MatrixOne Architecture Design
MatrixOne is a future-oriented hyperconverged cloud & edge native DBMS that supports transactional, analytical, and streaming workload with a simplified and distributed database engine working across multiple datacenters, clouds, edges, and other heterogenous infrastructures. This combination of engines is called HSTAP.
As a redefinition of the HTAP database, HSTAP aims to meet all the needs of Transactional Processing (TP) and Analytical Processing (AP) within a single database. Compared with the traditional HTAP, HSTAP emphasizes its built-in streaming capability used for connecting TP and AP tables. This provides users with an experience that a database can be used just like a Big Data platform, with which many users are already familiar thanks to the Big Data boom. With minimal integration efforts, MatrixOne frees users from the limitations of Big Data and provides one-stop coverage for all TP and AP scenarios for enterprises.
MatrixOne Architecture Layers
MatrixOne implements three independent layers, each with its object units and responsibilities. Different nodes can freely scale, no longer constrained by other layers. These three layers are:
- Compute Layer: Based on Compute Nodes (CNs), MatrixOne enables serverless computing and transaction processing with its cache, which is capable of random restarts and scaling.
- Transaction Layer: Based on Transaction Nodes and Log Services, MatrixOne provides complete logging services and metadata information, with built-in Logtail for recent data storage.
- Storage Layer: Full data is stored in object storage, represented by S3, implementing a low-cost, infinitely scalable storage method. A unified File Service enables seamless operations on underlying storage by different nodes.
After deciding on TAE as the sole storage engine, multiple design adjustments were made to the fused TAE engine, resulting in the TAE storage engine. This engine has the following advantages:
- Columnar Storage Management: Uniform columnar storage and compression methods provide inherent performance advantages for OLAP businesses.
- Transaction Processing: Shared logs and TN nodes jointly support transaction processing for compute nodes.
- Hot and Cold Separation: Using S3 object storage as the target for File Service, each compute node has its cache.
The compute engine is based on the fundamental goal of being compatible with MySQL, with higher requirements for node scheduling, execution plans, and SQL capabilities. The high-performance compute engine has both MPP (massively parallel processing) and experimental architecture:
- MySQL Compatible: Supports MySQL protocol and syntax. Fused Engine: Rebuilds execution plans based on DAG, capable of executing both TP and AP.
- Node Scheduling: Future support for adaptive intra-node and inter-node scheduling, meeting both concurrency and parallelism requirements.
- Comprehensive SQL Capability: Supports subqueries, window functions, CTE, and spill memory overflow processing.
MatrixOne Architecture Design
The MatrixOne architecture is as follows:
The architecture of MatrixOne is divided into several layers:
Cluster Management Layer
Being responsible for cluster management, it interacts with Kubernetes to obtain resources dynamically when in the cloud-native environment, while in the on-premises deployment, it gets hold of resources based on the configuration. Cluster status is continuously monitored with the role of each node allocated based on resource information. Maintenance works are carried out to ensure that all system components are up and running despite occasional node and network failures. It rebalances the loads on nodes when necessary as well. Major components in this layer are:
- Prophet Scheduler: take charge of load balancing and node keep-alive.
- Resource Manager: being responsible for physical resources provision.
Serverless Layer is a general term for a series of stateless nodes, which, as a whole, contains three categories:
- Background tasks: the most important one is called Offload Worker, which is responsible for offloading expensive compaction tasks and flushing data to S3 storage.
- SQL compute nodes: responsible for executing SQL requests, here divided into write nodes and read nodes. The former also provides the ability to read the freshest data.
- Stream task processing node: responsible for executing stream processing requests.
As MatrixOne's Single Source of Truth, data is considered as persistently stored in MatrixOne once it is written into the Log Layer. It is built upon our world-class expertise in the Replicated State Machine model to guarantee state-of-the-art high throughput, high availability, and strong consistency for our data. Following a fully modular and disaggregated design by itself, it is also the central component that helps to decouple the storage and compute layers. This in turn earns our architecture much higher elasticity when compared with traditional NewSQL architecture.
The storage layer transforms the incoming data from the Log Layer into an efficient form for future processing and storage. This includes cache maintenance for fast accessing data that has already been written to S3.
In MatrixOne, TAE (Transactional Analytic Engine) is the primary interface exposed by the Storage Layer, which can support both row and columnar storage together with transaction capabilities. Besides, the Storage Layer includes other internally used storage capabilities as well, e.g. the intermediate storage for streaming.
Storage Provision Layer
As an infrastructure agnostic DBMS, MatrixOne stores data in shared storage of S3 / HDFS, or local disks, on-premise servers, hybrid, and any cloud, or even smart devices. The Storage Provision Layer hides such complexity from upper layers by just presenting them with a unified interface for accessing such diversified storage resources.
MatrixOne System Components
In MatrixOne, to achieve the integration of distributed and multi-engine, a variety of different system components are built to complete the functions of the architecture-related layers:
File Service is the component of MatrixOne responsible for reading and writing all storage media. Storage media include memory, disk, object storage, and so on., which provide the following features:
- File Service provides a unified interface so that reading and writing of different media can use the same interface.
- The design of the interface follows the concept of immutable data. After the file is written, no further updates are allowed. The update of the data is realized by generating a new file.
- This design simplifies operations such as data caching, migration, and verification and is conducive to improving the concurrency of data operations.
- Based on a unified read-write interface, File Service provides a hierarchical cache and a flexible cache strategy to balance read-write speed and capacity.
Log Service is a component specially used to process transaction logs in MatrixOne, and it has the following features:
- The Raft protocol ensures consistency, and multiple copies are used to ensure availability.
- Save and process all transaction logs in MatrixOne, ensure that Log Service logs are read and written typically before the transaction is committed, and check and replay the log content when the instance is restarted.
- After the transaction is submitted and placed, truncate the content of the Log Service to control the size of the Log Service. The content that remains in the Log Service after truncation is called Logtail.
- If multiple Log Service copies are down at the same time, the entire MatrixOne will be down.
The Transaction Node (TN) is the carrier used to run MatrixOne's distributed storage engine TAE, which provides the following features:
- Manage metadata information in MatrixOne and transaction log content saved in Log Service.
- Receive distributed transaction requests sent by Computing Node (CN), adjudicate the read and write requests of distributed transactions, push transaction adjudication results to CN, and push transaction content to Log Service to ensure the ACID characteristics of transactions.
- Generate a snapshot according to the checkpoint in the transaction to ensure the snapshot isolation of the transaction, and release the snapshot information after the transaction ends.
The computing node (CN) is a component of Matrixone that accesses user requests and processes SQL. The toolkit includes the following modules:
- Frontend, it handles the client SQL protocol, accepts the client's message, parses it to get the executable SQL of MatrixOne, calls other modules to execute the SQL, organizes the query results into a message, and returns it to the client.
- Plan, parse the SQL processed by Frontend, generate a logical execution plan based on MatrixOne's calculation engine and send it to Pipeline.
- Pipeline, which parses the logical plan, converts the logical plan into an actual execution plan and then runs the execution plan through Pipeline.
- Disttae, responsible for specific read and write tasks, including synchronizing Logtail from TN and reading data from S3, and sending the written data to TN.
Stream Engine is a new component within MatrixOne, serving as an integrated stream engine designed for real-time querying, processing, and enriching data stored in a series of incoming data points, also known as data streams. With Stream Engine, you can employ SQL to define and create streaming processing pipelines, offering real-time data backend services. Additionally, you can utilize SQL to query data within streams and establish connections with non-streaming datasets, thereby further streamlining the data stack.
The Proxy component is a powerful tool mainly used for load balancing and SQL routing. It has the following functions:
- Through SQL routing, resource isolation between different accounts is realized, ensuring that the CNs of different accounts will not affect each other.
- Through SQL routing, users can do a second split in the resource group of the same account, improving resource utilization.
- The load balancing between different CNs is realized in the second split resource group, making the system more stable and efficient.
This page outlines the overall architecture design of MatrixOne. For information on other options that are available when trying out MatrixOne, see the following: