TiDB Data Migration (DM) Best Practices

TiDB Data Migration (DM) is a data migration tool developed by PingCAP. It supports full and incremental data migration from MySQL-compatible databases such as MySQL, Percona MySQL, MariaDB, Amazon RDS for MySQL, and Amazon Aurora into TiDB.

You can use DM in the following scenarios:

• Perform full and incremental data migration from a single MySQL-compatible database instance to TiDB
• Migrate and merge MySQL shards of small datasets (less than 1 TiB) to TiDB
• In the data hub scenario, such as the middle platform of business data, and real-time aggregation of business data, use DM as the middleware for data migration

This document introduces how to use DM in an elegant and efficient way, and how to avoid common mistakes when using DM.

Performance limitations

• DM supports managing 1000 work nodes simultaneously, and the maximum number of tasks is 600. To ensure the high availability of work nodes, you should reserve some work nodes as standby nodes. The recommended number of standby nodes is 20% to 50% of the number of the work nodes that have running migration tasks.
• A single work node can theoretically support replication QPS of up to 30K QPS/worker. It varies for different schemas and workloads. The ability to handle upstream binlogs is up to 20 MB/s/worker.
• If you want to use DM as a data replication middleware for long-term use, you need to carefully design the deployment architecture of DM components. For more information, see Deploy DM-master and DM-worker

Before data migration

Before data migration, the design of the overall solution is critical. The following sections describe best practices and scenarios from the business perspective and the implementation perspective.

Best practices for the business side

To distribute the workload evenly on multiple nodes, the design for the distributed database is different from traditional databases. The solution needs to ensure both low migration cost and logic correctness after migration. The following sections describe best practices before data migration.

Business impact of AUTO_INCREMENT in schema design

AUTO_INCREMENT in TiDB is compatible with AUTO_INCREMENT in MySQL. However, as a distributed database, TiDB usually has multiple computing nodes (entries for the client end). When the application data is written, the workload is evenly distributed. This leads to the result that when there is an AUTO_INCREMENT column in the table, the auto-increment IDs of the column might be inconsecutive. For more details, see AUTO_INCREMENT.

If your business has a strong dependence on auto-increment IDs, consider using the SEQUENCE function.

Usage of clustered indexes

When you create a table, you can declare that the primary key is either a clustered index or a non-clustered index. The following sections describe the pros and cons of each choice.

• Clustered indexes

  Clustered indexes use the primary key as the handle ID (row ID) for data storage. Querying using the primary key can avoid table lookup, which effectively improves the query performance. However, if the table is write-intensive and the primary key uses AUTO_INCREMENT, it is very likely to cause write hotspot problems, resulting in a mediocre performance of the cluster and the performance bottleneck of a single storage node.

• Non-clustered indexes + shard row id bit

  Using non-clustered indexes and shard row id bit, you can avoid the write hotspot problem when using AUTO_INCREMENT. However, table lookup in this scenario can affect the query performance when querying using the primary key.

• Clustered indexes + external distributed ID generators

  If you want to use clustered indexes and keep the IDs consecutive, consider using external distributed ID generators, such as the Snowflake algorithm and Leaf. The application program generates sequence IDs, which can guarantee that the IDs are consecutive to a certain extent. It also retains the benefits of using clustered indexes. But you need to customize the applications.

• Clustered indexes + AUTO_RANDOM

  This solution can retain the benefits of using clustered indexes and avoid the write hotspot problem. It requires less effort for customization. You can modify the schema attribute when you switch to use TiDB as the write database. In subsequent queries, if you have to use the ID column to sort data, you can use the AUTO_RANDOM ID column and left shift 5 bits to ensure the order of the query data. For example:

  CREATE TABLE t (a bigint PRIMARY KEY AUTO_RANDOM, b varchar(255));
  Select a, a<<5 ,b from t order by a <<5 desc
  

The following table summarizes the pros and cons of each solution.

Key points for MySQL shards

Splitting and merging

It is recommended that you use DM to migrate and merge MySQL shards of small datasets to TiDB.

Besides data merging, another typical scenario is data archiving. Data is constantly being written. As time goes by, large amounts of data gradually change from hot data to warm or even cold data. Fortunately, in TiDB, you can set different placement rules for data. The minimum granularity of the rules is a partition.

Therefore, it is recommended that for write-intensive scenarios, you need to evaluate from the beginning whether you need to archive data and store hot and cold data on different media separately. If you need to archive data, you can set the partitioning rules before migration (TiDB does not support Table Rebuild operations yet). It saves you from the need to recreate tables and import data in future.

The pessimistic mode and the optimistic mode

DM uses the pessimistic mode by default. In scenarios of migrating and merging MySQL shards, changes in upstream shard schemas can block DML writing to downstream databases. You need to wait until all the schemas are changed and have the same structure, and then continue the migration from the breakpoint.

• If the upstream schema changes take a long time, it might cause the upstream Binlog to be cleaned up. You can enable the relay log to avoid this problem. For more information, see Use the relay log.

• If you do not want to block data write due to upstream schema changes, consider using the optimistic mode. In this case, DM will not block the data migration even when it spots changes in the upstream shard schemas, but will continue to migrate the data. However, if DM spots incompatible formats in upstream and downstream, the migration task will stop. You need to resolve this issue manually.

The following table summarizes the pros and cons of optimistic mode and pessimistic modes.

Other restrictions and impact

Data types in upstream and downstream

TiDB supports most MySQL data types. However, some special types are not supported yet (such as SPATIAL). For the compatibility of data types, see Data Types.

Character sets and collations

Since TiDB v6.0.0, the new framework for collations is used by default. In earlier versions, if you want TiDB to support utf8_general_ci, utf8mb4_general_ci, utf8_unicode_ci, utf8mb4_unicode_ci, gbk_chinese_ci and gbk_bin, you need to explicitly declare it when creating the cluster by setting the value of new_collations_enabled_on_first_bootstrap to true. For more information, see New framework for collations.

The default character set in TiDB is utf8mb4. It is recommended that you use utf8mb4 for the upstream and downstream databases and applications. If the upstream database has explicitly specified a character set or collation, you need to check whether TiDB supports it.

Since TiDB v6.0.0, GBK is supported. For more information, see the following documents:

• Character Set and Collation
• GBK compatibility

Best practices for deployment

Deploy DM-master and DM-worker

DM consists of DM-master and DM-worker nodes.

• DM-master manages the metadata of migration tasks and schedules DM-worker nodes. It is the core of the whole DM platform. Therefore, you can deploy DM-master as clusters to ensure high availability of the DM platform.

• DM-worker executes upstream and downstream migration tasks. A DM-worker node is stateless. You can deploy at most 1000 DM-worker nodes. When using DM, it is recommended that you reserve some idle DM-workers to ensure high availability.

Plan the migration tasks

When migrating and merging MySQL shards, you can split a migration task according to the types of shards in the upstream. For example, if usertable_1~50 and Logtable_1~50 are two types of shards, you can create two migration tasks. It can simplify the migration task template and effectively control the impact of interruption in data migration.

For migration of large datasets, you can refer to the following suggestions to split the migration task:

• If you need to migrate multiple databases in the upstream, you can split the migration task according to the number of databases.

• Split the task according to the write pressure in the upstream, that is, split the tables with frequent DML operations in the upstream to a separate migration task. Use another migration task to migrate the tables without frequent DML operations. This method can speed up the migration progress, especially when there are a large number of logs written to a table in the upstream. But if this table that contains a large number of logs does not affect the whole business, this method still works well.

Note that splitting the migration task can only guarantee the final consistency of data. Real-time consistency may deviate significantly due to various reasons.

The following table describes the recommended deployment plans for DM-master and DM-worker in different scenarios.

Choose and configure the upstream data source

DM backs up the full data of the entire database when performing full data migration, and uses the parallel logical backup method. During backing up MySQL, it adds a global read lock FLUSH TABLES WITH READ LOCK. DML and DDL operations of the upstream database will be blocked for a short time. Therefore, it is strongly recommended to use a backup database in upstream to perform the full data backup, and enable the GTID function of the data source (enable-gtid: true). In this way, you can avoid the impact from the upstream, and switch to the master node in the upstream to reduce the latency during the incremental migration. For the instructions of switching the upstream MySQL data source, see Switch DM-worker Connection between Upstream MySQL Instances.

Note the following:

• You can only perform full data backup on the master node of the upstream database.

  In this scenario, you can set the consistency parameter to none in the configuration file, mydumpers.global.extra-args: "--consistency none", to avoid adding a global read lock to the master node. But this might affect the data consistency of the full backup, which may lead to inconsistent data between the upstream and downstream.

• Use backup snapshots to perform full data migration (only applicable to the migration of MySQL RDS and Aurora RDS on AWS)

  If the database to be migrated is AWS MySQL RDS or Aurora RDS, you can use RDS snapshots to directly migrate the backup data in Amazon S3 to TiDB to ensure data consistency. For more information, see Migrate Data from Amazon Aurora to TiDB.

Details of configurations

Capitalization

TiDB schema names are case-insensitive by default, that is, lower_case_table_names:2. But most upstream MySQL databases use Linux systems that are case-sensitive by default. In this case, you need to set case-sensitive to true in the DM task configuration file to ensure that the schema can be correctly migrated from the upstream.

In a special case, for example, if there is a database in the upstream that has both uppercase tables such as Table and lowercase tables such as table, then an error occurs when creating the schema:

ERROR 1050 (42S01): Table '{tablename}' already exists

Filter rules

You can configure the filter rules as soon as you start configuring the data source. For more information, see Data Migration Task Configuration Guide. The benefits of configuring the filter rules are:

• Reduce the number of Binlog events that the downstream needs to process, thereby improving migration efficiency.
• Reduce unnecessary relay log storage, thereby saving disk space.

Use the relay log

In the MySQL master/standby mechanism, the standby node saves a copy of relay logs to ensure the reliability and efficiency of asynchronous replication. DM also supports saving a copy of relay logs on DM-worker. You can configure information such as the storage location and expiration time. This feature applies to the following scenarios:

• During full and incremental data migration, if the amount of full data is large, the entire process takes more time than the time for the upstream binlogs to be archived. It causes the incremental replication task to fail to start normally. If you enable the relay log, DM-worker will start receiving relay logs when the full migration is started. This avoids the failure of the incremental task.

• When you use DM to perform long-time data replication, sometimes the migration task is blocked for a long time due to various reasons. If you enable the relay log, you can effectively deal with the problem of upstream binlogs being recycled due to the blocking of the migration task.

There are some restrictions on using the relay log. DM supports high availability. When a DM-worker fails, it will try to promote an idle DM-worker instance to a working instance. If the upstream binlogs do not contain the necessary migration logs, it may cause interruption. You need to intervene manually to copy the relay log to the new DM-worker node as soon as possible, and modify the corresponding relay meta file. For details, see Troubleshooting.

Use PT-osc/GH-ost in upstream

In daily MySQL operation and maintenance, usually you use tools such as PT-osc/GH-ost to change the schema online to minimize impact on the business. However, the whole process will be logged to MySQL Binlog. Migrating such data to TiDB downstream will result in a lot of unnecessary write operations, which is neither efficient nor economical.

To resolve this issue, DM supports third-party data tools such as PT-osc and GH-ost when you configure the migration task. When you use such tools, DM does not migrate redundant data and ensure data consistency. For details, see Migrate from Databases that Use GH-ost/PT-osc.

Best practices during migration

This section introduces how to troubleshoot problems you might encounter during migration.

Inconsistent schemas in upstream and downstream

Common errors include:

• messages: Column count doesn't match value count: 3 (columns) vs 2 (values)
• Schema/Column doesn't match

Usually such issues are caused by changed or added indexes in the downstream TiDB, or there are more columns in the downstream. When such errors occur, check whether the upstream and downstream schemas are inconsistent.

To resolve such issues, update the schema information cached in DM to be consistent with the downstream TiDB schema. For details, see Manage Table Schemas of Tables to be Migrated.

If the downstream has more columns, see Migrate Data to a Downstream TiDB Table with More Columns.

Interrupted migration task due to failed DDL

DM supports skipping or replacing DDL statements that cause a migration task to interrupt. For details, see Handle Failed DDL Statements.

Data validation after data migration

It is recommended that you validate the consistency of data after data migration. TiDB provides sync-diff-inspector to help you complete the data validation.

Now sync-diff-inspector can automatically manage the table list to be checked for data consistency through DM tasks. Compared with the previous manual configuration, it is more efficient. For details, see Data Check in the DM Replication Scenario.

Since DM v6.2.0, DM supports continuous data validation for incremental replication. For details, see Continuous Data Validation in DM.

Long-term data replication

If you use DM to perform a long-term data replication task, it is necessary to back up the metadata. On the one hand, it ensures the ability to rebuild the migration cluster. On the other hand, it can implement the version control of the migration task. For details, see Export and Import Data Sources and Task Configuration of Clusters.