作者：贾明国、元乙
大促前一周，老板拍板：这次促销要上新服务。
你打开电脑，开始排兵布阵。ECS 选什么规格、VPC 怎么划、SLB 怎么配、RDS 开多大、监控接哪些指标——五个控制台，十几张表单，中间哪个依赖关系没理清楚就得回头重来。三天后服务上线了，凌晨你还在补监控大盘和告警规则。又过了四天，促销开始，凌晨两点四十七分，告警响了。CPU 打满。
你摸出手机，看指标、翻日志、对着资源拓扑找线索。三张标签页四块大盘拼到一起，才勉强回答出三个问题：发生了什么，严重吗，现在该干什么。需要扩容，但哪些资源可以横向扩展？扩多少合适？有没有配额？又是一轮控制台切换。
选规格、配网络、接监控、查告警、做扩容——单看都不难，难在它们同时压在一个人身上。从部署到上线，从监控到扩容，每个阶段都在重复同样的模式：找控制台、填表单、等执行、再确认。
如果有一个搭档，能替你记住所有控制台的位置、所有资源的依赖关系、所有参数的默认值，你只需要说清楚要什么、然后在关键节点拍板——剩下的事，它来做。
这个搭档就是 STAROps 数字员工加上 ROS Agent。
STAROps 是阿里云基于大模型和智能体技术打造的全域智能运维平台。它深度融合跨域可观测数据与大语言模型推理能力，突破传统运维工具使用门槛高、数据孤岛严重等局限，支持用户通过自然语言定义目标，由运维智能体自主完成动态规划、安全执行与结果验证的全闭环。STAROps 实现运维模式从被动响应向主动自治的根本性转型，为企业提供 7x24 小时的自主运维能力，实时守护客户的业务系统韧性。
ROS 是阿里云的基础设施即代码（IaC）服务，通过模板描述云资源及其依赖关系，自动完成资源的创建、更新和删除。
ROS Agent 是面向云基础设施的资源管理智能体，通过自然语言交互，辅助完成资源规划与选型、ROS／Terraform 模板生成与校验、费用评估，以及授权后的资源创建、变更、查询和清理。
STAROps 数字员工提供统一的运维对话入口，结合日志、指标、告警等信息理解问题，并通过 alibabacloud-ros-agent 技能，将基础设施规划和变更任务委派给 ROS Agent。ROS Agent 围绕资源目标进行必要澄清，规划资源及其依赖关系，展示方案、关键参数和预估费用，并在获得授权后执行资源操作。双方通过会话上下文衔接多轮任务，执行范围受账号权限和授权策略约束。
对你来说，你面对的始终是 STAROps 的对话框。哪个控制台管什么、下一步该去哪个页面，这些路由归它管；说清楚要什么、然后拍板，这两件事归你。
下面用一个电商促销服务串联三个场景：大促前一周部署上线，上线后第三天接入数据和告警，促销当天应对容量瓶颈。每个场景展示你说了什么、Agent 做了什么、最终交付了什么。
场景一：大促倒计时 7 天，新服务怎么稳稳上线？
距离大促只剩一周，运维同学接到任务：新促销服务要上线，杭州地域，日常 2000 QPS，峰值 8000 QPS。机器选多大？要几台？网段怎么划？数据库扛得住吗？翻文档、问同事、照着去年的方案改参数，改完还是心里没底：万一开场被打挂了呢？
▍提问
杭州地域部署电商促销服务，日常 2000 QPS，峰值 8000 QPS，要负载均衡和数据库。参考去年的活动服务，但今年流量更大，帮我看看哪里要改。
▍Agent 回复
ROS Agent 收到需求后，分五步推进：
- 明确目标和参考依据： 确认历史方案、已有资源和可用的监控数据；针对高可用、预算及应用部署方式等关键要求进行必要澄清。历史运行数据可由 STAROps 查询后提供给 ROS Agent，作为规划参考。
- 提出基础设施方案： 围绕负载均衡、计算、数据库及弹性伸缩设计资源架构，说明提前准备容量与按需扩容的取舍。容量建议需结合应用特征和压测结果验证；如建议引入 Redis，还需明确业务接入和一致性要求，创建缓存资源本身并不等于完成业务改造。
- 核对资源约束： 查询相关规格、可用区和库存，结合可获取的网络规划及配额信息检查部署约束。发现冲突或资源不可用时，说明影响并提出替代方案；涉及网段、可用区等关键调整时，经用户确认后修改。
- 将方案变成可审查的执行计划： 生成并校验模板，展示资源架构、关键参数和预估费用。对可以合理推断的参数给出建议并说明依据，对需要用户决定的事项继续澄清，不预设固定的追问数量。
- 确认后执行，查询真实结果： 用户确认方案并授权后，ROS Agent 通过 ROS 执行部署，跟踪资源栈和资源状态，返回执行结果。应用发布、预发压测和业务验收，需要结合实际应用及相应工具另行完成；模板校验和资源部署成功不等于业务压测通过。
▍小结
交付物包括：一套可复用的基础设施模板、一份关键配置及预估费用说明，以及资源栈执行结果和相关资源信息，便于后续复用和迭代。若任务还包含应用部署，则进一步验证并返回应用访问入口。
点击链接，查看视频：mp.weixin.qq.com/s/SoIlSElG8…
场景二：一句话接入数据和告警
服务上线后，还要让运行状态看得见。先接入日志、指标和 APM 应用监控，才能知道请求慢在哪里、错误发生在哪个环节；随后配置告警，在异常发生时及时通知到人。
▍提问
/数据接入 帮我给刚才部署的促销服务接入日志、指标和 APM 应用监控，再配置告警，通知发到运维钉钉群。
▍Agent 回复
STAROps 提供专门的 /数据接入 技能。Agent 复用当前会话中的资源栈信息，确认服务、主机和依赖资源的范围，检查已有指标，并引导完成日志采集、所需客户端和 APM 探针的接入。待日志、指标和调用链数据能够正常查询后，再配置 CPU、内存、负载均衡及应用错误率、响应时间等告警，并确认阈值、持续时长和通知渠道。
建议先确认日志、指标和调用链数据能正常查询，再配置基础资源与应用告警。我会先列出采集范围、接入方式、告警规则和通知渠道，确认后执行。
▍小结
先有可查询的数据，再有可用的告警。 通过 /数据接入 技能，用户可以围绕同一个促销服务说明需求，由 Agent 引导完成数据接入和告警配置；后续排障也有了日志、指标和调用链作为依据。
场景三：告警响起，快速定位根因并恢复
促销正式开始。这里以开场前预热未及时完成的情况为例：伸缩组的 CPU 目标值为 60%，但流量到来时，只有 6 台 ECS 实例在承接请求，CPU 已超过 85%。伸缩组已经触发扩容，新实例还在启动和应用就绪过程中，告警先于扩容完成触发。下文数量和目标值用于说明决策过程，实际容量需结合实时流量和压测结果确定。
▍提问
促销服务 CPU 告警了，帮我查一下根因、业务影响，以及现在有哪些恢复方案。
▍Agent 回复
STAROps 关联告警、日志、指标、APM 调用链和资源拓扑，判断业务影响及瓶颈位置，并核对近期变更。ROS Agent 根据诊断信息，按需查询相关资源状态、可用性及约束，辅助制定基础设施变更方案。在这个示例中，诊断指向开场已就绪容量不足：流量增长快于新增实例投入服务的速度。下面给出三种处置思路，实际方案需结合查询结果、执行条件和业务要求确定。
当前 6 台 ECS 实例 CPU 均已超过 85%，负载均衡后端出现健康检查超时，数据库和网络指标未见明显异常。伸缩组已触发扩容，新实例尚未就绪。调低 CPU 目标值不能缩短这批实例的启动时间。下面按对现有运行环境的变更幅度，从激进到保守给出三种方案。
在这个场景中，Agent 优先推荐方案二，按容量缺口补足同规格实例。如果核对发现正在启动的实例已足够承接流量，且业务指标仍可接受，则可以选择方案三继续观察。
用户选定方案后，ROS Agent 说明目标容量、预估费用及执行条件，获得确认和授权后执行，并查询资源变更结果。STAROps 再结合应用就绪状态、后端健康检查、错误率和响应时间等观测数据验证恢复效果。资源变更成功不等于业务已经恢复。
三个场景，一条主线
适用范围
ROS Agent 适合需要资源规划、依赖编排和多步骤执行的基础设施任务。典型场景包括：根据业务目标规划云资源；生成、解释、校验和修改 ROS／Terraform 模板；创建、更新、查询和清理云资源；通过多轮对话澄清需求、调整方案和确认操作。资源数量不是能力门槛。
对于单个确定资源的简单查询或普通云产品知识问答，STAROps 可以使用自身已有能力直接处理，无需额外委派；与基础设施无关的任务不属于本文介绍的范围。
本文介绍和验证的集成范围为阿里云资源管理，涉及 ECS、RDS、负载均衡、VPC、Redis、OSS 等云产品。具体可执行操作取决于资源类型、服务能力及账号权限。涉及创建、更新、删除等资源变更时，Agent 会展示操作范围及影响，并按授权策略请求确认；执行仍受 RAM 权限约束。
快速开始
- 登录 STAROps 控制台（ starops.console.aliyun.com/ ），创建一个数字员工。
- 在技能中心搜索 alibabacloud-ros-agent 并挂载。
- 在对话框里描述你的部署或排障需求。
实际耗时与结果取决于你账号内的资源规模、监控数据完整度和所选方案，建议先用一个非生产需求验证流程。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Submitted by: Jamin, Yuan B
A week before the big push, the boss boarded: the promotion was for a new service.
You turn on the computer and start lined up. The ECS selects the specifications, how the VPC works, how the SLB fits, how large the RDS is, what indicators are the surveillance -- five control tables, a dozen forms, which depend on the middle, and which does not. The service's online in three days, and you're still monitoring the disk and the alarm rules in the morning. Four days later, the promotion began, at 2:47 a.m., the alarm went off. CPU full.
You touch the phone, look at the indicators, flip the logs, look for clues to the resources. Three tabs, four large discs put together, barely answered three questions: what happened, is serious, what should be done now. Expansion is needed, but what resources can be expanded horizontally? How much is appropriate? Is there a quota? Another round of the console switch.
Selecting specifications, matching networks, receiving surveillance, checking for alarms and building up-- it's easy to look at them alone, and it's hard to put them on one person at the same time. The same pattern is repeated at every stage, from deployment online, from surveillance to build-up: looking for a control desk, filling in forms, waiting for execution, confirmation.
If you have a partner who can remember the location of the console, the dependency of all resources, the default values for all parameters, you just have to say what you want and then board the key nodes — the rest, it does.
This partner is STAROPs digital employee plus ROS Agent.
STAROPs are a global smart platform based on LLM and AI agent technology. It deeply integrates inter-domain observational data and large-linguistic modelling capabilities, transcends the limitations of the use of traditional mobility tools with high thresholds, data isolations, etc., and supports users in defining their goals in natural languages, with the full closed loop of dynamic planning, safe implementation and validation of results autonomously by Transport AI agent. STAROPs achieves a fundamental transition from reactive response to active self-government, providing enterprises with 7x24 hours of autonomous mobility and real-time protection of client business system resilience.
ROS is the Alibaba Cloud infrastructure, i.e., IaC, which automatically completes the creation, updating and removal of cloud resources through templates describing them and their dependence.
ROS Agent is the resource management AI agent for cloud-based infrastructure, which interacts through natural languages to support the completion of the resource planning and selection, the production and validation of the ROS/Terraform template, the cost assessment, and the creation, change, query and clean-up of the authorized resources.
STAROPs digital staff members provide a unified access portal to transport dialogue, understand issues in conjunction with information such as logs, indicators, alarms, and delegate infrastructure planning and change tasks to ROS Agent through the Alibacloud-ros-agent skills. ROS Agent provides the necessary clarifications around resource objectives, plans resources and their dependencies, showcases programmes, key parameters and estimated costs, and implements resource operations when delegated. The parties were bound by account number permission and the enabling strategy by means of a session in which context was linked to multiple rounds of missions.
For you, you've always faced a STAROPs dialogue box. Which console is what, which page next step is going to go, and which route is going to go to it; and what is going to be done, and then board, is yours.
The following is a series of three scenarios with a vendor promotion service: deployment online a week before the launch, access data and alarms on the third day after the launch, and response to capacity bottlenecks on the day of the promotion. Every scene shows what you said, what Agent did, what he finally delivered.
Scene one: 7 days, how does the new service get on the line?
Only a week away from the big deal, the new promotion service is on line, Hangzhou territory, 2000 QPS, peak 8000 QPS. How big are the machines? How many? How many grids? If you look at the files, ask your colleagues, change the parameters according to last year's plan, or you don't know what's going on: what if you get hung up?
Question.
Hangzhou Geographically Deployed Electrician Promotion Service, Daily 2000 QPS, Peak 8000 QPS, Load Balance and Database. Reference is made to last year's activity service, but this year's traffic has increased, so help me see where to change.
Agent Reply
ROS Agent received the demand and advanced in five steps:
- Clarity of objectives and references: Identification of historical programmes, available resources and available monitoring data; necessary clarification of key requirements such as high availability, budget and application of deployment modalities. Historical running data can be made available to ROS Agent for planning purposes by STAROPs.
- Presentation of infrastructure options: Designing the resource architecture around load balance, computing, database and elasticity, describing the trade-off between pre-prepared capacity and needs-based expansion. The capacity proposal needs to be combined with the validation of application features and pressure results; if it is proposed to introduce Redis, it also needs to clarify business access and consistency requirements, and the creation of cache resources does not in itself amount to the completion of business transformation.
- Reconciling resource constraints: Checking relevant specifications, available areas and inventories, combined with available network planning and quota information check deployment constraints. When conflicts are detected or resources are not available, the impact is described and alternative options proposed; when key adjustments are involved, such as grids, available areas, etc., are modified after user confirmation.
- Translating the programme into a reviewable implementation plan: Generate and validate templates showing the resource architecture, key parameters and estimated costs. Recommendations and justifications are given for parameters that can reasonably be inferred, continuing to clarify matters that need to be determined by the user, without setting a fixed number of queries.
Following confirmation and authorization by the user, ROS Agent is deployed through ROS, tracking the resource bank and resource status and returning to implementation. Application of publish, pre-issuing and operational acceptance check needs to be completed in conjunction with practical applications and corresponding tools; template validation and resource deployment success does not amount to operational pressure adoption.
Zilong's knot.
Deliverables include a reusable infrastructure template, a key configuration and a statement of estimated costs, as well as results of implementation of the resource bank and related resource information to facilitate subsequent reuse and iterativeization. If the mission also includes an application deployment, further validation and return to the application access point.
Click on the link to view the video: mp.weixin.qq.com/s/SoilselG8...
Scene two: a sentence for access to data and alarms
Once the service is online, it will also be visible to the operational state. Access to logs, indicators and APM application monitoring is required to know where the request is slow and where the error occurred; the alarm is then configured to inform a person in a timely manner when an anomaly occurs.
Question.
/ Data access Help me monitor the distribution service access logs, indicators and APM applications that have just been deployed, configure the alarms and send them to the spikes.
Agent Reply
STAROPs provide specialized/data access skills. Agent repeats the resource stack information from the current session, confirms the range of services, hosts and resource dependence, checks the existing indicators and directs access to log collections, required clients and APM probes. The CPU, memory, load balance and application error rate, response time, etc., are then configured and the threshold, duration and channel of notification are confirmed once logs, indicators and call chain data are properly searchable.
It is recommended to confirm that logs, indicators and call chain data are routinely accessed, and then to allocate basic resources and application alarms. I will first set out the scope of collection, access, the rules for warning and the channels for notification, and then implement them.
Zilong's knot.
Accessible data are available, followed by available alarms. Through the /Data access skills, users can articulate their needs around the same promotional service, with Agent leading the data access and alert configuration; the subsequent barrier is also based on logs, indicators and call chains.
Scene three: Call the alarm, quickly locate the root causes and recover.
Promotions are officially launched. For example, preheating was not completed in time before opening: the CPU target for the stretching group was 60 per cent, but when the flow came, only 6 ECS examples were receiving requests, and CPU had exceeded 85 per cent. The condensation group has already triggered the build-up, and new examples are still in the process of being activated and applied, before the start-up occurs. The following numbers and target values are used to describe the decision-making process, with actual capacity to be determined in combination with real-time flow and pressure results.
Question.
The promotion service, CPU, called the police.
Agent Reply
STAROPs Link alerts, logs, indicators, APM transfer chains and resource scaling up to determine business impact and bottlenecks, and to reconcile recent changes. Based on diagnostic information, ROS Agent develops infrastructure change programmes by searching for resource status, availability and constraints, as required. In this example, diagnostics point to insufficient start-up capacity: traffic growth is faster than new examples are put into service. Three approaches are given below, and the actual programme needs to be determined in conjunction with the results of the inquiry, the conditions of implementation and the operational requirements.
There are currently more than 85 per cent of the 6 ECS examples CPUs, and there are health checks at the back end of the load balance timeout, and there are no apparent anomalies in the database and network indicators. The condensation group has triggered an expansion and new examples are not yet available. Lowering the CPU target does not shorten the start time of this batch. Three options, from radical to conservative, are given below, depending on the extent of the changes in the existing operating environment.
In this scenario, Agent gives priority to option II, which will complete the example of the same specification by capacity gap. Option III could be continued to observe if the reconciliation found that the start-up case was sufficient to absorb the flow and operational indicators were acceptable.
After the user selects the program, ROS Agent describes the target capacity, the estimated costs and the conditions of implementation, is validated and authorized, and the results of resource changes are checked. STAROPs are then combined with observations such as application readiness, back-end health check, error rate and response time to verify recovery. Successful resource changes do not mean that operations have resumed.
Three scenes, one main line.
Scope of application
ROS Agent is suited to infrastructure tasks that require resource planning, reliance on organization and multi-step implementation. Typical scenarios include the planning of cloud resources in line with operational objectives; the generation, interpretation, validation and revision of the ROS/terraform template; the creation, updating, querying and clearing of cloud resources; and the clarification of needs, adjustments to programmes and confirmation operations through multiple rounds of dialogue. The volume of resources is not the threshold of capacity.
For simple queries on individual identification of resources or questions and answers on knowledge of conventional cloud products, STAROPs can use their own capacity to handle them directly without additional assignment; tasks not related to infrastructure are outside the scope of this paper.
This paper presents and validates an integrated range of Alibaba Cloud resource management involving cloud products such as ECS, RDS, load balance, VPC, Redis, OSS. Specific implementable operations depend on the type of resources, service capacity and account number permission. In the case of resource changes such as creation, updating, deletion, Agent shows the scope and impact of the operation and confirms it in accordance with the authorization strategy request; execution is still subject to RAM permission.
Fast start.
- Login to STAROPs console (starops.console.aliyun.com/) to create a digital employee.
- Search and mount the alibacloud-ros-agent at the skills centre.
- Describe your deployment or barrier requirements in the dialogue box.
Actual time-consuming and results depend on the size of the resources in your account, the integrity of the monitored data and the options selected, and it is recommended that a non-production demand validation process be used first.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.