Ali cloud AnalyticDB for PostgreSQL (referred to as ADB PG), based on a kernel build distributed database PostgreSQL, Oracle compatible syntax section, which can scale horizontally, that is, support for distributed transaction processing high throughput, and also supports high-performance analysis of complex queries, It is the industry's most competitive in HTAP distributed database cluster. This article describes the key architectural ADB PG, and performance indicators under HTAP scene.
ADB PG Architecture
ADB PG PostgreSQL based core, into Segment Master data node and an access node. Master is responsible for parsing h and SQL optimization, distributed generation and execution plan, sent to the segment node execution.
- The coordinating node Master: divided into Main Master and Secondary Master, it accepts client requests and parsing and optimization of SQL. Meanwhile Main Master built GTM global transaction management module, maintains global unified Global XID and Snapshot, enabling strict isolation levels SI and distributed RC. Master node architecture optimizer using Cascade, the combined unified RBO and CBO, while automation solutions correlated subquery to ensure optimal execution plan to generate most queries.
- Computing node Segment: Segment supports multiple copies of data storage, support table stored in rows or columns, when a scene analysis for a pure line store providing aggregate performance analysis, etc. when TP transaction for the scene, and a variety of two columns of memory level index, comprising BTree, to provide a high performance transaction processing throughput. PG while the ADB in the compute engine on the basis of both PostgreSQL, developed a full research from the high-performance computing engine to quantify the Odyssey , the TPC-H performance analysis engine faster than the native PG 3 times
OLTP Distributed Transaction scene analysis and comparison TiDB
ADB PG ensuring strong uniform distributed, provided under the standard RC isolation levels, and horizontal expansion by Segmen Master can provide a good distributed transaction throughput. ADB PG supports standard TPC-C benchmark support. Contrast ADBPG and TPC-C Benchmark Performance analysis TiDB follows, both of test metrics can be found in the official documentation:
- ADBPG TPC-C 6.0 Evaluation: in total 64 Core, the 512GB memory 101231.3 tpmC, details can be found https://help.aliyun.com/document_detail/155810.html
- The TPC-C TiDB 3.0 Reviews: In total 120 Core, the 384GB memory 44068 tpmC, details can be found https://pingcap.com/docs-cn/stable/benchmark/tpcc/
The above-described configuration, since the whole memory are calculated, the difference memory negligible impact on performance, while the more CPU resources provided TiDB, ADB PG is twice or more of TiDB.
| ADB PG | TiDB |
|---|---|
| 101231.3 tpmC | 44068 tpmC |
OLAP comparative analysis of complex analysis scenarios and Greenplum
ADB PG research since the vectorization calculation engine, while building have the advantage of the column storage, the ICT Academy analytical database a large-scale performance tests in November 2019, the build ICT Academy of the test maximum cluster size 640 MPP node, verify 100TB analyze performance through support for TPC-DS.
ADB PG from the study to quantify calculation engine Odyssey , compared with the open-source computing Greenplum MPP database native PG engine performance fast approaching twice. The following is a comparison 1TB TPC-H performance under the same specifications.
Overview
ADB PG using horizontal architecture, with complete SQL database support functions, while GTM-based distributed transaction management framework to achieve a strong consistency in distributed, supports SI / RC isolation level. OLTP Benchmark TPC-C is TiDB twice, OLAP benchmark for the open-source database MPP twice Greenplum, on a framework to achieve a high-performance distributed HTAP ability.