Indexes

B+Tree

TDSQL uses the index supported by InnoDB. InnoDB uses B+ tree as the default index data structure.

Checkpoints

Blocking

Memory data need to be flushed to disk to make sure all data are saved. Read locks are need to lock on all tables before making the checkpoint.

Concurrency Control

Two-Phase Locking (Deadlock Detection)

Global wait-for-graphs are created to detect deadlocks. The timer is applied to count how long a transaction has been started. If a transaction is not finished within a given time period, the timer will indicate a potential deadlock.

Storage Architecture

Disk-oriented

3-level storage strategy is used to save storage resources and reduce query latency. According to the 3-level storage strategy, all data can be classified into Standard Storage (hot data), Infrequent Access Storage (warm data), and Archive (cold data). Automatic conversion is trigger between different storage levels under different lifecycles. Standard Storage is stored in three copies on SSD and Infrequent access Storage is stored in 2 copies on SATA. Archive is stored in 1.3 copies on SATA.

Storage Model

N-ary Storage Model (Row/Record)

Data Model

Relational Key/Value Document / XML

TDSQL supports multiple data models to meet different application requirements. MySQL API, Redis API, and MongoDB API are used with SQL execution engine, KV execution engine and Doc execution engine respectively. There is a data conversion layer between the execution layer and the storage layer to unify the data format.

Foreign Keys

Not Supported

Query Interface

HTTP / REST

TDSQL supports SDK for python, java, php, nodejs, .net. Tecent Cloud CLI and HTTPS/REST API can also be used to interact with the DBMS

Joins

Not Supported

It only supports joining on the same node, which requires join tables to be partitioned on the same key or the table is small enough to be copied across all partitions. Then the proxy/coordinator merges the results from different shards.

System Architecture

Shared-Nothing

The TDSQL framework can be divided into the following layers. Computer layer, which consists of the compute engine, is stateless. For the storage layer, data are stored in the unit of the replica set with multiple duplicates. Strong consistency is guaranteed by the Raft consistency model.