VLDB2024_Native Distributed Databases：Problems, Challenges and Opportunities_OceanBase.pdf

迹部景吾

780

4页

20次

2024-09-02

免费下载

Native Distributed Databases: Problems, Challenges and

Opportunities

Quanqing Xu

OceanBase, Ant Group

xuquanqing.xqq@oceanbase.com

Chuanhui Yang

∗

OceanBase, Ant Group

rizhao.ych@oceanbase.com

Aoying Zhou

East China Normal University

ayzhou@dase.ecnu.edu.cn

ABSTRACT

Native distributed databases, crucial for scalable applications, of-

fer transactional and analytical prowess but face data intricacies

and network challenges. Under the CAP theorem’s constraints,

latency and replication issues necessitate creative approaches to

maintenance, security, and upgrades. Progress in consistency al-

gorithms, network technology, automation, and machine learning

for optimization presents signicant potential. Embracing hybrid

transactional/analytical processing (HTAP), these databases repre-

sent an evolutionary leap in data management, aiming to reconcile

performance with the complexities inherent in distributed envi-

ronments. OceanBase is introduced as a case study, and its strong

TPC-C and TPC-H benchmark performances underscore Ocean-

Base as a top-tier distributed database. We also discuss possible

opportunities for native distributed databases.

PVLDB Reference Format:

Quanqing Xu, Chuanhui Yang, and Aoying Zhou. Native Distributed

Databases: Problems, Challenges and Opportunities. PVLDB, 17(12): 4217 -

4220, 2024.

doi:10.14778/3685800.3685839

1 INTRODUCTION

Native distributed databases are built to scale across intercon-

nected nodes, ensuring high availability and resilience against fail-

ures [

]. They use advanced replication algorithms (e.g., Raft [

]

and Paxos [

]) for data synchronization while preserving ACID

properties, essential for eliminating single points of failure. Such

databases, e.g., Google Spanner [

] and OceanBase [

], provide

robust solutions for consistency, data partitioning, and manage-

ment. Their capability to handle heavy data demands makes them

vital for businesses operating over distributed networks, oering

scalable, fault-tolerant database systems.

They excel in distributed computing, scaling elastically and en-

suring data durability with sophisticated replication. Designed

around the CAP theorem, they balance consistency, availability, and

partition tolerance, making them ideal for extensive, reliable data

management across vast computing environments. As they evolve,

these databases are incorporating cloud technologies and machine

learning to enhance scalability and reliability, adeptly meeting both

transactional and analytical needs [

]. They mark a new phase in

∗

Chuanhui Yang is the corresponding author.

This work is licensed under the Creative Commons BY-NC-ND 4.0 International

License. Visit https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of

this license. For any use beyond those covered by this license, obtain permission by

emailing info@vldb.org. Copyright is held by the owner/author(s). Publication rights

licensed to the VLDB Endowment.

Proceedings of the VLDB Endowment, Vol. 17, No. 12 ISSN 2150-8097.

doi:10.14778/3685800.3685839

data management systems, crucial for modern, scalable infrastruc-

ture demands, embodying resilience and eciency.

The proliferation of data and an increased reliance on robust

data management systems emphasize the signicance of native

distributed databases in modern technological ecosystems. They

are instrumental in revolutionizing data management by optimiz-

ing scalability and availability, beneting from the agility of cloud

technologies and the predictive power of machine learning. By sup-

porting both transactional and analytical processes, these databases

signify a shift towards more streamlined, cost-eective, and ca-

pable data management solutions, poised to meet the escalating

requirements for scalable and fault-tolerant data infrastructures.

We present a 1.5-hour tutorial, which is divided into seven

sections as follows: 1) Overview of Native Distributed Data-

base (

∼

5min). It oers scalable, resilient, and ecient large-scale

data management. 2) Data Replication and Synchronization

(

∼

15min). Distributed databases maintain data integrity through

advanced replication despite failures. 3) Consistency Models

(

∼

15min). It describes a variety of consistency models, ranging

from strict consistency to eventual consistency. 4) Distributed

Transactions (

∼

15min). It discusses distributed transactions by

ensuring atomicity, consistency, isolation, and durability (ACID)

across multiple nodes in a distributed environment. 5) Query Pro-

cessing (

∼

15min). It focuses on query processing by distributing

and executing queries eciently across various nodes, optimiz-

ing for reduced network latency and strategic data placement. 6)

Case Study: OceanBase (

∼

10min). It describes that OceanBase

oers a high-performance, scalable, shared-nothing architecture,

excelling in OLTP/OLAP integration. 7) Opportunities (

∼

15min).

Embracing Serverless architecture [

], AI4DB and DB4AI [

multi-model [

], and vector database capabilities [

], it presents

opportunities for unprecedented scalability, autonomous operation,

and versatile data handling in modern computing environments.

Target Audience. The intended audience includes database re-

searchers, developers, and students who aspire to study database

kernel techniques, as well as database administrators (DBAs) who

desire to better tune their database systems. The tutorial is self-

contained and does not require any prerequisite knowledge.

2 TUTORIAL OUTLINE

2.1 Overview of Native Distributed Database

Native distributed databases oer unied systems with high avail-

ability, fault tolerance, scalability, and performance across multiple

nodes and locations. Key problems of native distributed databases in-

clude: 1) Data Replication and Synchronization: These databases

handle synchronization among nodes to keep replicas up-to-date,

which is crucial for data accuracy and disaster recovery. 2) Consis-

tency Models: They oer various consistency models ranging from

4217

strong consistency to eventual consistency, allowing the system to

balance between consistency, availability, and partition tolerance

as dictated by the CAP theorem. 3) Distributed Transactions:

Support for transactions across multiple nodes is often provided,

although it may come with trade-os in terms of performance and

scalability. 4) Query Processing: They can execute queries across

nodes eciently, often optimizing query execution plans to mini-

mize network trac and data movement. Table 1 outlines dierent

mechanisms of popular distributed databases, and Figure 1 illus-

trates three distinct architectures of native distributed databases.

(a) Shared-Nothing (b) Shared-Nothing Disaggr. (c) Shared-Storage

Figure 1: Architectures of Native Distributed Databases

2.2 Data Replication and Synchronization

2.2.1 Data Replication. Replication in distributed databases strikes

a balance between synchronous methods, which ensure immediate

consistency but result in slower writes, and asynchronous methods,

which allow for faster access but may introduce data discrepancies.

The level of replication—be it at the row, block, or le level—is

chosen based on specic needs. Asynchronous replication resolves

conicts using timestamps or custom logic to ensure data accu-

racy. Techniques like asymmetric-partition replication can reduce

system load [

], whereas solutions like BatchDB [

] improve

OLTP/OLAP workloads by pairing logical replication with a lazy

strategy for enhanced performance.

2.2.2 Data Synchronization. Distributed databases balance consis-

tency and performance using models like eventual consistency [

which tolerates short-term discrepancies for assured long-term ac-

curacy. Data sync frequency and network latency [

] are crucial

to this consistency, inuencing system design for performance op-

timization. Moreover, to handle simultaneous transactions and data

conicts, strategies such as version control and timestamp [

] help

maintain orderly data sync and ensure steadfast consistency.

2.2.3 Challenges. In distributed databases, especially those span-

ning wide areas, data synchronization is essential but challenged

by bandwidth limits, network latency and uctuation, aecting

eciency and performance [

]. Optimizing bandwidth and main-

taining swift recovery post-failure are crucial for data integrity and

loss prevention. Ensuring transactional consistency amid partitions

and securing data against unauthorized changes during replication

are key hurdles. However, as demands for performance rise, evolv-

ing technologies are progressively tackling these issues, enhancing

the robustness and reliability of distributed database systems.

2.3 Consistency Mo dels

The data consistency models in distributed databases crucially im-

pact performance, reliability, and availability, as depicted in Figure 2.

2.3.1 Strong and Eventual Consistency. Strong consistency in dis-

tributed systems ensures that operations are immediately visible

and executed in sequence, providing a seamless experience but

potentially limiting performance due to the need for node syn-

chronization. Systems such as Calvin [

], PolarDB [

], and Ge-

oGauss [

] have improved transactional eciency and replication

to deliver this consistent state without signicantly aecting speed

or scalability. In contrast, eventual consistency allows for short-

term data anomalies in exchange for better responsiveness, with

systems like Dynamo [

] managing high-performance demands

through application-level conict resolution. BlockchainDB [

] in-

novatively combines the exibility of databases with the strength

of blockchain technology to oer a range of consistency levels, thus

optimizing data management for a variety of operational contexts.

2.3.2 Other Consistency Models. Causal consistency improves upon

eventual consistency by ensuring causally related operations fol-

low the same order across nodes, while independent operations

are not strictly ordered. Eciency gains come from minimizing

dependency checks and dening external causal relationships [

GentleRain [

] increases throughput with time-based protocols and

uses physical timestamps to save on storage and communication.

Orbe [

] leverages dependency matrices and transitive causality

for eective causal consistency in key-value systems. Achieving

strong consistency in partitioned, replicated systems is challenging.

Google Spanner [

] clusters servers and uses Paxos for log repli-

cation within groups, maintaining a consistent prex order across

data replicas to uphold its consistency standard.

2.3.3 Challenges. Choosing the right consistency for distributed

databases is crucial for balancing performance, availability, and

precision. Financial systems often require strong consistency, while

CDNs may opt for eventual consistency, accepting brief data discrep-

ancies. The CAP theorem advises a trade-o between consistency,

availability, and partition tolerance, inuenced by application needs.

Databases like OceanBase [

] adapt consistency options for diverse

scenarios. Data replication, key for fault tolerance, faces latency

issues with synchronous methods and potential inconsistency with

asynchronous ones. Developers must navigate these complexities

to ensure optimal system performance and data reliability.

2.4 Distributed Transactions

2.4.1 Distributed Transaction Commit Protocols. Native distributed

databases employ distributed transaction commit protocols to up-

hold the ACID properties across distributed nodes, ensuring data

integrity and consistency. The 2PC protocol is a classic example, op-

erating in two distinct stages. ROCOCO [

] optimizes this process

by treating transactions as collections of atomic blocks, tracking

dependencies before execution to allow for serializable ordering

upon commit. Primo [

] avoids concurrency conicts by ensuring

transactions are conict-free after the commit phase. We proposed

OceanBase 2PC [

], a Paxos-enhanced 2PC protocol, to strengthen

fault tolerance and reduce transaction latency in distributed envi-

ronments, streamlining synchronizations for eciency.

2.4.2 Distributed Version Control. One of the core principles of

SAP HANA is full support for distributed query capabilities and

horizontal expansion [

]. It employs MVCC to provide distributed

4218