SIGMOD 2023-PolarDB-IMCI- A Cloud-Native HTAP Database System at Alibaba.pdf

波风水门

429

14页

6次

2023-05-31

免费下载

PolarDB-IMCI: A Cloud-Native HTAP Database System at Alibaba

Jianying Wang

†

, Tongliang Li

†

, Haoze Song

†

, Xinjun Yang

†

, Wenchao Zhou

†

, Feifei Li

†

Baoyue Yan

†

, Qianqian Wu

†

, Yukun Liang

†

, Chengjun Ying

†§

, Yujie Wang

†

, Baokai Chen

†

Chang Cai

†

, Yubin Ruan

†

, Xiaoyi Weng

†

, Shibin Chen

†

, Liang Yin

†

, Chengzhong Yang

†

, Xin Cai

†

Hongyan Xing

†

, Nanlong Yu

†

, Xiaofei Chen

†

, Dapeng Huang

†

, Jianling Sun

†§

†

Alibaba Group

Zhejiang University

{beilou.wjy, litongliang.ltl, songhaoze.shz, xinjun.y, zwc231487, lifeifei, baoyue.yby, daisy.wqq, liangyukun.lyk

zhencheng.wyj, baokai.cbk, caichang.cc, yubin.ryb, echo.wxy, wuha.csb, allen.yinl, chengzhong.ycz

frank.cx, diane.xhy, nanlong.ynl, chenxiaofei.cxf, wuzang.hdp}@alibaba-inc.com

{yingcj,sunjl}@zju.edu.cn

ABSTRACT

Cloud-native databases have become the de-facto choice for mission-

critical applications on the cloud due to the need for high availabil-

ity, resource elasticity, and cost eciency. Meanwhile, driven by

the increasing connectivity between data generation and analysis,

users prefer a single database to eciently process both OLTP and

OLAP workloads, which enhances data freshness and reduces the

complexity of data synchronization and the overall business cost.

In this paper, we summarize ve crucial design goals for a cloud-

native HTAP database based on our experience and customers’

feedback, i.e., transparency, competitive OLAP performance, min-

imal perturbation on OLTP workloads, high data freshness, and

excellent resource elasticity. As our solution to realize these goals,

we present PolarDB-IMCI, a cloud-native HTAP database system

designed and deployed at Alibaba Cloud. Our evaluation results

show that PolarDB-IMCI is able to handle HTAP eciently on

both experimental and production workloads; notably, it speeds up

analytical queries up to

149 on TPC-H (100

𝐺𝐵

). PolarDB-IMCI

introduces low visibility delay and little performance perturbation

on OLTP workloads (< 5%), and resource elasticity can be achieved

by scaling out in tens of seconds.

CCS CONCEPTS

• Information systems → DBMS engine architectures

; Data-

base transaction processing; Online analytical processing engines.

KEYWORDS

cloud databases, hybrid transactional and analytical processing

ACM Reference Format:

Jianying Wang, Tongliang Li, Haoze Song, Xinjun Yang, Wenchao Zhou,

Feifei Li, Baoyue Yan, Qianqian Wu, Yukun Liang, Chengjun Ying, Yujie

Wang, Baokai Chen, Chang Cai, Yubin Ruan, Xiaoyi Weng, Shibin Chen,

Liang Yin, Chengzhong Yang, Xin Cai, Hongyan Xing, Nanlong Yu, Xiaofei

Chen, Dapeng Huang, Jianling Sun. 2023. PolarDB-IMCI: A Cloud-Native

HTAP Database System at Alibaba. In Procee dings of International Conference

This work is licensed under a Creative Commons Attribution Inter-

national 4.0 License.

SIGMOD’23, June 18–23, 2023, Seattle, WA, USA

ACM ISBN 978-x-xxxx-xxxx-x/YY/MM.

https://doi.org/10.1145/3589785

on Management of Data (SIGMOD’23). ACM, New York, NY, USA, 14 pages.

https://doi.org/10.1145/3589785

1 INTRODUCTION

In recent years, cloud-native databases [

] have become

an inexorable trend in the database industry. Dierent from on-

premise databases, a cloud-native database decouples its architec-

ture into two layers: a computation layer and a storage layer, allow-

ing resources to scale independently. Nodes equipped with disks

(in the storage layer) form a shared storage pool that serves as a

unied data interface for nodes in the computation layer. This dis-

aggregation architecture enables database systems to oer extreme

elasticity, exible on-demand charging models, and low operat-

ing costs for customers. As a result, the market of cloud-native

databases has quickly taken o [36].

Meanwhile, we have witnessed another trend that the line be-

tween classic OLTP and OLAP databases started to blur: there is a

growing need for a database to provide sucient support for both

transactional processing and analytical processing, especially in

the elds of business intelligence [

], social media [

], fraud

detection [

], and marketing [

]. To provide such capability,

traditional solutions often deploy data and application logic into

two databases, one specialized in OLTP and the other in OLAP (e.g.,

MySQL [

] for OLTP, and ClickHouse [

] for OLAP), and rely

on data synchronization techniques (such as Extract-Transform-

Load [

] (ETL) workow) for ensuring consistencies between them,

as shown in Figure 1. According to our statistics, nearly 30% of the

customers of PolarDB, an OLTP database, synchronize data to an

independent data warehouse system for data analytics needs.

Such solutions are costly, as it negatively impacts the OLTP per-

formance, and introduces a time-consuming data synchronization

process, which further leads to delays or even inconsistencies be-

tween the data maintained at the TP/AP databases. In practice, these

issues lead to sub-optimal user experience and a large number of

user inquiries. To address these issues, it calls for a cloud-native Hy-

brid Transactional and Analytical Processing (HTAP) database. In

this paper, we present PolarDB-IMCI, a cloud-native HTAP database

deployed at Alibaba Cloud. We summarize the crucial design goals

of PolarDB-IMCI below, which are also applicable to the design of

a general cloud-native HTAP database.

• G#1: Transparent Query Execution.

To serve mixed work-

loads in a single database, database users should not be required

arXiv:2305.08468v1 [cs.DB] 15 May 2023

SIGMOD’23, June 18–23, 2023, Seale, WA, USA Jianying Wang et al.

OLTP DBMS

(e.g., MySQL)

OLAP DBMS

(e.g., ClickHouse)

ETL

REDO

HTAP DBMS

(PolarDB-IMCI)

T Q

Figure 1: Comparison of ETL and PolarDB-IMCI.

to understand the working logic of the database, nor should they

identify query types manually. That is, users should not perceive

two isolated systems (e.g., engines, indexes, interfaces, etc.) for

OLAP and OLTP queries respectively. Our system should provide

a unied SQL interface for both OLAP and OLTP workloads.

• G#2: Advanced OLAP Performance.

As a major goal of any

HTAP database, the OLAP performance (e.g., execution latency)

of PolarDB-IMCI should be comparable to typical databases spe-

cialized in processing OLAP queries (typically through the intro-

duction of columnar data storage).

• G#3: Minimal Perturbation on OLTP Workloads.

While the

performance of OLAP queries is signicantly improved, it should

have a minimal negative impact on the performance of OLTP

queries. In fact, as we have practically validated in real application

scenarios, OLTP queries are usually more mission-critical and

are more sensitive to performance degradation. This requires

eective resource isolation for OLTP and OLAP queries.

• G#4: High Data Freshness.

High data freshness is an important

property of HTAP databases, which is a distinguishing advan-

tage compared to the traditional Extract-Transform-Load (ETL)

method. In this paper, we follow earlier similar work [

]

using the visibility delay as a freshness score for a query. By

denition, the visibility delay is the time interval during which

updates to the database can be visible to OLAP queries.

• G#5: Excellent Resource Elasticity.

In HTAP scenarios, the

consumption of CPU/IO resources uctuates signicantly, from

hundreds to thousands of times. As a key feature of cloud-native

databases, our system should ensure high resource elasticity

(e.g., scale-out in minutes or even seconds) to adaptively serve

the changing data volume and analytical workloads with stable

performance and high resource utilization.

PolarDB-IMCI meets all desired goals (i.e., G#1-5) with the fol-

lowing innovations. First, to meet G#1 and G#2, we implemented

in-memory column index (IMCI, §4) as complementary storage.

PolarDB-IMCI absorbs diverse advanced optimizations from the

OLAP community and derives a new SQL engine (§6.3) to match

the execution mode on columns. Further, PolarDB-IMCI proposes

a new query routing mechanism (§6.1) that dispatches queries trans-

parently.

Second, to meet G#3, PolarDB-IMCI resides column indexes on

separated read-only (RO) nodes (§3.1) with a shared storage archi-

tecture to provide eective resource isolation between OLTP and

OLAP requests. Updates are propagated to RO nodes by reusing

REDO logs (§5.3) (i.e., the dierential logging for the row store)

instead of shipping additional logical logs (i.e., MySQL Binlogs).

Third, to meet G#4, we enhance our update propagation frame-

work with commit-ahead log shipping (CALS, §5.1) and 2-Phase

conict-free log replay (2P-COFFER, §5.2). CALS ships transaction

logs before committing. 2P-COFFER eciently parses and applies

REDO logs to RO nodes. Furthermore, we implemented the column

index as append-only storage (§4): records are organized in insert or-

der rather than primary key order. Thus, updates to column indexes

are performed out-place and quickly.

Finally, to meet G#5, the checkpoint mechanism of the colum-

nar index is seamlessly built into PolarDB’s original storage en-

gine. Therefore fast scale-out capability can be achieved by quickly

pulling up a RO node using the checkpoint on shared storage (§7).

We started the design and development of cloud-native PolarDB

in 2017, and seek for an HTAP solution (i.e., PolarDB-IMCI) in 2019.

By now, PolarDB-IMCI is severing a large number of internal and

external customers (Table 3). The key contributions of this work

are listed as follows:

•

We propose PolarDB-IMCI, an HTAP solution for cloud-

native relational database systems. To the best of our knowl-

edge, PolarDB-IMCI is the rst cloud-native HTAP database

to satisfy all of the aforementioned design goals.

•

We design an architecture that provides dual-format stor-

age on read-only nodes under the storage-computation sep-

aration architecture, which enables ecient execution of

analytical queries and minimizes the impact on OLTP load.

Additionally, PolarDB-IMCI is the rst practical template

to demonstrate that it is possible and applicable to imple-

ment replication from row-store to dual-format storage with

physical redo logs while reducing replication latency to mil-

lisecond levels.

•

We evaluate PolarDB-IMCI with diverse experiments (in

both experimental and production environments). The ex-

perimental results show that PolarDB-IMCI outperforms

row-based PolarDB up to

149 on a standard analytical

workload TPC-H (100

𝐺𝐵

), and its performance is compa-

rable to the advanced OLAP databases (e.g., ClickHouse).

Performance degradation on OLTP is tiny (less than 5%),

even when OLAP workloads increase continuously. The visi-

bility delay of PolarDB-IMCI at is <5

𝑚𝑠

on typical workloads,

and <30

𝑚𝑠

under heavy workloads. PolarDB-IMCI can scale

out in tens of seconds.

The remainder of the paper is organized as follows. §2 introduces

the background of HTAP and cloud-native databases. §3 presents

the architecture. PolarDB-IMCI’s components and update prop-

agation framework are introduced in §4 and §5 respectively. §6

discusses query dispatch, optimization, and execution. §7 intro-

duces the checkpoint mechanism. §8 details the experiments and

evaluation. §9 concludes the paper.

2 BACKGROUND AND RELATED WORK

2.1 Hybrid Transactional/Analytical

Processing

For long decades, OLTP and OLAP databases are dedicatedly de-

signed for their respective workloads. For instance, OLTP engines

(e.g., MySQL [

]) prefer row-based data formats, row-at-a-time

operators, and early materialization strategy, favoring data mod-

ication and point queries. On the contrary, OLAP engines (e.g.,

of 14

免费下载

关注

评论