ClickHouse/docs/en/engines/database-engines/materialized-mysql.md
Kseniia Sumarokova 764bb1181f
Merge pull request #33017 from stigsb/table-overrides-drop-add-column-support
Remove the possibility of adding columns with table overrides
2021-12-22 22:30:21 +03:00

13 KiB

toc_priority toc_title
29 MaterializedMySQL

[experimental] MaterializedMySQL

!!! warning "Warning" This is an experimental feature that should not be used in production.

Creates ClickHouse database with all the tables existing in MySQL, and all the data in those tables.

ClickHouse server works as MySQL replica. It reads binlog and performs DDL and DML queries.

Creating a Database

CREATE DATABASE [IF NOT EXISTS] db_name [ON CLUSTER cluster]
ENGINE = MaterializedMySQL('host:port', ['database' | database], 'user', 'password') [SETTINGS ...]
[TABLE OVERRIDE table1 (...), TABLE OVERRIDE table2 (...)]

Engine Parameters

  • host:port — MySQL server endpoint.
  • database — MySQL database name.
  • user — MySQL user.
  • password — User password.

Engine Settings

  • max_rows_in_buffer — Maximum number of rows that data is allowed to cache in memory (for single table and the cache data unable to query). When this number is exceeded, the data will be materialized. Default: 65 505.
  • max_bytes_in_buffer — Maximum number of bytes that data is allowed to cache in memory (for single table and the cache data unable to query). When this number is exceeded, the data will be materialized. Default: 1 048 576.
  • max_rows_in_buffers — Maximum number of rows that data is allowed to cache in memory (for database and the cache data unable to query). When this number is exceeded, the data will be materialized. Default: 65 505.
  • max_bytes_in_buffers — Maximum number of bytes that data is allowed to cache in memory (for database and the cache data unable to query). When this number is exceeded, the data will be materialized. Default: 1 048 576.
  • max_flush_data_time — Maximum number of milliseconds that data is allowed to cache in memory (for database and the cache data unable to query). When this time is exceeded, the data will be materialized. Default: 1000.
  • max_wait_time_when_mysql_unavailable — Retry interval when MySQL is not available (milliseconds). Negative value disables retry. Default: 1000.
  • allows_query_when_mysql_lost — Allows to query a materialized table when MySQL is lost. Default: 0 (false).
CREATE DATABASE mysql ENGINE = MaterializedMySQL('localhost:3306', 'db', 'user', '***')
     SETTINGS
        allows_query_when_mysql_lost=true,
        max_wait_time_when_mysql_unavailable=10000;

Settings on MySQL-server Side

For the correct work of MaterializedMySQL, there are few mandatory MySQL-side configuration settings that must be set:

  • default_authentication_plugin = mysql_native_password since MaterializedMySQL can only authorize with this method.
  • gtid_mode = on since GTID based logging is a mandatory for providing correct MaterializedMySQL replication.

!!! attention "Attention" While turning on gtid_mode you should also specify enforce_gtid_consistency = on.

Virtual Columns

When working with the MaterializedMySQL database engine, ReplacingMergeTree tables are used with virtual _sign and _version columns.

  • _version — Transaction counter. Type UInt64.
  • _sign — Deletion mark. Type Int8. Possible values:
    • 1 — Row is not deleted,
    • -1 — Row is deleted.

Data Types Support

MySQL ClickHouse
TINY Int8
SHORT Int16
INT24 Int32
LONG UInt32
LONGLONG UInt64
FLOAT Float32
DOUBLE Float64
DECIMAL, NEWDECIMAL Decimal
DATE, NEWDATE Date
DATETIME, TIMESTAMP DateTime
DATETIME2, TIMESTAMP2 DateTime64
ENUM Enum
STRING String
VARCHAR, VAR_STRING String
BLOB String
BINARY FixedString
BIT UInt64

Nullable is supported.

Other types are not supported. If MySQL table contains a column of such type, ClickHouse throws exception "Unhandled data type" and stops replication.

Specifics and Recommendations

Compatibility Restrictions

Apart of the data types limitations there are few restrictions comparing to MySQL databases, that should be resolved before replication will be possible:

  • Each table in MySQL should contain PRIMARY KEY.

  • Replication for tables, those are containing rows with ENUM field values out of range (specified in ENUM signature) will not work.

DDL Queries

MySQL DDL queries are converted into the corresponding ClickHouse DDL queries (ALTER, CREATE, DROP, RENAME). If ClickHouse cannot parse some DDL query, the query is ignored.

Data Replication

MaterializedMySQL does not support direct INSERT, DELETE and UPDATE queries. However, they are supported in terms of data replication:

  • MySQL INSERT query is converted into INSERT with _sign=1.

  • MySQL DELETE query is converted into INSERT with _sign=-1.

  • MySQL UPDATE query is converted into INSERT with _sign=-1 and INSERT with _sign=1 if the primary key has been changed, or INSERT with _sign=1 if not.

Selecting from MaterializedMySQL Tables

SELECT query from MaterializedMySQL tables has some specifics:

  • If _version is not specified in the SELECT query, the FINAL modifier is used, so only rows with MAX(_version) are returned for each primary key value.

  • If _sign is not specified in the SELECT query, WHERE _sign=1 is used by default. So the deleted rows are not included into the result set.

  • The result includes columns comments in case they exist in MySQL database tables.

Index Conversion

MySQL PRIMARY KEY and INDEX clauses are converted into ORDER BY tuples in ClickHouse tables.

ClickHouse has only one physical order, which is determined by ORDER BY clause. To create a new physical order, use materialized views.

Notes

  • Rows with _sign=-1 are not deleted physically from the tables.
  • Cascade UPDATE/DELETE queries are not supported by the MaterializedMySQL engine, as they are not visible in the MySQL binlog.
  • Replication can be easily broken.
  • Manual operations on database and tables are forbidden.
  • MaterializedMySQL is affected by the optimize_on_insert setting. Data is merged in the corresponding table in the MaterializedMySQL database when a table in the MySQL server changes.

Table Overrides

Table overrides can be used to customize the ClickHouse DDL queries, allowing you to make schema optimizations for your application. This is especially useful for controlling partitioning, which is important for the overall performance of MaterializedMySQL.

These are the schema conversion manipulations you can do with table overrides for MaterializedMySQL:

CREATE DATABASE db_name ENGINE = MaterializedMySQL(...)
[SETTINGS ...]
[TABLE OVERRIDE table_name (
    [COLUMNS (
        [col_name [datatype] [ALIAS expr] [CODEC(...)] [TTL expr], ...]
        [INDEX index_name expr TYPE indextype[(...)] GRANULARITY val, ...]
        [PROJECTION projection_name (SELECT <COLUMN LIST EXPR> [GROUP BY] [ORDER BY]), ...]
    )]
    [ORDER BY expr]
    [PRIMARY KEY expr]
    [PARTITION BY expr]
    [SAMPLE BY expr]
    [TTL expr]
), ...]

Example:

CREATE DATABASE db_name ENGINE = MaterializedMySQL(...)
TABLE OVERRIDE table1 (
    COLUMNS (
        userid UUID,
        category LowCardinality(String),
        timestamp DateTime CODEC(Delta, Default)
    )
    PARTITION BY toYear(timestamp)
),
TABLE OVERRIDE table2 (
    COLUMNS (
        client_ip String TTL created + INTERVAL 72 HOUR
    )
    SAMPLE BY ip_hash
)

The COLUMNS list is sparse; existing columns are modified as specified, extra ALIAS columns are added. It is not possible to add ordinary or MATERIALIZED columns. Modified columns with a different type must be assignable from the original type. There is currently no validation of this or similar issues when the CREATE DATABASE query executes, so extra care needs to be taken.

You may specify overrides for tables that do not exist yet.

!!! warning "Warning" It is easy to break replication with table overrides if not used with care. For example:

* If an ALIAS column is added with a table override, and a column with the same name is later added to the source
    MySQL table, the converted ALTER TABLE query in ClickHouse will fail and replication stops.
* It is currently possible to add overrides that reference nullable columns where not-nullable are required, such as in
  `ORDER BY` or `PARTITION BY`. This will cause CREATE TABLE queries that will fail, also causing replication to stop.

Examples of Use

Queries in MySQL:

mysql> CREATE DATABASE db;
mysql> CREATE TABLE db.test (a INT PRIMARY KEY, b INT);
mysql> INSERT INTO db.test VALUES (1, 11), (2, 22);
mysql> DELETE FROM db.test WHERE a=1;
mysql> ALTER TABLE db.test ADD COLUMN c VARCHAR(16);
mysql> UPDATE db.test SET c='Wow!', b=222;
mysql> SELECT * FROM test;
┌─a─┬───b─┬─c────┐
│ 2 │ 222 │ Wow! │
└───┴─────┴──────┘

Database in ClickHouse, exchanging data with the MySQL server:

The database and the table created:

CREATE DATABASE mysql ENGINE = MaterializedMySQL('localhost:3306', 'db', 'user', '***');
SHOW TABLES FROM mysql;
┌─name─┐
│ test │
└──────┘

After inserting data:

SELECT * FROM mysql.test;
┌─a─┬──b─┐
│ 1 │ 11 │
│ 2 │ 22 │
└───┴────┘

After deleting data, adding the column and updating:

SELECT * FROM mysql.test;
┌─a─┬───b─┬─c────┐
│ 2 │ 222 │ Wow! │
└───┴─────┴──────┘

Original article