Community Blog Integrating Serverless with SaaS-based Cloud Data Warehouses

Integrating Serverless with SaaS-based Cloud Data Warehouses

This article describes business scenarios and resource usage requirements of modern cloud data warehouses and analyzes different resource delivery modes.

By Ning Qu

This article describes business scenarios and resource usage requirements of modern cloud data warehouses, and differentiation analysis of different resource delivery modes. Based on the features and values of Serverless MaxCompute, this article also introduces best practices of Serverless MaxCompute.

1. Introduction to Serverless

The following figure shows the Serverless architecture of MaxCompute, which mainly includes modules of the ingest service, multi-computing environment, storage service and management.


The main features of each module are as follows:

Serverless ingest service

  • Support Tunnel batch ingest and streaming ingest to convert data format into MC column storage format, which can automatically scale and is free of charge;
  • Provide Load and Unload commands to import and export OSS data free of charge.

Serverless multi-computing environment

  • Serverless computing resource pool is an on-demand large-scale computing resource pool in pay-as-you-go billing mode;
  • Exclusive computing resources support subscription billing mode and workload management, including load isolation, priority, and time-sharing scaling;
  • ETL, OLAP, ML, and other big data analysis scenarios are supported during runtime.

Serverless storage service

  • Support independent scaling from GB level to EB level;
  • Charge the storage space in pay-as-you-go mode to reduce costs;
  • Perform analysis-oriented optimizations by default, such as column pressure and compression;
  • Support partitions, buckets, Zorder, and other optimization methods.

Serverless management

  • Provide comprehensive built-in management capabilities, which manage resources through API, SDK, or web-console;
  • Require no user O&M on the platform, which reduces costs.

The above is a brief introduction to the Serverless architecture. The focus of this article is how to use the Serverless computing resources of MaxCompute to meet the requirements of data warehouses.

The following figure shows the logic model for management and usage of MaxCompute computing resources. A Project in MaxCompute corresponds to a logical isolation unit of data warehouses. Different projects can be created for different management objectives. For example, a test-oriented project and a development-oriented project can be created respectively. The two projects have independent data and permission management systems, which are not related. Therefore, the management of them is isolated. However, such isolation space is not enough, because computing tasks need to bind computing resources. By binding projects with payment methods, different payment methods can be set for different projects according to the demands, so that different isolation spaces use different computing resources.


Under the above system, MaxCompute has some unique characteristics. First, it has a multi-tenant environment. Multiple isolated data warehouse spaces can be created according to different management requirements in MaxCompute. Enterprises can purchase multiple groups of logical computing resources, which provide multiple computing resources in isolated environments to better meet the needs of different scenarios.

As shown in the following figure, the ideal Serverless resource model requires well-planned resource utilization methods to fully adapt to actual requirements (the black line).


However, customers have different resource requirements and diversified demand scenarios. Main scenarios are as follows:

  • Scenarios of stable periodic jobs;
  • Scenarios of fast-growing business and fast-changing demands;
  • Scenarios of routine demands with emergencies;
  • Scenarios of test or development demands.

For these scenarios, big data computing demands for computing resources not in the completely Serverless on-demand allocation mode. In fact, it has different demands at different stages, while different types of demands have different requirements for computing resources.

Main features of computing resource demands are as follows:

Business agility requirements

  • Business is in the long-term growing stage, and the processing capability should meet the needs of natural business growth, especially in the stage of rapid business change;
  • Computing resources are required for business in the initial stage or the innovative business of the innovation department.

Significant differences between periodic peaks and valleys

  • The daily and monthly periodic fluctuations of peak and valley traffics are huge. If enterprises plan resources based on the peak capacity, the cost and SLA are difficult to balance;
  • Conventional computing power and elastic computing power with job resource policies are scheduled or manually specified.

Stable business with focus on the SLA-based output of critical jobs

  • Unlike non-critical jobs, baseline jobs have different requirements for SLA. The output time of baseline jobs must be guaranteed;
  • Non-critical jobs are processed at a cost as low as possible, without affecting critical jobs.

Resource governance: Computing power demands become stably predictable rather than rapidly changing

  • CU capacity planning, mutual conversion, and calculation;
  • Refined workload management of fixed resources.

In general, the ultimate goal of computing power is to minimize costs on the premise of meeting differentiated needs in reality.

2. Serverless Benefits for Business Agility

Then, how does MaxCompute Serverless meet the demands in the above scenarios? For an enterprise with rapid business development and business change, on-demand computing resources of MaxCompute Serverless are recommended. From a management point of view, different projects can be created for isolation, such as development test environment and production environment.

For some analysts, they often need to do a lot of exploration or machine learning analysis on some detailed data randomly. However, there are often some sudden computing power demands, which may be very large. In this case, these jobs need to be isolated from other environments because they are infrequent but require analysis of massive amounts of data.

Environments can also be isolated based on organizations. For example, many enterprises have relatively large organizations and they can isolate environments according to departments. By doing so, each department has an isolated environment and each department is an independent organization. As they require relatively independent data and computing resources, Serverless on-demand allocation mode can be adopted. With this mode, enterprises do not need capacity planning. In the initial stage, they can use the pay-as-you-go mode. The large resource pool meets the resource demands of various departments, avoiding resource competition.

In conclusion, Serverless can meet all demands well in a variety of job scenarios. In the single-job scenarios, Serverless can meet resource demands for jobs of different scales. In the scenarios of multi-job concurrency, Serverless can also meet multi-job resource demands and avoid resource competition. In the scenarios of cost control, MaxCompute can achieve cost estimation and control by preventing costly jobs. Through the preceding methods, MaxCompute and Serverless can greatly improve business agility and accelerate value realization.


In addition, in consideration of their daily management environment, some enterprises prefer a relatively stable resource pool, because they have certain resource planning and resource governance capabilities. In this case, they can purchase resources of a fixed specification and isolate the environments by functions or organizations. Then, they can use the quota group management capability provided by MaxCompute to divide the resources into multiple resource groups. By doing so, these resources groups can meet the demands of different businesses and organizations on the premise of predictable finance.

The key technical features of this mode are as follows:

  • Load isolation for avoiding competition and focusing on preferential critical projects and organizations;
  • Job priority: Procedure detection of critical dynamic jobs;
  • Time-sharing scaling: Daytime and nighttime resource allocation policies to maximize utilization.

MaxCompute Management

The third scenario is the balance of costs and business agility. For example, data platform managers may often face a variety of jobs. Some are daily jobs. This kind of job usually requires fixed-size resources with controllable and predictable costs. Some are key jobs. For this kind of job, certain costs are required and allowed to accelerate job processes for business needs. Additional computing power is required for these jobs. Others are exploration jobs for data scientists.

These jobs are expected to not interfere with production jobs, while at the same time data scientists can use powerful computing power to quickly make business assumptions. These jobs can be allocated with an on-demand resource pool. Some enterprises may have innovative business. They need a new environment for data development and application innovation. Thus, an isolated environment for data warehouses can be created to allocate resources on demand and help them quickly verify business assumptions.

Two capabilities are provided for users:

  • Pay-as-you-go project: The initiated job uses Serverless resources and can switch the resource group bound to the project;
  • Personal setting: Computing resource and job-level resource routing can be set according to the actual needs.


The above three scenarios are real scenarios in daily operations. Another scenario is that business stabilizes after the customer uses resources in pay-as-you-go mode for a period of time. Therefore, the customer wants to put the project on a fixed resource pool in subscription mode. However, the customer wonders how to assess the resource demand. After all, resource demand estimation was not required in the previous pay-as-you-go mode. MaxCompute provides the capacity planning feature to solve this problem. It uses information schema to estimate the overall computing power demand of a project based on historical computing power consumption.

The key information includes:

  • Number of compute hours (unit: cu•h) consumed by recent project jobs on a daily basis, based on information schema;
  • Hourly computing power demand (unit: cu•h) in the day with the highest consumption by recent project jobs, based on information schema.

Based on the preceding information, MaxCompute can predict the computing power demands of the business and conduct capacity planning based on certain rules. For more information about this feature, see the relevant articles in Alibaba Cloud community.

3. Summary

This article describes how to use Serverless to better manage resources and meet the resource demands of different businesses at low costs. It can be summarized as follows:

(1) The pay-as-you-go billing mode is applicable to fast-growing business and fast-changing demands. With cost control of MaxCompute, the cost can be controlled while meeting the business computing power needs.

(2) For subscription resources, multiple computing resources are segmented for corresponding load isolation and time-sharing scaling through quota management. The feature of baseline job priority provided by DataWorks and MaxCompute can guarantee the SLA of key jobs.

(3) For combination of pay-as-you-go and subscription modes, different computing resources can be chosen based on the job level. For emergency jobs, resources in pay-as-you-go mode can meet the demand for sudden computing power. For routine periodic jobs with spikes, resources in pay-as-you-go mode are also available to realizing efficiently utilize resources and reduce costs.

(4) The metadata can be used to assess the computing power demand for capacity planning. Thus, customers can choose the pay-as-you-go or subscription mode. Besides, metadata can also be used to analyze resource consumption, optimize resources, reduce the number of jobs with high resource consumption, and manage resources accordingly.

0 0 0
Share on

Alibaba Cloud MaxCompute

135 posts | 18 followers

You may also like


Alibaba Cloud MaxCompute

135 posts | 18 followers

Related Products