打破技术孤岛：融合 DevOps 与 MLOps，构建统一软件供应链
打破技术孤岛：融合 DevOps 与 MLOps，构建统一软件供应链
Breaking silos: unifying DevOps and MLOps into a unified software supply chain
siloed/ ˈsaɪloʊ / 孤岛式
作者：尤瓦尔・芬巴赫
As businesses realized the potential of artificial intelligence (AI), the race began to incorporate machine learning operations (MLOps) into their commercial strategies. But integrating machine learning (ML) into the real world proved challenging, and the vast gap between development and deployment was made clear. In fact, research from Gartner tells us 85% of AI and ML fail to reach production.
随着企业意识到人工智能（AI）蕴藏的巨大价值，各大企业纷纷争先将机器学习运维（MLOps）纳入自身商业战略。但将机器学习（ML）落地到真实业务场景困难重重，开发与部署之间巨大的鸿沟也暴露无遗。高德纳（Gartner）的调研数据显示，85% 的人工智能与机器学习项目无法成功上线投产。
In this piece, we’ll discuss the importance of blending DevOps best practices with MLOps, bridging the gap between traditional software development and ML to enhance an enterprise’s competitive edge and improve decision-making with data-driven insights. We’ll expose the challenges of separate DevOps and MLOps pipelines and outline a case for integration.
本文将探讨融合 DevOps 最佳实践与 MLOps 的重要意义，弥合传统软件开发与机器学习之间的断层，帮助企业提升核心竞争力，依托数据洞察优化业务决策。同时梳理 DevOps、MLOps 两套流水线割裂运行带来的各类痛点，并论证二者整合的必要性。
bridging the gap：弥合差距
competitive edge：竞争优势
expose/ ɪkˈspoʊz / v.暴露，使显露；使面临
一、两套流水线独立运行的各类痛点
1. 工作流集成效率低下
DevOps pipelines are designed to optimize the software development lifecycle (SDLC), focusing on continuous integration, continuous delivery (CI/CD), and operational reliability.
DevOps 流水线专为优化软件开发生命周期（SDLC）打造，核心聚焦持续集成、持续交付（CI/CD）与生产运行稳定性。
While there are certainly overlaps between the traditional SDLC and that of model development, MLOps pipelines involve unique stages like data preprocessing, model training, experimentation, and deployment, which require specialized tools and workflows. This distinct separation creates bottlenecks when integrating ML models into traditional software applications.
传统软件开发生命周期与模型研发流程虽存在重合环节，但 MLOps 流水线包含数据预处理、模型训练、实验迭代、模型部署等独有环节，需配套专用工具与工作流程。两套体系完全割裂，会导致机器学习模型嵌入传统软件应用时出现流程瓶颈。
For example, data scientists may work on Jupyter notebooks, while software engineers use CI/CD tools like Jenkins or GitLab CI. Integrating ML models into the overall application often requires a manual and error-prone process, as models need to be converted, validated, and deployed in a manner that fits within the existing DevOps framework.
举例来说，数据科学家使用 Jupyter Notebook 开展实验，软件工程师则使用 Jenkins、GitLab CI 等 CI/CD 工具。若要将机器学习模型集成至整体应用，往往依靠人工操作，极易出错 —— 模型需要完成格式转换、效果验证、适配现有 DevOps 框架后才能部署上线。
2. 工具与资源大量冗余
DevOps and MLOps have similar automation, versioning, and deployment goals, but they rely on separate tools and processes. DevOps commonly leverages tools such as Docker, Kubernetes, and Terraform, while MLOps may use ML-specific tools like MLflow, Kubeflow, and TensorFlow Serving.
DevOps 与 MLOps 在自动化、版本管理、部署交付上目标高度相似，却各自使用独立工具与流程。DevOps 常用 Docker、Kubernetes、Terraform 等工具；MLOps 则选用 MLflow、Kubeflow、TensorFlow Serving 等机器学习专用工具。
This lack of unified tooling means teams often duplicate efforts to achieve the same outcomes.
工具体系不统一，导致团队为实现相同目标重复投入人力。
For instance, versioning in DevOps is typically done using source control systems like Git, while MLOps may use additional versioning for datasets and models. This redundancy leads to unnecessary overhead in terms of infrastructure, management, and cost, as both teams need to maintain different systems for essentially similar purposes—version control, reproducibility, and tracking.
例如 DevOps 依靠 Git 等代码仓库完成版本管理，而 MLOps 还需额外搭建数据集、模型版本管理系统。两套体系重复建设，带来基础设施、运维管理、资金成本的额外负担；但二者底层诉求完全一致：版本管控、流程可复现、全链路追踪。
3. 团队缺乏协同合力
The lack of integration between DevOps and MLOps pipelines also creates silos between engineering, data science, and operations teams. These silos result in poor communication, misaligned objectives, and delayed deployments. Data scientists may struggle to get their models production-ready due to the absence of consistent collaboration with software engineers and DevOps.
DevOps 与 MLOps 流水线互不打通，会在研发、数据科学、运维团队之间形成信息孤岛。孤岛效应造成沟通不畅、团队目标脱节、交付延期。数据科学家因缺少与软件工程师、运维团队常态化协作，模型难以完成投产适配。
Moreover, because the ML models are not treated as standard software artefacts, they may bypass crucial steps of testing, security scanning, and quality assurance that are typical in a DevOps pipeline. This absence of consistency can lead to quality issues, unexpected model behavior in production, and a lack of trust between teams.
此外，机器学习模型未被视作标准软件制品，会跳过 DevOps 流水线中标准化的测试、安全扫描、质量校验关键环节。标准不统一极易引发线上模型异常、质量缺陷，还会造成跨团队互不信任。
bypass/ ˈbaɪpæs / v.绕过，避开；
artefact/ ˈɑːrtɪfækt / n.手工制品，人造物；
4. 部署受阻，迭代周期拉长
The disjointed state of DevOps and MLOps also affects deployment speed and flexibility. In a traditional DevOps setting, CI/CD ensures frequent and reliable software updates. However, with ML, model deployment requires retraining, validation, and sometimes even re-architecting the integration. This mismatch results in slower iteration cycles, as each pipeline operates independently, with distinct sets of validation checks and approvals.
DevOps 与 MLOps 割裂运行还会拖慢部署速度、降低交付灵活性。传统 DevOps 依靠 CI/CD 实现高频、稳定的软件版本更新；但机器学习模型部署需要重新训练、效果验证，有时还要重构集成逻辑。两套流水线独立运行、校验与审批流程互不通用，流程错配直接拉长迭代周期。
disjointed/ dɪsˈdʒɔɪntɪd / adj.脱节的
distinct/ dɪˈstɪŋkt / adj.不同的，有区别的
For instance, an engineering team might be ready to release a new feature, but if an updated ML model is needed, it might delay the release due to the separate MLOps workflow, which involves retraining and extensive testing. This leads to slower time-to-market for features that rely on machine learning components. Our State of the Union Report found organizations using our platform brought over 7 million new packages into their software supply chains in 2024, highlighting the scale and speed of development.
举例：研发团队已完成新功能开发准备发布，但如果配套机器学习模型需要更新，独立的 MLOps 流程要重新训练、全量测试，会直接拖慢整体上线节奏，导致依赖 AI 模型的业务功能推向市场的周期变长。我们发布的行业全景报告显示，2024 年使用本平台的企业累计向软件供应链导入超 700 万个软件包，足以体现当下研发交付的规模与速度。
extensive/ ɪkˈstensɪv / adj.广阔的；广泛的；巨大的，大量的
5. 难以保障全链路一致性与可追溯性
Having separate DevOps and MLOps configurations makes it difficult to maintain a consistent approach to versioning, auditing, and traceability across the entire software system. In a typical DevOps pipeline, code changes are tracked and easily audited. In contrast, ML models have additional complexities like training data, hyperparameters, and experimentation, which often reside in separate systems with different logging mechanisms.
DevOps、MLOps 两套独立配置，很难在整套软件体系中统一版本管理、审计、链路追溯标准。标准 DevOps 流水线可完整记录代码变更，审计便捷；但机器学习模型还关联训练数据集、超参数、实验记录等复杂资产，这类数据分散在独立系统，日志记录规范各不相同。
traceability/ ˌtreɪsəˈbɪləti / n.[统计] 可追溯性；
This lack of end-to-end traceability makes troubleshooting issues in production more complicated. For example, if a model behaves unexpectedly, tracking down whether the issue lies in the training data, model version, or a specific part of the codebase can become cumbersome without a unified pipeline.
缺少端到端可追溯能力，线上故障排查难度大幅提升。例如模型线上表现异常时，若无统一流水线，很难快速定位问题根源是训练数据、模型版本还是代码逻辑缺陷。
behave/ bɪˈheɪv / v.行为，表现
cumbersome/ ˈkʌmbərsəm / adj.笨重的；复杂的
track down 追踪、寻找
二、整合势在必行：为何要统一 DevOps 与 MLOps？
As you can see, maintaining siloed DevOps and MLOps pipelines results in inefficiencies, redundancies, and a lack of collaboration between teams, leading to slower releases and inconsistent practices. Integrating these pipelines into a single, cohesive Software Supply Chain would help address these challenges by bringing consistency, reducing redundant work, and fostering better cross-team collaboration.
综上，DevOps 与 MLOps 割裂运行会造成效率低下、资源冗余、团队协作断裂，最终交付缓慢、流程标准混乱。将两套流水线整合为一体化软件供应链，能够统一规范、减少重复工作、促进跨团队高效协作，从根源解决上述痛点。
siloed/ ˈsaɪloʊ / 孤岛式
cohesive/ koʊˈhiːsɪv / adj.团结的，有凝聚力的；使团结的，使凝聚的
foster/ ˈfɑːstər / v.促进，培养
DevOps 与 MLOps 拥有共同核心目标
DevOps and MLOps share the same overarching goals: rapid delivery, automation, and reliability. Although their areas of focus differ—DevOps concentrates on traditional software development while MLOps focuses on machine learning workflows—their core objectives align in the following ways:
DevOps 与 MLOps 的顶层目标完全一致：快速交付、流程自动化、生产稳定可靠。二者侧重点虽有区分（DevOps 面向传统软件开发，MLOps 面向机器学习全流程），但核心诉求高度契合：
overarching/ ˌoʊvərˈɑːrtʃɪŋ / adj.首要的；支配一切的
concentrate/ ˈkɑːns(ə)ntreɪt / v.集中注意力；使集中；聚焦于
1. 快速交付
Both DevOps and MLOps strive to enable frequent, iterative releases to accelerate time-to-market. DevOps achieves this through the continuous integration and delivery of code changes, while MLOps aims to expedite the cycle of model development, training, and deployment.
DevOps、MLOps 均追求高频迭代发布，缩短产品上市周期。DevOps 通过代码持续集成与交付实现快速迭代；MLOps 则致力于压缩模型研发、训练、部署全流程周期。
strive/ straɪv / v.努力，力争；
expedite/ ˈekspədaɪt / v.加速，促进
Rapid delivery in DevOps ensures that new software features are shipped as quickly as possible. Similarly, in MLOps, the ability to deliver updated models with improved accuracy or behaviour allows businesses to respond swiftly to changes in data or business needs.
DevOps 的快速交付保障软件新功能快速上线；MLOps 可快速迭代精度、表现更优的模型，帮助企业快速适配数据变化与业务需求。
Similarly/ ˈsɪmələrli / adv.相似地；同样
2. 自动化
Automation is central to both practices as it reduces manual intervention and minimises the potential for human error. DevOps automates testing, building, and deploying software to ensure consistency, efficiency, and reliability.
自动化是二者的核心基石，能够减少人工操作、降低人为失误。DevOps 自动化软件测试、构建、部署流程，保障交付统一、高效、稳定。
In MLOps, automation is equally crucial. Automating data ingestion, model training, hyperparameter tuning, and deployment allows data scientists to focus more on experimentation and improving model performance rather than dealing with repetitive tasks. Automation in MLOps also ensures reproducibility, which is critical for managing ML models in a production environment.
自动化对 MLOps 同样至关重要。自动化数据接入、模型训练、超参调优、部署流程，让数据科学家从重复工作中解放，专注实验优化模型效果；同时自动化保障流程可复现，这是线上模型运维的核心要求。
ingestion/ ɪnˈdʒestʃən / n.摄取；吸收；咽下；接入
repetitive/ rɪˈpetətɪv / adj.重复乏味的；多次重复的
3. 稳定可靠
Both DevOps and MLOps emphasize reliability in production. DevOps uses practices like automated testing, monitoring, and infrastructure as code to maintain software stability and mitigate downtime.
DevOps 与 MLOps 均重视线上运行稳定性。DevOps 依靠自动化测试、系统监控、基础设施即代码（IaC）保障软件稳定，减少服务中断。
emphasize/ ˈemfəsaɪz / v.强调，着重
mitigate/ ˈmɪtɪɡeɪt / v.<正式>减轻，缓和
MLOps aims to maintain the reliability of deployed models, ensuring that they perform as expected in changing environments. Practices such as model monitoring, automatic retraining, and drift detection are part of MLOps that ensure the ML system stays robust and reliable over time.
MLOps 聚焦保障上线模型持续稳定，确保模型在动态变化的数据环境中正常输出。模型监控、自动重训、数据 / 概念漂移检测等 MLOps 实践，可长期维持机器学习系统稳健运行。
三、将机器学习模型视作软件供应链标准制品
In traditional DevOps, the concept of treating all software components as artefacts such as binaries, libraries, and configuration files, is well-established. These artifacts are versioned, tested, and promoted through different environments (e.g., staging, production) as part of a cohesive software supply chain. Applying the same approach to ML models can significantly streamline workflows and improve cross-functional collaboration. Here are four key benefits of treating ML models as artifacts:
传统 DevOps 早已形成成熟理念：将二进制文件、依赖库、配置文件等所有软件组件统一视为制品。各类制品纳入统一版本管理、标准化测试，按流程流转至预发、生产等多环境，构成完整软件供应链。将这套逻辑复用至机器学习模型，可大幅简化工作流、提升跨职能团队协作效率。将模型标准化为软件制品，四大核心价值如下：
1. 统一全量资产视图
Treating ML models as artifacts means integrating them into the same systems used for other software components, such as artifact repositories and CI/CD pipelines. This approach allows models to be versioned, tracked, and managed in the same way as code, binaries, and configurations. A unified view of all artifacts creates consistency, enhances traceability, and makes it easier to maintain control over the entire software supply chain.
将模型视作制品，代表把模型纳入与代码、二进制包、配置文件相同的制品仓库、CI/CD 流水线体系统一管理。模型与代码、依赖包、配置文件采用同一套版本管控、追踪机制。全资产统一视图实现流程标准化、提升链路可追溯性，便于企业完整管控整条软件供应链。
For instance, versioning models alongside code means that when a new feature is released, the corresponding model version used for the feature is well-documented and reproducible. This reduces confusion, eliminates miscommunication, and allows teams to identify which versions of models and code work together seamlessly.
举例：模型与代码同步版本管理，发布新业务功能时，配套使用的模型版本会完整留存、可复现。避免版本混乱、跨团队信息偏差，团队可快速确认适配的代码与模型组合。
2. 全流程自动化提效
Integrating ML models into the larger software supply chain ensures that the automation benefits seen in DevOps extend to MLOps as well. By automating the processes of training, validating, and deploying models, ML artifacts can move through a series of automated steps—from data preprocessing to final deployment—similar to the CI/CD pipelines used in traditional software delivery.
将模型纳入整体软件供应链，可让 DevOps 成熟的自动化能力全面赋能 MLOps。自动化模型训练、验证、部署环节，机器学习制品可像传统软件一样，经由数据预处理直至最终部署的全自动化流水线流转。
This integration means that when software engineers push a code change that affects the ML model, the same CI/CD system can trigger retraining, validation, and deployment of the model. By leveraging the existing automation infrastructure, organizations can achieve end-to-end delivery that includes all components—software and models—without adding unnecessary manual steps.
整合后，软件工程师提交影响模型的代码变更时，同一套 CI/CD 系统可自动触发模型重训、校验与上线。企业复用现有自动化基础设施，即可实现代码、模型全组件端到端交付，无需新增大量人工操作。
3. 强化跨团队协同
A major challenge of maintaining separate DevOps and MLOps pipelines is the lack of cohesion between data science, engineering, and DevOps teams. Treating ML models as artifacts within the larger software supply chain fosters greater collaboration by standardizing processes and using shared tooling. When everyone uses the same infrastructure, communication improves, as there is a common understanding of how components move through development, testing, and deployment.
两套流水线独立运行的最大痛点是数据科学、研发、运维团队割裂。将模型纳入统一软件供应链、标准化流程、共享工具链，可大幅提升协同效率。全员共用同一套基础设施，统一理解资产从开发、测试到上线的流转规则，沟通成本显著降低。
cohesion/ koʊˈhiːʒ(ə)n / n.凝聚力，团结；内聚力；
standardize/ ˈstændərdaɪz / v.使标准化，使符合标准
For example, data scientists can focus on developing high-quality models without worrying about the nuances of deployment, as the integrated pipeline will automatically take care of packaging and releasing the model artifact. Engineers, on the other hand, can treat the model as a component of the broader application, version-controlled and tested just like other parts of the software. This shared perspective enables more efficient handoffs, reduces friction between teams, and ensures alignment on project goals.
举例：数据科学家只需专注打磨模型效果，无需关心部署细节，一体化流水线会自动完成模型打包与发布；研发工程师可将模型视作应用的标准组件，像管理其他代码一样做版本管控与测试。统一视角简化团队工作交接、减少协作矛盾，保障全体成员目标一致。
nuance/ ˈnuːɑːns / n.细微差别
friction/ ˈfrɪkʃ(ə)n / n.不和，分歧；摩擦；摩擦力
4. 完善合规、安全与治理能力
When models are treated as standard artifacts in the software supply chain, they can undergo the same security checks, compliance reviews, and governance protocols as other software components. DevSecOps principles—embedding security into every part of the software lifecycle—can now be extended to ML models, ensuring that they are verified, tested, and deployed in compliance with organizational security policies.
若模型成为标准化软件制品，即可和其他软件组件同步执行安全扫描、合规审核、治理流程。将安全贯穿软件全生命周期的 DevSecOps 理念，可同步落地至机器学习模型，保障模型上线前完成校验、测试，完全符合企业安全规范。
undergo/ ˌʌndərˈɡoʊ / v.经历，经受
embedding/ ɪmˈbedɪŋ / v.把（物体）嵌入；使（观点）深深植入
This is particularly important as models become increasingly integral to business operations. By ensuring that models are scanned for vulnerabilities, validated for quality, and governed for compliance, organizations can mitigate risks associated with deploying AI/ML in production environments.
当下 AI 模型已深度嵌入企业业务，这一点尤为关键。对模型开展漏洞扫描、质量校验、合规治理，能够降低 AI/ML 系统上线带来的各类业务风险。integral/ ˈɪntɪɡrəl / adj.必需的，必要的，不可或缺的
结语
Treating ML models as artifacts within the larger software supply chain transforms the traditional approach of separating DevOps and MLOps into a unified, cohesive process. This integration streamlines workflows by leveraging existing CI/CD pipelines for all artifacts, enhances collaboration by standardizing processes and infrastructure, and ensures that both code and models meet the same standards for quality, reliability, and security. As organizations race to deploy more software and models, we need holistic governance.
将机器学习模型纳入统一软件供应链、视作标准制品，能够彻底打破 DevOps 与 MLOps 割裂的传统模式，构建一体化完整流程。依托统一 CI/CD 流水线管理全部资产简化工作流，标准化流程与基础设施提升团队协作，同时让代码、模型遵循同一套质量、稳定性、安全标准。如今企业加速落地软件与 AI 模型，全局化治理体系必不可少。
holistic/ hoʊˈlɪstɪk / adj.整体的，全面的
Currently, only 60% of companies have full visibility into software provenance in production. By combining DevOps and MLOps into a single Software Supply Chain, organizations can better achieve their shared goals of rapid delivery, automation, and reliability, creating an efficient and secure environment for building, testing, and deploying the <u>entire spectrum of software</u>, from application code to machine learning models.
目前仅有 60% 的企业能够完整追溯线上软件资产来源。将 DevOps 与 MLOps 整合为统一软件供应链，企业能更好落地快速交付、自动化、稳定运行三大共同目标，搭建高效、安全的研发交付环境，覆盖从应用代码到机器学习模型的<u>全品类软件</u>资产的构建、测试与部署。
visibility/ ˌvɪzəˈbɪləti / n.能见度，能见距离；可见性，明显性
provenance/ ˈprɑːvənəns / n.出处，起源
spectrum/ ˈspektrəm / n.范围，幅度；光谱