POLARDB architecture
summary
POLARDB uses a shared storage layer called Polar Store with chunk servers, DMA networking, NVDIMM-like low-latency media and NVMe SSDs; a user-space libpolofs library talks direct to chunks to avoid syscalls and copies. One master writes redo to shared storage while replicas physical apply redo in parallel, enabling fast replica provisioning, non-blocking DDL, and lower IO (no binary-log copies). Designed for 100 million transactions/day and 11/11 spikes ~100x, it scales reads well but writes remain single-master limited until multi-master work completes.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:48Databases inside Alibaba Group
- 3:47Data Explosion
- 4:12Cloud Native Database - Requirements
- 5:10POLARDB — Cloud Native Database
- 5:50Storage Revolution: PolarStore
- 7:06PolarStore: Design for Emerging Hardware
- 9:03PolarFS: posix distributed file system closely with DB
- 9:35Database Architecture Revolution: Separation of Storage and Computation
- 11:42Dynamic Scaling
- 16:37Physical Replication by Redo Log
- 22:09Physical Replication - Page from Past
- 24:52Physical Replication - Page from Future
- 25:50Single Master
- 27:02Read and Write Separation - Session Consistent
- 29:34HTAP - Hybrid Transaction and Analytical Processing
related talks
21mPolarDB: Alibaba Cloud-Native Relational Database Service
16mPolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases
01h 12mFrom Storage Formats to Open Governance: The Evolution to Apache Polaris
42mPostgreSQL Distributed: Architectures & Best practices
37mPolars 1.0 and beyond
44mPolars; DataFrames in the multi-core era
13mBuilding reliable infrastructure with Postgres
20mThe Polars vs SQL differences nobody is talking about