How PolarDB MySQL Limitless Redefines Cloud‑Native Database Performance

In recent years, cloud‑native architecture has driven elastic scaling and low‑cost operations, breaking single‑node performance limits and supporting massive data and high‑concurrency workloads. PolarDB MySQL 8.0.2’s Limitless multi‑master cluster achieved a world‑record TPC‑C performance of 20.55 billion tpmC with a cost of 0.8 CNY per tpmC, showcasing extreme performance and cost‑effectiveness.

1. Overview

PolarDB MySQL Limitless supports multi‑master clusters (Limitless), allowing customers to add read‑write (RW) nodes to transform a single‑master‑multiple‑read setup into a multi‑master configuration.

2. Technical Deep Dive

(1) Overall Architecture

PolarDB MySQL clusters consist of multiple RW compute nodes, multiple RO compute nodes, and optional Cache Coordinator nodes, all built on distributed shared storage. Each RW node can read and write, communicating with Cache Coordinators via high‑speed RDMA/CXL networks. Cache Coordinators manage cluster metadata and transaction coordination, and also provide CDC functionality with a unified global Binlog.

The architecture distributes different partitions of a single table across RW nodes, enabling concurrent writes and horizontal write scaling. It also supports cross‑node distributed queries, DML, and DDL, with global RO nodes accelerating global and multi‑table queries.

Improves resource utilization by eliminating idle CN/DN components.

Reduces communication overhead between CN and DN, boosting performance.

Maintains 100% MySQL compatibility by using MySQL’s native parser, optimizer, and executor.

(2) High‑Performance Horizontal Scaling

PolarTrans, the cloud‑native transaction system, replaces InnoDB’s active‑transaction list with a lightweight CTS (Commit Timestamp) log implemented as a ring buffer. Transactions map to slots via modulo, storing pointers and CTS values, eliminating global locks and enabling lock‑free operation.

(3) Distributed Snapshot Consistency

All nodes interconnect via RDMA, sharing the CTS Log for global transaction state synchronization, avoiding a central active‑transaction list. Transactions use CTS timestamps for visibility checks, ensuring consistent snapshots across nodes.

(4) Multi‑Node Transaction Consistency

A hardware‑assisted distributed transaction scheme leverages RDMA and LSN mechanisms to accelerate the Prepare and Commit phases, reducing coordination overhead and enabling asynchronous commit.

(5) Multi‑Node Parallel DDL

Global DDL uses an XB lock to prevent concurrent DDL on the same table and a multi‑phase commit protocol to guarantee atomicity, rolling back on failures.

(6) High Availability

Two HA options are provided: private RO nodes per RW node for instant failover, or RW‑to‑RW failover where another RW node takes over the workload, ensuring minimal performance impact.

(7) Global RO Nodes

Global read‑only nodes aggregate data from all RW nodes, enabling cross‑table and cross‑database queries without extra data replication, routed automatically via PolarProxy.

(8) Extreme Performance and Scalability

Using the TPC‑C benchmark on a 2,340‑node Limitless cluster (48‑core, 512 GB per node), PolarDB achieved 20.55 billion tpmC with a fluctuation of only 0.16% over 8 hours, setting a new world record. The system also demonstrated sub‑10‑second failover and sub‑2% performance impact during disaster recovery tests, and achieved second‑level horizontal scaling when adding RW nodes.

PolarDB is the first cloud‑native relational database deeply integrating RDMA/CXL, supporting massive scale, high‑performance cross‑node transactions, and near‑linear performance scaling, solidifying its position as a global leader in database technology.