TiCDC Overview

TiCDC is a tool used to replicate incremental data from TiDB. Specifically, TiCDC pulls TiKV change logs, sorts captured data, and exports row-based incremental data to downstream databases.

Usage scenarios

TiCDC has multiple usage scenarios, including:

  • Providing high availability and disaster recovery solutions for multiple TiDB clusters. TiCDC ensures eventual data consistency between primary and secondary clusters in case of a disaster.
  • Replicating real-time data changes to homogeneous systems. This provides data sources for various scenarios, such as monitoring, caching, global indexing, data analysis, and primary-secondary replication between heterogeneous databases.

Major features

Key capabilities

TiCDC has the following key capabilities:

  • Replicating incremental data between TiDB clusters with second-level RPO and minute-level RTO.
  • Bidirectional replication between TiDB clusters, allowing the creation of a multi-active TiDB solution using TiCDC.
  • Replicating incremental data from a TiDB cluster to a MySQL database or other MySQL-compatible databases with low latency.
  • Replicating incremental data from a TiDB cluster to a Kafka cluster. The recommended data format includes Canal-JSON, Avro, and Debezium.
  • Replicating incremental data from a TiDB cluster to storage services, such as Amazon S3, GCS, Azure Blob Storage, and NFS.
  • Replicating tables with the ability to filter databases, tables, DMLs, and DDLs.
  • High availability with no single point of failure, supporting dynamically adding and deleting TiCDC nodes.
  • Cluster management through Open API, including querying task status, dynamically modifying task configuration, and creating or deleting tasks.

Replication order

  • For all DDL or DML statements, TiCDC outputs them at least once.

  • When the TiKV or TiCDC cluster encounters a failure, TiCDC might send the same DDL/DML statement repeatedly. For duplicated DDL/DML statements:

    • The MySQL sink can execute DDL statements repeatedly. For DDL statements that can be executed repeatedly in the downstream, such as TRUNCATE TABLE, the statement is executed successfully. For those that cannot be executed repeatedly, such as CREATE TABLE, the execution fails, and TiCDC ignores the error and continues with the replication process.
    • The Kafka sink provides different strategies for data distribution.
      • You can distribute data to different Kafka partitions based on the table, primary key, or timestamp. This ensures that the updated data of a row is sent to the same partition in order.
      • All these distribution strategies send Resolved TS messages to all topics and partitions periodically. This indicates that all messages earlier than the Resolved TS have already been sent to the topics and partitions. The Kafka consumer can use the Resolved TS to sort the messages received.
      • The Kafka sink sometimes sends duplicated messages, but these duplicated messages do not affect the constraints of Resolved Ts. For example, if a changefeed is paused and then resumed, the Kafka sink might send msg1, msg2, msg3, msg2, and msg3 in order. You can filter out the duplicated messages from Kafka consumers.

Replication consistency

  • MySQL sink

    • TiCDC enables the redo log to ensure eventual consistency of data replication.

    • TiCDC ensures that the order of single-row updates is consistent with the upstream.

    • TiCDC does not ensure that the downstream transactions are executed in the same order as the upstream transactions.

      Overview - 图1

      Note

      Since v6.2, you can use the sink URI parameter transaction-atomicity to control whether to split single-table transactions. Splitting single-table transactions can greatly reduce the latency and memory consumption of replicating large transactions.

TiCDC architecture

TiCDC is an incremental data replication tool for TiDB, which is highly available through PD’s etcd. The replication process consists of the following steps:

  1. Multiple TiCDC processes pull data changes from TiKV nodes.
  2. TiCDC sorts and merges the data changes.
  3. TiCDC replicates the data changes to multiple downstream systems through multiple replication tasks (changefeeds).

The architecture of TiCDC is illustrated in the following figure:

TiCDC architecture

The components in the architecture diagram are described as follows:

  • TiKV Server: TiKV nodes in a TiDB cluster. When data changes occur, TiKV nodes send the changes as change logs (KV change logs) to TiCDC nodes. If TiCDC nodes detect that the change logs are not continuous, they will actively request the TiKV nodes to provide change logs.
  • TiCDC: TiCDC nodes where TiCDC processes run. Each node runs a TiCDC process. Each process pulls data changes from one or more tables in TiKV nodes and replicates the changes to the downstream system through the sink component.
  • PD: The scheduling module in a TiDB cluster. This module is responsible for scheduling cluster data and usually consists of three PD nodes. PD provides high availability through the etcd cluster. In the etcd cluster, TiCDC stores its metadata, such as node status information and changefeed configurations.

As shown in the architecture diagram, TiCDC supports replicating data to TiDB, MySQL, Kafka, and storage services.

Best practices

  • When you use TiCDC to replicate data between two TiDB clusters, if the network latency between the two clusters is higher than 100 ms:

    • For TiCDC versions earlier than v6.5.2, it is recommended to deploy TiCDC in the region (IDC) where the downstream TiDB cluster is located.
    • With a series of improvements introduced starting from TiCDC v6.5.2, it is recommended to deploy TiCDC in the region (IDC) where the upstream TiDB cluster is located.

  • TiCDC only replicates tables that have at least one valid index. A valid index is defined as follows:

    • A primary key (PRIMARY KEY) is a valid index.
    • A unique index (UNIQUE INDEX) is valid if every column of the index is explicitly defined as non-nullable (NOT NULL) and the index does not have a virtual generated column (VIRTUAL GENERATED COLUMNS).
  • To ensure eventual consistency when using TiCDC for disaster recovery, you need to configure redo log and ensure that the storage system where the redo log is written can be read normally when a disaster occurs in the upstream.

Implementation of processing data changes

This section mainly describes how TiCDC processes data changes generated by upstream DML operations.

For data changes generated by upstream DDL operations, TiCDC obtains the complete DDL SQL statement, converts it into the corresponding format based on the sink type of the downstream, and sends it to the downstream. This section does not elaborate on this.

Overview - 图3

Note

The logic of how TiCDC processes data changes might be adjusted in subsequent versions.

MySQL binlog directly records all DML SQL statements executed in the upstream. Unlike MySQL, TiCDC listens to the real-time information of each Region Raft Log in the upstream TiKV, and generates data change information based on the difference between the data before and after each transaction, which corresponds to multiple SQL statements. TiCDC only guarantees that the output change events are equivalent to the changes in the upstream TiDB, and does not guarantee that it can accurately restore the SQL statements that caused the data changes in the upstream TiDB.

Data change information includes the data change types and the data values before and after the change. The difference between the data before and after the transaction can result in three types of events:

  1. The DELETE event: corresponds to a DELETE type data change message, which contains the data before the change.

  2. The INSERT event: corresponds to a PUT type data change message, which contains the data after the change.

  3. The UPDATE event: corresponds to a PUT type data change message, which contains the data both before and after the change.

Based on the data change information, TiCDC generates data in the appropriate formats for different downstream types, and sends the data to the downstream. For example, it generates data in formats such as Canal-JSON and Avro, and writes the data to Kafka, or converts the data back into SQL statements and sends them to the downstream MySQL or TiDB.

Currently, when TiCDC adapts data change information for the corresponding protocol, for specific UPDATE events, it might split them into one DELETE event and one INSERT event. For more information, see Split update events into delete and insert events.

When the downstream is MySQL or TiDB, TiCDC cannot guarantee that the SQL statements written to the downstream are exactly the same as the SQL statements executed in the upstream. This is because TiCDC does not directly obtain the original DML statements executed in the upstream, but generates SQL statements based on the data change information. However, TiCDC ensures the consistency of the final results.

For example, the following SQL statement is executed in the upstream:

  1. Create Table t1 (A int Primary Key, B int);
  3. BEGIN;
  4. Insert Into t1 values(1,2);
  5. Insert Into t1 values(2,2);
  6. Insert Into t1 values(3,3);
  7. Commit;
  9. Update t1 set b = 4 where b = 2;

TiCDC generates the following two SQL statements based on the data change information, and writes them to the downstream:

  1. INSERT INTO `test.t1` (`A`,`B`) VALUES (1,2),(2,2),(3,3);
  2. UPDATE `test`.`t1`
  3. SET `A` = CASE
  4.         WHEN `A` = 1 THEN 1
  5.         WHEN `A` = 2 THEN 2
  6. END, `B` = CASE
  7.         WHEN `A` = 1 THEN 4
  8.         WHEN `A` = 2 THEN 4
  9. END
  10. WHERE `A` = 1 OR `A` = 2;

Unsupported scenarios

Currently, the following scenarios are not supported:

TiCDC only partially supports scenarios involving large transactions in the upstream. For details, refer to the TiCDC FAQ, where you can find details on whether TiCDC supports replicating large transactions and any associated risks.