This article provides a blueprint for scaling LLM inference in production by integrating open-source engines like vLLM into existing microservice architectures. It highlights the trade-offs between raw performance and operational flexibility when managing high-scale GPU workloads.
By AI Platform’s Model Runtime team and Inference team
Most organizations consume LLMs through hosted APIs. Netflix went further — we run the full stack ourselves, from model deployment through inference, inside our existing production environment rather than a separate ML silo. Some of those decisions weren’t obvious, and a few revealed their trade-offs only under production load.
This post focuses on the choices where alternatives were seriously considered: engine selection, model packaging, API surface design, deployment strategy, and output constraints enforcement. The goal is to share not just what was built, but why — and what production revealed that the design phase didn’t anticipate.
Member-scale ML at Netflix is fronted by a unified JVM-based serving system that handles the end-to-end flow for downstream consumers: routing and A/B test logic, candidate generation, feature fetching, inference, post-processing, and logging at each stage. Both real-time and cached batch paths are supported. Figure 1 shows the two ways callers reach inference today: the gRPC path through this serving system and a direct HTTP path used by newer LLM-driven applications.
Where inference runs depends on the model. Small CPU models run in-process, avoiding remote-call overhead. Larger models need GPUs — the serving system handles pre- and post-processing locally but delegates inference to a remote service, Model Scoring Service (MSS). MSS is the shared inference backend, supporting XGBoost, TensorFlow, PyTorch, and LLMs behind a unified interface, with NVIDIA Triton Inference Server underneath managing model loading, batching, and GPU scheduling.
On top of Triton sits a Java control plane that handles deployment, versioning, health checking, autoscaling, and multi-region rollout. Model authors package their artifacts and configure the deployment; the control plane provisions GPU instances, configures Triton, and orchestrates zero-downtime upgrades.
Four decisions shape this platform — engine, packaging, API surface, and rollout — presented in dependency order, since each one constrains the next.
The platform was originally built on TensorRT-LLM, a performant inference engine at the time and already integrated with Triton — the compute backend in use within MSS.
By summer 2025, two things had shifted: open-source engines had largely closed the performance gap with specialized stacks, and our workload mix had broadened to include embedding generation, prefill-only inference for ranking and retrieval, autoregressive decoding, and custom models with non-trivial per-step constraint logic. We re-benchmarked against this mix and selected vLLM as our paved-path engine on operational fit:
With vLLM picked, the next decision was how to package models for it. Triton supports two ways, and the choice has significant implications for maintainability — specifically, how tightly model artifacts are coupled to frontend upgrades.
The vLLM backend is the architecturally correct default. Two things bit us in production:
With engine and packaging settled, the next question is how callers reach the system. A key design goal of our system was that LLM models should NOT be special snowflakes. Every model — XGBoost ensemble or large-scale LLMs — is scored via the same gRPC call, so we reuse the same client libraries, health checking, and deployment pipelines. Given that the OpenAI-compatible API interface has become the de facto interface for the LLM ecosystem — inference engines, orchestration frameworks, evaluation tools, and client libraries all speak it — so we expose the OpenAI-compatible API as an additional frontend alongside gRPC.
The payoff shows up in the experimentation-to-production path: graduating from a hosted model to a fine-tuned self-hosted one — for quality, latency, cost, or data privacy — is nearly seamless. Same API, minimal code changes.
Behind the API, the implementation reuses NVIDIA’s Triton OpenAI-compatible frontend. It starts an embedded Triton server, wraps it in a TritonLLMEngine that converts request schemas into Triton inference requests, and serves responses through FastAPI. KServe HTTP/gRPC frontends are enabled alongside, so the same Triton instance remains accessible to the Java control plane over gRPC. Adopting Triton’s frontend directly exposed one gap: response_format — accepted by the schema — was silently dropped before reaching vLLM, so that a caller requesting JSON output proceeded without guided decoding constraints and could receive malformed JSON with no error surfaced by the platform. We git-subtreed and patched the frontend to translate response_format into vLLM’s guided decoding parameters at request time.
With API surface and engine in place, the question that remains is how new versions roll out without dropping requests. GPU deployments take longer to bring up than CPU services, and the I/O schema may change between model versions — adding a coordination problem on top. The platform offers two strategies:
We recommend embedding variable configurations (e.g., tensor shapes) directly into the inference model to make it version-agnostic, so it can use the cheaper Red-Black path. Versioned is reserved for the rare cases where a breaking interface change is unavoidable.
Beyond those four decisions, two operational details are worth flagging — both hit production gaps the design phase didn’t anticipate.
Bringing a vLLM-on-Triton instance up involves several coordinated steps before the gRPC port opens. Two are non-routine.
The rest of the boot sequence is mechanical: extracting the model package, installing custom vLLM plugins via Python entry_points, cleaning the Prometheus multiprocess directory, and gating the gRPC port until the engine is ready.
The Prometheus cleanup above hints at a wider observability gap. vLLM writes metrics to PROMETHEUS_MULTIPROC_DIR as .db files; Triton reports server-level metrics through its own Prometheus endpoint. Neither is aware of the other, and Triton’s built-in bridge surfaces only 9 of 40+ vLLM metrics — missing critical ones like token throughput, KV cache utilization, and prefix cache hit rates.
We added a lightweight HTTP proxy that merges both into a single /metrics endpoint: it fetches Triton metrics via HTTP, reads vLLM metrics from disk using Prometheus’s MultiProcessCollector, and returns the combined output. Existing dashboards and alerts work without modification.
Some Netflix production workloads rely heavily on fine-grained control over token generation. Rather than applying business logic after inference — paying for invalid generations, then retrying or repairing — we push constraints inside the decode loop, so the model generates outputs that are compliant by construction. We implement this via vLLM’s custom logits processor interface, modeling each constraint as a state machine that evolves with the generated token history and emits token-eligibility masks at each step. Each request gets its own configured processor, since different requests apply different rules.
Getting this to scale ran across two engine versions: we initially deployed on vLLM V0 (V1 had feature gaps), then migrated to V1 in Q4 2025 once it matured. The two subsections that follow are the before-and-after.
Our initial pure-Python implementation worked functionally but hit a scaling bottleneck. In vLLM V0, custom logits processors run per-request: the GPU produces logits for the whole batch, the CPU copies them across and waits for the transfer, and then constraint logic runs sequentially for each request — sequentially because the GIL prevents Python from parallelizing the per-request work. CPU time in logit processing therefore grows linearly with batch size, hitting tail latencies. End-to-end latency becomes CPU-bound even though the model’s forward pass is batched efficiently on GPU. It’s a bottleneck invisible in single-request benchmarks that only surfaces under realistic concurrency. Figure 2 makes the serial pattern visible.
The structural fix arrived in vLLM V1, which moved logits processing to batch level. We rewrote our custom processor to operate on batch-level data structures, computing masks across many requests together, and reimplemented the hot path in C++ with multi-threading to step around the GIL. The V1 API requires explicit tracking of batch membership changes via update_state(batch_update) — more complex than V0’s per-request interface, but necessary to maintain correct state in a dynamically evolving batch. Figure 3 shows logits processing time staying flat as batch size grows.
Now, performance was no longer the bottleneck. But stateful constraint logic in the decode loop introduced two issues the design phase didn’t anticipate:
We set out to build an LLM serving platform for broad production ML requirements — low latency, deep customization, and integration with existing infrastructure. The result is a system on vLLM and Triton, unified behind a consistent API, designed to give ML practitioners a fast path from experimentation to production.
The lessons were often in the details — version pinning, silent API gaps, packaging trade-offs — but addressing them has made the platform meaningfully more robust and the developer experience smoother. Next investments reflect where we expect friction:
We’ll continue working closely with the open-source community as this space evolves.
This system is the result of close collaboration and contributions from many teams within the AI Platform org at Netflix. In particular, Liping Peng designed and developed the model packaging workflow and drove the integration of Triton and vLLM with MSS to enable a unified pathway for serving LLMs. Hakan Baba, Nicolas Hortiguera, and ZQ Zhang led GPU capacity planning, system performance tuning, application integration and observability, as well as A/B test readiness and operational excellence efforts for all production models. Santino Ramos enabled vLLM for production models and optimized constrained decoding performance. Binh Tang developed the initial version of custom model serving and benchmarked different LLM serving frameworks. Lanxi Huang and Daneo Zhang built the serving development tools to enable user self-service. Lingyi Liu drove the overall system architecture and core technical decisions. Abhishek Agrawal and Shaojing Li provide management leadership to ensure alignment, prioritization and execution.
This work heavily leverages open-source ML libraries, such as Triton, vLLM and PyTorch, etc. We’re especially grateful to the teams and contributors from the community. We also thank our partner teams in Netflix AI for Member Systems for their close collaborations and innovation on the modeling side.
In-House LLM Serving at Netflix was originally published in Netflix TechBlog on Medium, where people are continuing the conversation by highlighting and responding to this story.
Continue reading on the original blog to support the author
Read full articleThis article demonstrates how to build a scalable ML platform that decouples model innovation from client applications. It provides a blueprint for managing complex routing, A/B testing, and high-throughput inference (1M+ RPS) in a distributed microservices environment.
Understanding service dependencies is critical for incident response. This architecture demonstrates how to process massive telemetry data in real-time using backpressure and local aggregation to provide actionable observability without overwhelming storage systems.
Netflix demonstrates that generative transformers can replace complex recommendation stacks. This approach simplifies architecture, reduces maintenance, and enables whole-page optimization through RL, leading to better user engagement and lower serving latency.
This migration shows how to replace complex homegrown infrastructure with standardized Kubernetes-native tools like Kueue. It enables advanced batch features like preemption and fair sharing at scale while maintaining a transparent experience for developers.
