Configuring Load Balancing Between GPU Instances

Imagine: you launch four GPU instances with vLLM, but 80% of requests go to the first server. The rest sit idle while users complain about timeouts. The reason — load balancing isn't configured. For LLMs this is critical: one long request of 4000 tokens can block a server for a minute, while the others remain idle. As a result, p99 latency skyrockets to 30 seconds, and GPU utilization drops to 25%. A typical cluster of 4 GPUs without balancing loses up to 50% throughput.

Proper load balancing reduces GPU infrastructure costs by up to 40% through uniform utilization. Average GPU-hour savings after implementation — 30% under the same load. P99 latency drops 1.7x compared to Round Robin. If you face similar issues, contact us — we will select the optimal configuration for your scenario.

Comparison of Balancing Algorithms for LLMs

Least Pending Tokens is the optimal choice for services with heterogeneous load. It uses Prometheus metrics from vLLM (vllm:num_requests_waiting) to select the least loaded instance. Our experience shows that Least Pending Tokens outperforms Round Robin by 1.7x in p99 latency.

Example: Nginx with Health Checks and Custom Balancer

Below is a basic Nginx configuration for an upstream of four vLLM servers, with active health checks and timeouts for streaming.

upstream vllm_cluster {
    least_conn;

    server 10.0.1.10:8000 max_fails=3 fail_timeout=30s weight=1;
    server 10.0.1.11:8000 max_fails=3 fail_timeout=30s weight=1;
    server 10.0.1.12:8000 max_fails=3 fail_timeout=30s weight=1;
    server 10.0.1.13:8000 max_fails=3 fail_timeout=30s weight=1;

    keepalive 100;
    keepalive_requests 1000;
    keepalive_timeout 60s;
}

server {
    listen 443 ssl http2;
    server_name llm-api.internal;

    location /v1/ {
        proxy_pass http://vllm_cluster;
        proxy_http_version 1.1;
        proxy_set_header Connection "";

        # Timeout для длинных streaming ответов
        proxy_read_timeout 600s;
        proxy_send_timeout 600s;
        proxy_connect_timeout 5s;

        # Streaming: отключаем буферизацию
        proxy_buffering off;
        proxy_cache off;
        chunked_transfer_encoding on;

        # Circuit breaker
        proxy_next_upstream error timeout http_500 http_502 http_503;
        proxy_next_upstream_tries 2;
        proxy_next_upstream_timeout 10s;
    }

    location /health {
        proxy_pass http://vllm_cluster/health;
    }
}

If a more intelligent backend selection is needed, we write a custom balancer in FastAPI that polls metrics in real time.

from fastapi import FastAPI, Request
import httpx
import asyncio

class LLMLeastPendingBalancer:
    def __init__(self, backends: list[str]):
        self.backends = {url: {"pending": 0, "healthy": True} for url in backends}
        self.client = httpx.AsyncClient(timeout=300)

    async def get_backend(self) -> str:
        """Выбираем backend с наименьшим числом pending токенов."""
        healthy = {url: info for url, info in self.backends.items() if info["healthy"]}
        if not healthy:
            raise RuntimeError("No healthy backends")

        metrics = await self._fetch_metrics(list(healthy.keys()))
        best = min(metrics.items(), key=lambda x: x[1].get("vllm_num_requests_waiting", 0))
        return best[0]

    async def _fetch_metrics(self, backends: list[str]) -> dict:
        tasks = [self._get_backend_queue(url) for url in backends]
        results = await asyncio.gather(*tasks, return_exceptions=True)
        return {url: result for url, result in zip(backends, results)
                if not isinstance(result, Exception)}

    async def _get_backend_queue(self, url: str) -> dict:
        response = await self.client.get(f"{url}/metrics")
        for line in response.text.split('\n'):
            if line.startswith('vllm:num_requests_waiting'):
                return {"vllm_num_requests_waiting": float(line.split()[-1])}
        return {"vllm_num_requests_waiting": 0}

    async def forward(self, request: Request) -> httpx.Response:
        backend = await self.get_backend()
        url = f"{backend}{request.url.path}"
        self.backends[backend]["pending"] += 1
        try:
            return await self.client.request(
                method=request.method,
                url=url,
                content=await request.body(),
                headers=dict(request.headers)
            )
        finally:
            self.backends[backend]["pending"] -= 1

app = FastAPI()
balancer = LLMLeastPendingBalancer(["http://gpu1:8000", "http://gpu2:8000", "http://gpu3:8000"])

@app.api_route("/v1/{path:path}", methods=["GET", "POST"])
async def proxy(path: str, request: Request):
    return await balancer.forward(request)

Why Sticky Sessions Are Critical for LLMs?

If your LLM uses KV cache prefix reuse (e.g., a common system prompt in a chatbot), without sticky sessions each request may land on a different server — the cache becomes useless. The solution — consistent hashing by prefix and sticky sessions.

def get_backend_by_prefix(prompt: str, backends: list[str]) -> str:
    prefix_hash = hashlib.md5(prompt[:256].encode()).hexdigest()
    idx = int(prefix_hash, 16) % len(backends)
    return backends[idx]

Applying sticky sessions increases cache hit ratio by 30-50%, reducing latency by 20%. Without them, a typical service with a shared system prompt loses up to 60% of cache efficiency.

Typical Mistakes in GPU Balancing

• Using Round Robin for LLMs — leads to uneven load.
• Lack of health checks — traffic goes to a failed server.
• Ignoring streaming timeouts — clients get 502 errors during long generations.
• Incorrect proxy_buffering configuration — increases latency.
• No GPU failover — all traffic is lost when one instance fails.

How to Set Up Health Checks for GPU Instances?

What's Included in Turnkey Load Balancing Configuration

1. Analysis of load scenarios (number of requests, token length, latency requirements).
2. Selection of algorithm and stack (Nginx, custom balancer, Envoy).
3. Configuration of health checks, circuit breaker, timeouts.
4. Implementation of sticky sessions (if KV cache is needed).
5. Integration with monitoring (Prometheus + Grafana dashboards).
6. Operational documentation and Incident Playbook.

Process of Work

• Analytics — collection of current infrastructure metrics, request profiling.
• Design — balancing architecture, algorithm selection, failover scheme.
• Implementation — deploying configs or writing custom module.
• Testing — load testing with p50/p99/p999 latency measurements.
• Deployment — phased rollout with canary release.

Timelines and Cost

Basic configuration on Nginx — from 1 day. Custom balancer with Least Pending Tokens support — from 3 to 5 days. Cost is calculated individually, based on infrastructure complexity and fault tolerance requirements. Guaranteed service stability after implementation — our engineers with 10+ years of experience in ML infrastructure deliver turnkey work. Typical ROI after implementation — 6 months.

Load Distribution Monitoring

After implementation, track: RPS distribution (should be uniform ±20%), queue depth on each backend, error rate, p99 latency. Set an alert: "one backend receives >80% traffic" — a sign of failure. With proper configuration, p99 latency drops to 5 seconds, and GPU utilization increases to 95%. Cache hit ratio reaches 70% with sticky sessions. We also train your team to work with dashboards.

Contact us for a preliminary audit — we will assess your current configuration and propose the optimal solution. Order a consultation — we will help you choose a balancing strategy for your GPU cluster.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.