--- title: Reliable GPU compute for large training & inference jobs description: Netflix describes its layered approach to keeping GPU infrastructure reliable for large-scale training and inference, which power personalization and... canonical: https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Reliable GPU compute for large training & inference jobs | daily.dev og:description: Netflix describes its layered approach to keeping GPU infrastructure reliable for large-scale training and inference, which power personalization and... og:url: https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8 og:image: https://api.daily.dev/og/posts/PCELksbh8.png og:image:alt: Reliable GPU compute for large training & inference jobs og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Reliable GPU compute for large training & inference jobs **[Netflix TechBlog](https://daily.dev/sources/netflix)** · 15 min read · 0 upvotes · 0 comments ## Summary Netflix describes its layered approach to keeping GPU infrastructure reliable for large-scale training and inference, which power personalization and generative AI features. Basic cloud health checks and GPU-driver checks catch only obvious failures, missing degraded network adapters, latent memory faults, and tools that silently stop working. Netflix built node-level health checks on node-problem-detector and NVIDIA DCGM, with fail-closed behavior, dmesg fallbacks, a custom fatal-error map per GPU generation, and dependency checks that prevent silent no-ops. Faulty machines are automatically remediated through a graded response (kill job, reset GPU, reboot, or replace) using a customized Karpenter node auto-repair controller, and new machines are validated twice before accepting work. Jobs checkpoint and use elastic workers to survive machine loss, with failure attribution distinguishing platform from application causes. GPUs are shared safely via NVIDIA MIG partitioning, with plans to move to Kubernetes Dynamic Resource Allocation. Since these changes, Netflix reports over 99.95% GPU node availability and job success rates rising from about 98% to above 99.5%. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why do basic GPU driver checks fail to catch hardware problems in distributed training clusters? A GPU can respond to a status query and still be faulty once a workload stresses it, since faults in on-chip memory or firmware stay invisible until sustained compute or memory allocation triggers them. A network adapter can also be present but show fewer links than expected, letting collective operations run at a fraction of bandwidth without crashing, which silently slows the job instead. _Teams debugging flaky distributed training runs can track detailed GPU reliability engineering writeups on daily.dev._ ### How does Netflix decide whether to replace a faulty GPU machine or just reboot it? Netflix matches the remediation action to the fault type rather than defaulting to replacement, which is the slowest option. Killing the offending job clears a software wedge, resetting a single GPU clears many transient device faults, and rebooting clears drifted firmware state; full termination and replacement is reserved only for hardware that fails a trusted test, using an immutable hardware identifier to block repeat offenders from rejoining the pool. _Engineers designing graded remediation for GPU fleets can follow similar infrastructure patterns on daily.dev._ ### What GPU availability and job success rates has Netflix achieved with automated hardware remediation? Netflix reports consistently achieving over 99.95% availability of its GPU nodes, and job success rate improved from around 98% to above 99.5% after moving from manual draining and basic checks to layered health checks with automatic, graded remediation and checkpointed, elastic training jobs that survive machine loss. _Anyone benchmarking GPU reliability targets for ML infrastructure can find comparable case studies on daily.dev._ ## Similar posts on daily.dev - [How we keep GPUs reliable across Databricks AI](https://daily.dev/posts/how-we-keep-gpus-reliable-across-databricks-ai-um3yzqfzg) · databricks · 0 upvotes · 0 comments - [Keeping 20,000 GPUs healthy](https://daily.dev/posts/keeping-20-000-gpus-healthy-if9axbund) · Modal · 40 upvotes · 3 comments - [Self-healing GPU nodes in Kubernetes: What we learned building the EKS node monitoring agent](https://daily.dev/posts/self-healing-gpu-nodes-in-kubernetes-what-we-learned-building-the-eks-node-monitoring-agent-nivfgvnnn) · The New Stack · 2 upvotes · 0 comments - [Best GPU Optimization Tools for Kubernetes and AI Workloads \(2026\)](https://daily.dev/posts/best-gpu-optimization-tools-for-kubernetes-and-ai-workloads-2026--xv0s5w89i) · Cast AI · 0 upvotes · 0 comments - [New in Together GPU Clusters: Reliability and control for production GPU clusters](https://daily.dev/posts/new-in-together-gpu-clusters-reliability-and-control-for-production-gpu-clusters-azg0okgzb) · Together AI · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#kubernetes](https://daily.dev/tags/kubernetes) [View this post on daily.dev](https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Reliable GPU compute for large training & inference jobs","url":"https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8"},"datePublished":"2026-10-15T07:38:16.501Z","dateModified":"2026-10-15T07:38:42.386Z","description":"Netflix describes its layered approach to keeping GPU infrastructure reliable for large-scale training and inference, which power personalization and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ac99c1cc77ce93dde8b9bf1d5f381d05?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ac99c1cc77ce93dde8b9bf1d5f381d05?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Netflix TechBlog","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Netflix TechBlog","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/netflix","url":"https://daily.dev/sources/netflix"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,kubernetes","timeRequired":"PT15M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Netflix TechBlog","item":"https://daily.dev/sources/netflix"},{"@type":"ListItem","position":3,"name":"Reliable GPU compute for large training & inference jobs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/reliable-gpu-compute-for-large-training-inference-jobs-pcelksbh8#faq","mainEntity":[{"@type":"Question","name":"Why do basic GPU driver checks fail to catch hardware problems in distributed training clusters?","acceptedAnswer":{"@type":"Answer","text":"A GPU can respond to a status query and still be faulty once a workload stresses it, since faults in on-chip memory or firmware stay invisible until sustained compute or memory allocation triggers them. A network adapter can also be present but show fewer links than expected, letting collective operations run at a fraction of bandwidth without crashing, which silently slows the job instead. Teams debugging flaky distributed training runs can track detailed GPU reliability engineering writeups on daily.dev."}},{"@type":"Question","name":"How does Netflix decide whether to replace a faulty GPU machine or just reboot it?","acceptedAnswer":{"@type":"Answer","text":"Netflix matches the remediation action to the fault type rather than defaulting to replacement, which is the slowest option. Killing the offending job clears a software wedge, resetting a single GPU clears many transient device faults, and rebooting clears drifted firmware state; full termination and replacement is reserved only for hardware that fails a trusted test, using an immutable hardware identifier to block repeat offenders from rejoining the pool. Engineers designing graded remediation for GPU fleets can follow similar infrastructure patterns on daily.dev."}},{"@type":"Question","name":"What GPU availability and job success rates has Netflix achieved with automated hardware remediation?","acceptedAnswer":{"@type":"Answer","text":"Netflix reports consistently achieving over 99.95% availability of its GPU nodes, and job success rate improved from around 98% to above 99.5% after moving from manual draining and basic checks to layered health checks with automatic, graded remediation and checkpointed, elastic training jobs that survive machine loss. Anyone benchmarking GPU reliability targets for ML infrastructure can find comparable case studies on daily.dev."}}]} ```