Loading…
Topic
378 posts about Scalability. Every summary links to the original.
See the idea
Consolidating agent memory, retrieval, and governance onto MongoDB Atlas removes sync and vendor layers, but ties scaling and control to one platform. This trade-off view shows the payoff, the cost, and the condition that decides it.
Autoscaling trades responsiveness, stability, and cost: this explainer shows how signal choice, scaling direction, and capacity limits shape the right control loop for a workload.
Apache Pinot is a distributed real-time OLAP datastore for low-latency analytics. This architecture view shows how streams, processing, partitioning, indexed storage, and replica-aware serving fit together.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
NVIDIA released nvmath-python v1.0 to give Python users direct access to CUDA-X math libraries without requiring low-level C or C++ interfaces. The library acts as an abstraction layer over GPU engines like cuFFT and cuBLASLt, CPU backends like NVPL and Intel MKL, and distributed libraries such as cuBLASMp. It interfaces directly with NumPy, CuPy, and PyTorch tensors, structuring routines into generic interfaces for broad interoperability and specialized advanced modules for hardware-specific optimizations. Users can define custom sparse formats with Universal Sparse Tensors or invoke compiled device APIs directly inside custom numba-cuda kernels. Built-in logging captures execution space decisions and flags potentially expensive memory transfers across host and device environments.
Netflix developed GenRec, an LLM-backed recommendation ranker built by adapting an internal foundation model for large-scale personalization. Traditional recommendation stacks rely on thousands of hand-crafted features and specialized architectures that are costly to maintain across diverse content types. To replace manual feature pipelines, GenRec verbalizes user histories, metadata, and contexts into natural-language prompts and trains with multi-objective losses, including catalog-aware ranking and reward-weighted alignment. At inference time, the model executes in prefill-only mode on vLLM without decoding text. In large-scale online A/B testing against a mature production ranker, GenRec achieved statistically significant improvements in short-term and long-term metrics while using fewer labeled examples.
Enterprise AI is shifting from a primary focus on model capability to the utilization of scarce, expensive GPU capacity. Like aircraft, GPUs accrue financing, depreciation, power, and cooling costs by calendar hour, while producing value only during useful compute hours; unlike aircraft, they cannot freely absorb workloads with different memory, latency, and duration profiles. The article explains how peak-sized clusters, mixed workloads such as training, fine-tuning, quantization, real-time and batch inference, embeddings, and evaluation can leave capacity stranded even when dashboards show high occupancy, and argues that API pricing can motivate enterprises to own infrastructure. It presents specialized models and GPU orchestration as complementary levers: specialization reduces each workload’s footprint, while management reallocates freed capacity across workloads and queues. The conclusion is that installed capacity alone does not determine returns; enterprises need both strategies to close the gap between GPUs owned and useful output.
Shopify’s Application Security team built an agentic harness that scans software for vulnerabilities, proves candidates with real tests, and opens Shopify-specific remediation pull requests. Its Dispatch orchestrator partitions repositories, runs category-specific Hunting agents in parallel, verifies findings sequentially with a different model, and persists architecture artifacts for later diff-based scans. Test oracles require realistic integration or functional coverage, including cross-tenant fixtures for IDOR checks, while unproven candidates are rejected or downgraded. Across more than 80 applications and thousands of scans, the workflow produced over 300 findings, including two rated Critical by Shopify’s severity calculator, with full scans costing $50–$300 and incremental scans $5–$50. The post concludes that durable value lies in the harness’s verification, partitioning, cross-model review, and deterministic infrastructure rather than in any single model.
Databricks AI Search now offers generally available high-QPS scaling for Standard endpoints, addressing production workloads such as search bars, recommendations, and real-time entity resolution. Operators set a human-readable target_qps value when creating or updating an endpoint through the Python SDK, REST API, or UI, while Databricks provisions the required compute without manual replica counts, node sizing, or load balancers. Existing Unity Catalog governance and Delta Sync remain in place, and scaling progress is exposed through scaling_info as it moves from SCALING_CHANGE_IN_PROGRESS to SCALING_CHANGE_APPLIED; endpoint observability shows requests per second, latency, and health. Service principal authentication is recommended for high-QPS traffic, while personal access tokens are capped at a few tens of QPS, and new capacity applies when an index is created or synced.
AI prototypes rely on simple prompt-response interactions, but production AI systems handling thousands of records must manage distributed execution issues like worker crashes, rate limits, and transient failures. Rather than treating long-running operations as single large jobs or rebuilding state from logs, systems require durable workflows that preserve execution history across failures. In Agentforce Grid, execution is organized into a hierarchy where column runs divide into recoverable batches and rows composed of discrete activities. Each batch checkpoints after every successful row, retrying up to ten times with exponential backoff and resuming strictly at incomplete work to prevent duplicated model calls. Exposing progress across these hierarchical boundaries ensures that both users and engineers can distinguish completed, failing, and retryable units.
Online services and AI agents require low-latency point queries over massive datasets that are economically impractical to store entirely in key-value systems like Bigtable or DynamoDB. While cloud object storage offers rapid request latency, traditional distributed query engines like Trino and BigQuery introduce seconds of planning and scheduling overhead along with dependent file reads. Random Access Parquet addresses this limitation by using an external index mapping keys directly to exact Parquet files and row locations. Readers execute O(1) index lookups and issue parallel ranged reads to fetch only the required bytes without scanning footers or metadata. By running directly against existing data lake files, this architecture eliminates separate storage copies and allows shared analytical data to serve interactive online lookups.
Increasing artificial intelligence workloads and slowing geometric scaling require rapid semiconductor innovation across materials engineering and manufacturing. Applied Materials collaborated with NVIDIA to build an end-to-end digital development pipeline that accelerates atomic-scale modeling, process development, and factory optimization. The Ginestra defect-centric simulation platform integrates NVIDIA cuDSS sparse solvers to achieve up to a 10x speedup, while NVIDIA cuEST accelerates density functional theory calculations on B200 systems by approximately 55x compared to 64 CPU cores. In process engineering, the ACE+ multiphysics platform pairs with NVIDIA PhysicsNeMo to run topography simulations up to 35x faster, and NVIDIA Omniverse generates fab digital twins to optimize material flows. This unified digital workflow enables engineers to perform virtual experiments rapidly, mitigate production bottlenecks, and shorten transitions to high-volume manufacturing.
Salesforce needed to deliver major releases across 34 languages as product localization volume surged by more than 35% under fixed budgets and unchanged delivery windows. To address these scaling constraints, the engineering team replaced their legacy translation workflows with an LLM-powered orchestration pipeline integrating prompt engineering, context engineering, and multi-stage validation. The architecture transforms existing style guides, glossaries, and terminology databases into structured context, routing interface strings through specialized prompt stages—up to 85 stages for some languages—to enforce product context, grammar, and branding rules. Combined with an automated validation loop spanning AI editing, AI validation, and human review, this system reduced localization costs by 50 to 90% while accelerating turnaround times across enterprise releases.
Databricks built the Field Engineering Vending Machine (FEVM) to replace shared environments with isolated, centrally governed infrastructure for its growing field organization. Users or agents describe a use case and configuration details, receiving an environment assembled through a React frontend, Python backend, Databricks Apps, Terraform, Git Runner, and Lakebase across AWS, Azure, and GCP. FEVM supports templates and add-ons such as Lakebase, notebooks, and pre-packaged assets from UC Volumes, applies configurable time-to-live policies, and sends Slack notifications for provisioning, expiration, and deletion. It also manages shared resources independently, including catalogs that persist across workspace lifecycles, and provides administrative controls for limits and audits. At the time described, it managed more than 2,600 active deployments across three clouds and over 5,000 active users, after handling nearly 1,200 requests during one internal event.
Canva gateways check in-memory session revocations on every request, but pulling over a million revocations from MySQL during deployments created severe database stampedes. To eliminate this bottleneck, the team transitioned to storing a twelve-hour sliding window of revocations in Amazon S3 partitioned into 30-minute chunks. Each revocation is packed into a 16-byte binary structure within a sorted flat array, allowing gateways to query records via direct binary search without object deserialization. Asynchronous workers batch updates into S3 using ZooKeeper leader election and conditional PUT requests for optimistic concurrency control. This architecture reduced the in-memory cache footprint by 87.5% and allowed the team to reduce their MySQL database to just two read replicas.
Frontier AI pre-training increasingly relies on mixture of experts architectures, which lower compute per token while introducing severe all-to-all communication bottlenecks. To address these scaling limits, the NVIDIA GB300 NVL72 rack-scale system co-designs Blackwell Ultra GPUs, fifth-generation NVLink scale-up fabrics, and 800 Gbps SuperNIC scale-out networking. Testing on DeepSeek-V3 671B across 256 GPUs achieved a pre-training benchmark of 1,648 TFLOPs per GPU using Megatron Core. In addition to hardware co-design, ongoing software optimizations across Megatron Core, TorchTitan, and JAX contributed major throughput gains over time. Scaling from 256 to 1,024 GPUs sustained over 97% per-GPU performance efficiency across all evaluated frameworks by overlapping scale-out gradient traffic behind compute windows.
Agentic artificial intelligence workloads require sustained multistep inference, demanding low per-step latency, high decode throughput, and efficient long-context processing. NVIDIA designed the Rubin GPU to deliver up to ten times more agentic throughput per unit of energy than Blackwell. The hardware combines 336 billion transistors across two dies unified by NV-HBI, up to 288 GB of HBM4 memory providing 22 TB/s peak bandwidth, and the third-generation Transformer Engine reaching up to 50 petaflops of NVFP4 compute. Performance optimizations include inline descriptor updates for the Tensor Memory Accelerator, doubled K-dimension instruction throughput, 3-bit lookup-table weights, and structured 2:4 sparse attention compression. At the rack level, Vera Rubin NVL72 leverages Intelligent Power Smoothing and DSX MaxLPS to enable up to 40% more GPUs within a fixed power budget.
Scale-up networking has emerged as a critical architectural component for AI factories supporting trillion-parameter models, mixture-of-experts architectures, and disaggregated serving. NVIDIA presents its sixth-generation NVLink scale-up fabric and NVLink 6 Switch to address latency-sensitive all-to-all GPU communication and offload collective operations. In simulation benchmarks on models like DeepSeek-R1, Qwen 235B, and a simulated 2T parameter model, NVLink achieved up to 2.3X higher decode throughput than off-the-shelf Ethernet in a 72-accelerator domain. The architecture integrates NVLink-C2C to supply 1.8 TB/s of coherent CPU-GPU bandwidth in the Vera Rubin NVL72 platform, as well as NVLink Fusion for integrating semi-custom silicon XPUs. Through co-designed hardware, SHARP in-network compute, and NCCL libraries, the platform optimizes token throughput, factory uptime, and deployment economics.
Dropbox developed Riviera, a centralized content processing platform that evolved from an internal file preview generator into a shared infrastructure supporting over 300 file formats. By decomposing complex conversion tasks into modular, reusable transformations, the system now powers search indexing, media transcoding, and AI context extraction across Dropbox products.
Netflix established an in-house serving platform to run large language model inference directly inside existing production environments alongside traditional machine learning models. Built upon NVIDIA Triton Inference Server and vLLM, the unified architecture handles member-scale routing, candidate generation, feature fetching, and model execution over gRPC and OpenAI-compatible HTTP endpoints. The team shifted their primary engine from TensorRT-LLM to vLLM to support custom architectures, simpler debugging, and non-trivial constraint logic. Addressing production obstacles required patching Triton's frontend for guided decoding, pinning dependent library versions, rewriting logits processors in C++ for vLLM V1, and handling state machine resets during engine preemptions. The resulting platform unifies deployment flows while preserving operational stability across evolving model schemas.
Deploying Eclipse Dataspace Components (EDC) connectors on AWS presents challenges in forecasting infrastructure expenses without clear benchmarks. Estimated costs differ substantially between business-critical setups and non-critical environments based on compute and database sizing. A business-critical configuration relying on Amazon Aurora PostgreSQL db.r6g.large and standard Amazon ECS with AWS Fargate totals an estimated 387.00 USD monthly in the eu-south-2 region. Sizing down to db.t4g.medium and switching to AWS Fargate Spot lowers the non-critical monthly expense to 164.00 USD, representing a 58% reduction. Core cost drivers remain baseline infrastructure components, while storage, API calls, and data transfers contribute marginally under baseline assumptions.
Agentic artificial intelligence workflows transform data center infrastructure by turning context management, security enforcement, and inter-service communication into active components of the inference data path. To prevent host central processing units and graphics processing units from stalling on these tasks, NVIDIA introduces BlueField-4 data processing units, Vera BlueField-4 STX storage processors, and the DOCA software platform. The BlueField-4 processor offloads Layer 3 routing, encryption, and telemetry using a sixty-four-core Grace processor with up to eight hundred gigabits per second connectivity. Simultaneously, the Vera BlueField-4 STX powers the NVIDIA CMX context memory platform alongside DOCA Memos to store, protect, and share key-value cache across flash storage tiers. This co-designed infrastructure isolates tenant traffic and accelerates context reuse without degrading host-side agent execution.
Modal rebuilt its sandbox platform to support millions of concurrent sandboxes and tens of thousands of creations per second, removing all central bottlenecks from the control plane. The new architecture eliminates central coordination and durable datastores from the critical path, using horizontally scaled scheduling servers that run a fast load-balancing algorithm against in-memory cached worker state from Redis streams. Workers are contacted directly via RPC to create sandboxes, and sandbox metadata is written asynchronously, avoiding O(sandboxes) and O(nodes) operations that limited previous systems. In a benchmark, Modal created 1 million sandboxes in under a minute with median time-to-interactivity under half a second, though the long tail is attributed to kernel and network contention like the rtnl lock. The system is designed to scale without practical limits, supporting use cases like reinforcement learning and background agents that require massive concurrent sandbox creation.
Stripe describes how its Document Database fleet previously handled degraded MongoDB shards with single-step, hard-coded plugins that struggled with combined failures, intermediate states, and more than 40 shard layouts. The replacement models shard topology and health as a directed graph, with infrastructure operations as edges and safety invariants defining valid states. A simulator uses breadth-first search to find minimum-operation recovery paths, while an abstraction over Temporal contexts lets the same remediation logic run in memory without real infrastructure calls; later, the system also uses Dijkstra's algorithm. The system reduced pager volume by 30%—about 200 pages annually—reclaimed 12 days of fleet health each year, and handled a new layout with zero code changes, according to the post.
The best engineering reads, once a week.
