{"id":1489,"date":"2026-11-17T08:40:00","date_gmt":"2026-11-17T08:40:00","guid":{"rendered":"https:\/\/cotocus.cn\/blog\/?p=1489"},"modified":"2026-11-20T18:40:00","modified_gmt":"2026-11-20T18:40:00","slug":"mlops-faq","status":"publish","type":"post","link":"https:\/\/cotocus.cn\/blog\/mlops-faq\/","title":{"rendered":"MLOps FAQ: Fifteen Questions Engineers Keep Asking"},"content":{"rendered":"<p>When engineers and managers first meet MLOps, the same dozen questions arrive in nearly the same order. Instead of another lecture, this article answers them one by one, as a colleague might during a break in a workshop. The answers lean on the Ops Atlas entry for MLOps, and where a claim is the Atlas&#8217;s own, I say so.<\/p>\n<h2>What is MLOps in one sentence?<\/h2>\n<p>It is the engineering work of getting a trained machine-learning model into production and keeping it correct: versioning data, code, settings and environment, automating the training and release pipelines, watching for mismatch and drift, and retraining in a controlled way. It is not data science itself, and nobody sells it as a ready-made package.<\/p>\n<h2>Who defines it, and is there a standard?<\/h2>\n<p>There is no standards body publishing a normative definition, according to the Atlas. What exists is a mix of peer-reviewed papers, frameworks from the large cloud providers, and open-source projects hosted by the Cloud Native Computing Foundation. A continuous-delivery foundation described it in a 2024 roadmap as an extension of DevOps that brings machine-learning and data-science assets into continuous delivery. The entry even says that no reliable source records who first used the word, so be wary of anyone who claims to know.<\/p>\n<h2>How is it different from DevOps?<\/h2>\n<p>DevOps owns the culture and the software delivery pipeline. MLOps adds what plain software does not have: tests on data and models, continuous training, and release gates for models. The practical overlap is large, which is why teams with strong delivery habits usually pick MLOps up faster. If you are building a custom application and want a sense of the planning questions that come before any of this, our earlier piece on <a href=\"https:\/\/cotocus.cn\/blog\/questions-to-ask-before-building-a-custom-web-application-the-complete-guide\/\">the questions worth asking before a bespoke web application is built<\/a> sets out a comparable checklist for plain software.<\/p>\n<h2>What is the difference between MLOps and AIOps?<\/h2>\n<p>The direction of travel is opposite. AIOps applies machine learning to IT operations events, such as correlating alerts. MLOps is operations for machine-learning systems themselves. Related words, different jobs.<\/p>\n<h2>Do prompts and retrieval for language models count?<\/h2>\n<p>Per the Atlas, those belong to a sibling field often called LLMOps: prompts, retrieval-augmented generation, language-model evaluations and guardrails. Fine-tuning pipelines, however, remain inside MLOps. Similarly, portfolio-level governance of models, including risk tiers and independent validation, sits under ModelOps, while MLOps builds and operates the registry that governance relies on. Upstream ingestion, data contracts and general data quality are the territory of DataOps, though MLOps owns training snapshots, feature definitions and the checks that compare training with serving.<\/p>\n<h2>Why does everyone say &#8220;the model is a build artifact&#8221;?<\/h2>\n<p>Because it is a useful mental picture: a model is produced by a process, and the inputs to that process include the data. If you can name the code commit, the data fingerprint, the parameters and the environment, you can rebuild or at least explain a model. If you cannot, you are guessing. Note that training may be non-deterministic, so the standard is metrics within tolerance rather than identical bytes.<\/p>\n<h2>What does &#8220;train like you serve&#8221; mean?<\/h2>\n<p>Each input gets exactly one definition, and the identical code computes it during training and during live use. Keep a log of the values the running system actually received, then push a sample of them through both routes and compare. In the Atlas&#8217;s reference setup the disagreement rate from that comparison has to be exactly zero. If it is not, the pipelines are inconsistent, and the correct remedy is repair, not retraining.<\/p>\n<h2>How should a new model be approved?<\/h2>\n<p>With a gate. Freeze a test set split by time and compare the new model with the current one on the overall result, on important slices, on latency and on schema. Set the rules before you see the numbers. In the Atlas&#8217;s executed example the gate demanded at least 0.005 of improvement, allowed no slice to fall by more than 0.02, and capped the 95th-percentile latency at 50 milliseconds; a candidate with a good overall gain but a sagging slice was refused until it was fixed.<\/p>\n<h2>How do you release without gambling?<\/h2>\n<p>Stage it. Shadow mode lets the new model see live requests while only the existing model&#8217;s answers are returned. A canary then gives it a small fraction of traffic; for one popular serving framework the Atlas notes that canary support applies only in its serverless mode, so check your deployment style first. Finally, test against a business outcome, because a better offline metric does not guarantee a better result. A well-known travel-booking case is cited for exactly this lesson.<\/p>\n<h2>How do we tell skew, a data defect and drift apart?<\/h2>\n<p>By asking different questions in order. Do the training and live pipelines produce the same numbers from the same raw record? If not, you have skew. Did an upstream system change a format, drop a column or start sending nulls? Then you have a defect, to be fixed at its source. Only if both checks pass should you consider drift, and the Atlas splits that into three: a change in the inputs, a change in how often each outcome occurs, and a change in how inputs relate to outcomes. The remedies differ: retrain if fresh labels exist, adjust thresholds, or rethink the features and gather new labels.<\/p>\n<p>Be cautious with the distance measures too. A popular stability index depends on how the data is binned and on sample size, and its familiar rule-of-thumb bands are described in the entry as lacking academic support. Whatever measure you choose, calibrate it on your own history.<\/p>\n<h2>What should we monitor?<\/h2>\n<p>Think in three layers. First, is the service healthy? Second, is the incoming data sane? Third, are the predictions still any good? Quality is the hard one because outcomes arrive late; in a churn example the labels take about thirty days. Until they arrive, watch proxies such as the prediction distribution and the predicted rate against a matured base rate, and define action limits ahead of time.<\/p>\n<h2>When should we retrain?<\/h2>\n<p>The Atlas passes along five triggers named by one cloud provider: a manual request, a calendar schedule, the arrival of new training data, a measured fall in performance, and when the incoming data changes character in a significant way. To choose a cadence, run a staleness test: train up to a fixed date and measure how quality decays each week afterwards. And diagnose first. If training and serving disagree, fix the pipelines; if an upstream source broke, fix it there; only genuine drift justifies a retrain, and some kinds only need a threshold recalibrated.<\/p>\n<h2>Which tools should we look at?<\/h2>\n<p>The entry surveys many, and its status notes are time-stamped, which is a sensible habit. Two examples: the experiment-tracking and registry project MLflow has deprecated its older stage labels in favour of aliases and tags, and the workflow platform Kubeflow graduated within the CNCF in July 2026. Treat such facts as dated, because they will change. The sturdier advice is to choose tools after you have defined your process, not before.<\/p>\n<h2>What is the simplest sensible start?<\/h2>\n<p>Begin with one repeatable pipeline, a registry that remembers lineage, and a way to step back. The Atlas argues outright that such a trio is worth more than a grand platform that nobody opens. Rollback should happen by moving an alias or traffic split, never by retraining, and a plain non-ML fallback should stand behind it, rehearsed quarterly.<\/p>\n<h2>Is there a longer reference I can read?<\/h2>\n<p>The full entry, with its scope table, maturity levels and worked examples, lives at <a href=\"https:\/\/www.xopsschool.com\/mlops\/\">the Ops Atlas reference page on MLOps<\/a>, if you want every detail in one place. For a broader engineering view of how requirements get turned into dependable systems, see <a href=\"https:\/\/cotocus.cn\/blog\/deep-dive-the-complete-engineering-blueprint-for-web-project-requirements\/\">this deep dive into an engineering blueprint for web requirements<\/a>, which applies a comparable plan-first logic to site builds.<\/p>\n<p>If a question is missing from this list, it probably belongs to one of the neighbouring disciplines. Ask it anyway; the boundary questions are often the most useful ones to settle early.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Plain answers to the questions that come up first about MLOps: definitions, scope, release gates, monitoring, retraining, tools and the smallest sensible start.<\/p>\n","protected":false},"author":4,"featured_media":1490,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[331],"tags":[],"class_list":["post-1489","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-devsecops"],"_links":{"self":[{"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/posts\/1489","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/comments?post=1489"}],"version-history":[{"count":1,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/posts\/1489\/revisions"}],"predecessor-version":[{"id":1614,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/posts\/1489\/revisions\/1614"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/media\/1490"}],"wp:attachment":[{"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/media?parent=1489"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/categories?post=1489"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cotocus.cn\/blog\/wp-json\/wp\/v2\/tags?post=1489"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}