash function rand())
primary_key Like order_by, a ClickHouse primary key expression. If not specified, ClickHouse will use the order by expression as the primary key
unique_key A tuple of column names that uniquely identify rows. Used with incremental models for updates.
inserts_only If set to True for an incremental model, incremental updates will be inserted directly to the target table without creating intermediate table. It has been deprecated in favor of the append incremental strategy, which operates in the same way. If inserts_only is set, incremental_strategy is ignored.
incremental_strategy Incremental model update strategy: delete+insert, append, or insert_overwrite. See the following Incremental Model Strategies default
incremental_predicates Additional conditions to be applied to the incremental materialization (only applied to delete+insert strategy
settings A map/dictionary of "TABLE" settings to be used to DDL statements like 'CREATE TABLE' with this model
query_settings A map/dictionary of ClickHouse user level settings to be used with INSERT or DELETE statements in conjunction with this model
ttl A TTL expression to be used with the table. The TTL expression is a string that can be used to specify the TTL for the table.

Column Configuration

Option Description Default if any
codec A string consisting of arguments passed to CODEC() in the column's DDL. For example: codec: "Delta, ZSTD" will be interpreted as CODEC(Delta, ZSTD).

ClickHouse Cluster

The cluster setting in profile enables dbt-clickhouse to run against a ClickHouse cluster.

Effective Scope

if cluster is set in profile, on_cluster_clause now will return cluster info for:

  • Database creation
  • View materialization
  • Distributed materializations
  • Models with Replicated engines

table and incremental materializations with non-replicated engine will not be affected by cluster setting (model would be created on the connected node only).

Compatibility

If a model has been created without a cluster setting, dbt-clickhouse will detect the situation and run all DDL/DML without on cluster clause for this model.

A Note on Model Settings

ClickHouse has several types/levels of "settings". In the model configuration above, two types of these are configurable. settings means the SETTINGS clause used in CREATE TABLE/VIEW types of DDL statements, so this is generally settings that are specific to the specific ClickHouse table engine. The new query_settings is use to add a SETTINGS clause to the INSERT and DELETE queries used for model materialization (including incremental materializations). There are hundreds of ClickHouse settings, and it's not always clear which is a "table" setting and which is a "user" setting (although the latter are generally available in the system.settings table.) In general the defaults are recommended, and any use of these properties should be carefully researched and tested.

Known Limitations

  • Ephemeral models/CTEs don't work if placed before the "INSERT INTO" in a ClickHouse insert statement, see ClickHouse/ClickHouse#30323. This should not affect most models, but care should be taken where an ephemeral model is placed in model definitions and other SQL statements.

Incremental Model Strategies

As of version 1.3.2, dbt-clickhouse supports three incremental model strategies.

The Default (Legacy) Strategy

Historically ClickHouse has had only limited support for updates and deletes, in the form of asynchronous "mutations." To emulate expected dbt behavior, dbt-clickhouse by default creates a new temporary table containing all unaffected (not deleted, not changed) "old" records, plus any new or updated records, and then swaps or exchanges this temporary table with the existing incremental model relation. This is the only strategy that preserves the original relation if something goes wrong before the operation completes; however, since it involves a full copy of the original table, it can be quite expensive and slow to execute.

The Delete+Insert Strategy

ClickHouse added "lightweight deletes" as an experimental feature in version 22.8. Lightweight deletes are significantly faster than ALTER TABLE ... DELETE operations, because they don't require rewriting ClickHouse data parts. The incremental strategy delete+insert utilizes lightweight deletes to implement incremental materializations that perform significantly better than the "legacy" strategy. However, there are important caveats to using this strategy:

  • Lightweight deletes must be enabled on your ClickHouse server using the setting allow_experimental_lightweight_delete=1 or you must set use_lw_deletes=true in your profile (which will enable that setting for your dbt sessions)
  • Lightweight deletes are now production ready, but there may be performance and other problems on ClickHouse versions earlier than 23.3.
  • This strategy operates directly on the affected table/relation (with creating any intermediate or temporary tables), so if there is an issue during the operation, the data in the incremental model is likely to be in an invalid state
  • When using lightweight deletes, dbt-clickhouse enabled the setting allow_nondeterministic_mutations. In some very rare cases using non-deterministic incremental_predicates this could result in a race condition for the updated/deleted items (and related log messages in the ClickHouse logs). To ensure consistent results the incremental predicates should only include sub-queries on data that will not be modified during the incremental materialization.

The Append Strategy

This strategy replaces the inserts_only setting in previous versions of dbt-clickhouse. This approach simply appends new rows to the existing relation. As a result duplicate rows are not eliminated, and there is no temporary or intermediate table. It is the fastest approach if duplicates are either permitted in the data or excluded by the incremental query WHERE clause/filter.

The insert_overwrite Strategy (Experimental)

[IMPORTANT]
Currently, the insert_overwrite strategy is not fully functional with distributed materializations.

Performs the following steps:

  1. Create a staging (temporary) table with the same structure as the incremental model relation: CREATE TABLE <staging> AS <target>.
  2. Insert only new records (produced by SELECT) into the staging table.
  3. Replace only new partitions (present in the staging table) into the target table.

This approach has the following advantages:

  • It is faster than the default strategy because it doesn't copy the entire table.
  • It is safer than other strategies because it doesn't modify the original table until the INSERT operation completes successfully: in case of intermediate failure, the original table is not modified.
  • It implements "partitions immutability" data engineering best practice. Which simplifies incremental and parallel data processing, rollbacks, etc.

The strategy requires partition_by to be set in the model configuration. Ignores all other strategies-specific parameters of the model config.

Additional ClickHouse Macros

Model Materialization Utility Macros

The following macros are included to facilitate creating ClickHouse specific tables and views:

  • engine_clause -- Uses the engine model configuration property to assign a ClickHouse table engine. dbt-clickhouse uses the MergeTree engine by default.
  • partition_cols -- Uses the partition_by model configuration property to assign a ClickHouse partition key. No partition key is assigned by default.
  • order_cols -- Uses the order_by model configuration to assign a ClickHouse order by/sorting key. If not specified ClickHouse will use an empty tuple() and the table will be unsorted
  • primary_key_clause -- Uses the primary_key model configuration property to assign a ClickHouse primary key. By default, primary key is set and ClickHouse will use the order by clause as the primary key.
  • on_cluster_clause -- Uses the cluster profile property to add an ON CLUSTER clause to certain dbt-operations: distributed materializations, views creation, database creation.
  • ttl_config -- Uses the ttl model configuration property to assign a ClickHouse table TTL expression. No TTL is assigned by default.

s3Source Helper Macro

The s3source macro simplifies the process of selecting ClickHouse data directly from S3 using the ClickHouse S3 table function. It works by populating the S3 table function parameters from a named configuration dictionary (the name of the dictionary must end in s3). The macro first looks for the dictionary in the profile vars, and then in the model configuration. The dictionary can contain any of the following keys used to populate the parameters of the S3 table function:

Argument Name Description
bucket The bucket base url, such as https://datasets-documentation.s3.eu-west-3.amazonaws.com/nyc-taxi. https:// is assumed if no protocol is provided.
path The S3 path to use for the table query, such as /trips_4.gz. S3 wildcards are supported.
fmt The expected ClickHouse input format (such as TSV or CSVWithNames) of the referenced S3 objects.
structure The column structure of the data in bucket, as a list of name/datatype pairs, such as ['id UInt32', 'date DateTime', 'value String'] If not provided ClickHouse will infer the structure.
aws_access_key_id The S3 access key id.
aws_secret_access_key The S3 secret key.
role_arn The ARN of a ClickhouseAccess IAM role to use to securely access the S3 objects. See this documentation for more information.
compression The compression method used with the S3 objects. If not provided ClickHouse will attempt to determine compression based on the file name.
See the S3 test file for examples of how to use this macro.

Contracts and Constraints

Only exact column type contracts are supported. For example, a contract with a UInt32 column type will fail if the model returns a UInt64 or other integer type. ClickHouse also support only CHECK constraints on the entire table/model. Primary key, foreign key, unique, and column level CHECK constraints are not supported. (See ClickHouse documentation on primary/order by keys.)

Materialized Views (Experimental)

A materialized_view materialization should be a SELECT from an existing (source) table. The adapter will create a target table with the model name and a ClickHouse MATERIALIZED VIEW with the name <model_name>_mv. Unlike PostgreSQL, a ClickHouse materialized view is not "static" (and has no corresponding REFRESH operation). Instead, it acts as an "insert trigger", and will insert new rows into the target table using the defined SELECT "transformation" in the view definition on rows inserted into the source table. See the [test file] (https://github.com/ClickHouse/dbt-clickhouse/blob/main/tests/integration/adapter/materialized_view/test_materialized_view.py) for an introductory example of how to use this functionality.

Dictionary materializations (experimental)

See the tests in https://github.com/ClickHouse/dbt-clickhouse/blob/main/tests/integration/adapter/dictionary/test_dictionary.py for examples of how to implement materializations for ClickHouse dictionaries

Distributed materializations

Notes:

  • dbt-clickhouse queries now automatically include the setting insert_distributed_sync = 1 in order to ensure that downstream incremental materialization operations execute correctly. This could cause some distributed table inserts to run more slowly than expected.

Distributed table materialization

Distributed table created with following steps:

  1. Creates temp view with sql query to get right structure
  2. Create empty local tables based on view
  3. Create distributed table based on local tables.
  4. Data inserts into distributed table, so it is distributed across shards without duplicating.

Distributed table model example

{{
    config(
        materialized='distributed_table',
        order_by='id, created_at',
        sharding_key='cityHash64(id)',
        engine='ReplacingMergeTree'
    )
}}

select id, created_at, item from {{ source('db', 'table') }}

Generated migrations

CREATE TABLE db.table_local on cluster cluster
(
    `id` UInt64,
    `created_at` DateTime,
    `item` String
)
ENGINE = ReplacingMergeTree
ORDER BY (id, created_at)
SETTINGS index_granularity = 8192;


CREATE TABLE db.table on cluster cluster
(
    `id` UInt64,
    `created_at` DateTime,
    `item` String
)
ENGINE = Distributed('cluster', 'db', 'table_local', cityHash64(id));

Distributed incremental materialization

Incremental model based on the same idea as distributed table, the main difficulty is to process all incremental strategies correctly.

  1. The Append Strategy just insert data into distributed table.
  2. The Delete+Insert Strategy creates distributed temp table to work with all data on every shard.
  3. The Default (Legacy) Strategy creates distributed temp and intermediate tables for the same reason.

Only shard tables are replacing, because distributed table does not keep data. The distributed table reloads only when the full_refresh mode is enabled or the table structure may have changed.

Distributed incremental model example

{{
    config(
        materialized='distributed_incremental',
        engine='MergeTree',
        incremental_strategy='append',
        unique_key='id,created_at'
    )
}}

select id, created_at, item from {{ source('db', 'table') }}

Generated migrations

CREATE TABLE db.table_local on cluster cluster
(
    `id` UInt64,
    `created_at` DateTime,
    `item` String
)
ENGINE = MergeTree
SETTINGS index_granularity = 8192;


CREATE TABLE db.table on cluster cluster
(
    `id` UInt64,
    `created_at` DateTime,
    `item` String
)
ENGINE = Distributed('cluster', 'db', 'table_local', cityHash64(id));

Contributing

We welcome contributions from the community to help improve the dbt-ClickHouse adapter. Whether you’re fixing a bug, adding a new feature, or enhancing documentation, your efforts are greatly appreciated!

Please take a moment to read our Contribution Guide to get started. The guide provides detailed instructions on setting up your environment, running tests, and submitting pull requests.

Original Author

ClickHouse wants to thank @silentsokolov for creating this connector and for their valuable contributions.