Every post, summarised

Sep 10, 2026

• Stack Overflow

  (Re)introducing Developer Story

  Stack Overflow is bringing back Developer Story, a profile feature that records a developer’s career and presents expertise alongside contributions. The relaunch introduces “specialties,” areas of proven expertise that users can choose to highlight, with planned integrations to show verified contributions from other sites and Stack Overflow’s network. The feature is intended to become the center of a fuller technical identity and the foundation of “Stack Identity,” described as a verified developer proof-of-work layer for showcasing skills and career. The release is privacy-centric, excludes the old Developer Story data because it was deleted for privacy compliance, and is being developed through frequent changes, including improvements to specialties and integrations.

• NVIDIA

  How Full-Stack NIM Optimizations Deliver 2.5x More Users on Nemotron 3 Ultra

  Production serving for agentic AI must maximize concurrent users on fixed GPU infrastructure while preserving latency, especially with long prompts, reused context, and streamed responses. The article presents NVIDIA NIM as a deployable microservice that packages validated model- and GPU-aware choices, then measures Nemotron 3 Ultra NIM 2.0.12 against an open-source baseline on four B200 GPUs using a 64K/400 workload, 76% KV reuse, and 50 TPS per user (20 ms ITL). The optimized stack combines autotuned MoE and Mamba kernels, tensor and expert-aware parallelism, prefix and Mamba state-cache reuse, scheduling, batching, memory tuning, and MTP speculative decoding. At the stated target, throughput rises from 718 to 1,997 output tokens per second—more than 2.5x—while the article recommends replaying representative traffic with NVIDIA AIPerf to select a Pareto point meeting the application’s latency constraint.

• AWS

  Building resilient real-time streaming workers with Amazon DynamoDB leases

  Managing hundreds of persistent outbound WebSocket connections requires explicit ownership because a failed worker can drop connections and stop data flow until they are reassigned. The described fleet-management pattern uses Amazon DynamoDB leases, conditional writes, and a global secondary index to ensure one worker owns each connection, renew ownership with heartbeats, and find expired leases for reconciliation. Amazon ECS on AWS Fargate workers poll Amazon SQS for notifications, maintain the WebSockets, and release or transfer ownership during graceful shutdown, failures, and scaling. The design also records desired state, the upstream URL, and the last processed sequence number in the lease item, while clock synchronization supports its time-based expiry. The post identifies heartbeat writes as the dominant DynamoDB cost and recommends provisioned capacity with Auto Scaling for sustained high connection counts, while noting that longer intervals reduce cost but extend failover.

• Lyft

  Refreshing the Travel-Time Map Behind Lyft’s Marketplace: Rebuilding Neighborhood Reachability…

  Lyft’s Neighborhood Reachability Signals dataset supplies precomputed travel-time estimates between geohash-6 cells and a vocabulary of usable cells for pricing, driver heatmaps, and marketplace features. Production ETAs had remained largely based on 2018–2019 data, systematically underestimating travel times and omitting genuinely matchable pairs, while consumers maintained separate allowlists and denylists. The FORTOP team redesigned the Airflow-generated files by rebuilding the geohash universe from demand and supply history, estimating dense pairs from trip history, using Routing Simulation System (RSS) fallbacks for sparse pairs, and blending partial observations. In the second half of 2025, the refreshed dataset was adopted by Pricing after validation, with reported improvements in accuracy and completeness, a six-month refresh cadence, and plans for nine time-category files per region to represent traffic patterns.

• Clickhouse

  Introducing WalShadow: Sub-second Postgres replication to ClickHouse from physical WAL

  ClickHouse announces WalShadow, an open-source engine that replicates Postgres data to ClickHouse directly from physical WAL, targeting sub-second analytical availability without logical replication. It consumes the physical WAL stream outside the source database, tracks schema in a shadow Postgres instance, decodes heap records through parallel Rust-based decoders, builds ClickHouse-native blocks, and inserts them concurrently. Because parallel blocks can arrive out of order, each row carries the source WAL position (_lsn), while barriers preserve ordering for schema changes and truncations. In benchmarks using modest c8i.2xlarge instances, WalShadow reached approximately 200 ms commit-to-visible latency and 289,000 rows per second, compared with PeerDB’s roughly 10 seconds and 120,000 rows per second; it supports initial load, continuous replication, schema evolution, restart recovery, and planned switchovers.

• Thumbtack

  Three principles for building a vector platform at Thumbtack

  Thumbtack describes a vector platform that lets ML engineers stand up production vector search without separately negotiating database access, building ETL, or writing a query service. The design uses an existing Aurora Postgres cluster with pgvector, while Airflow jobs generate embeddings into BigQuery and managed reverse-ETL syncs them into vector tables; a query API validates inputs and returns ranked results. Three principles guide the layers: reuse trusted infrastructure, treat embeddings as ordinary data flowing through established pipelines, and lower adoption costs for subsequent teams. Production examples include semantic customer-support escalation using cosine similarity and an internal knowledge assistant combining vector retrieval, often with keyword search, before a language model responds; planned directions include self-serve provisioning, vector-backed feature retrieval, and framework-neutral agent memory.

• Clickhouse

  ClickHouse is a launch partner for the Data agent in ChatGPT Work

  ClickHouse is a launch partner for the Data agent in ChatGPT Work, with a plugin that connects ClickHouse Cloud to natural-language questions, reports, and interactive dashboards. The listing combines a fully managed ClickHouse Remote MCP server with open-source ClickHouse Agent Skills, and OAuth scopes access to the organizations and services visible to the connected user without requiring an API key. The MCP server exposes 13 read-only tools for metadata, SELECT queries, and service, backup, ClickPipes, and billing information, while the skills guide SQL and code around schema design, optimization, and ingestion. ChatGPT Work can use query results to build branded dashboards, with fresh answers coming from the ClickHouse Cloud service; availability depends on ChatGPT plan and workspace settings.

• Databricks

  Unifying governance across engines and catalogs in the Open Lakehouse

  Apache Iceberg’s REST Catalog additions—read restrictions and catalog labels—address different cross-platform governance problems: delegating enforcement to trusted engines and carrying governance context across federated catalogs. Read restrictions let a catalog evaluate policy for a principal and return bounded column-projection actions and row-filter expressions that a trusted engine must enforce, while leaving trust establishment outside the protocol. Catalog labels exchange opaque key-value metadata at table and column level, allowing a consuming catalog to map classifications locally, preserve native policy expressiveness, and maintain its own audit trail. Together with centralized enforcement through Cross-Engine ABAC, they provide distinct options for direct access and catalog federation, although read restrictions cannot express policies outside their standardized vocabulary.

• AWS

  Why don’t machine learning research agents overfit?

  Repeatedly optimizing against long-lived machine-learning benchmarks appears to invite overfitting, yet the post notes that gains on heavily reused benchmarks have largely transferred to newly created test sets. It describes a paper that resets LLM-based research agents, gives an explorer repeated access to validation feedback, and passes the resulting strategy through a token bottleneck to a fresh reproducer. Across eight datasets, strategies from hundreds of experiments were reproducible from prompts as short as 16 to 32 tokens; at shorter budgets, performance diverged, showing that the compressed message carried data-dependent choices rather than only prior knowledge. Deliberate overfitting occurred in 38 of 102 runs, with validation accuracy more than 10% above held-out accuracy, but those gains disappeared after compression; possible pretraining side channels remain unresolved.

• NVIDIA

  High-Throughput Structure Prediction with BioNeMo Inference Runtime

  Proteome-scale biomolecular structure prediction requires moving large worklists efficiently, and NVIDIA BioNeMo Inference Runtime (BioIR) accelerates supported models on NVIDIA GPUs while retaining a PyTorch workflow. Its end-to-end processor passes InputRequest objects through parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF writing; a Ray backend places complete model replicas on GPUs for concurrent independent inputs, while direct PyTorch integration supports custom code. In a matched 1,000-target human-dimer benchmark, BioIR completed all targets at 58.5K successfully folded residues per allocated GPU-hour versus 20.2K for the tested public implementation, a 2.90× improvement, which ran out of memory on 29 targets. Rated-power-equivalent estimates for one million comparable targets were 11 versus 35 MWh at 8-GPU TDP and 21 versus 64 MWh at full-node maximum power, excluding data-center overhead and several pipeline costs.

• Clickhouse

  Announcing On-Demand Compute: Instant compute for your most intensive workloads

  ClickHouse Cloud’s On-Demand Compute is announced in private preview as a way to run eligible queries on a shared pool of workers outside a service’s primary cluster, avoiding production contention and reactive autoscaling delays. It uses ClickHouse’s multistage distributed query execution and cost-based optimizer to distribute complex workloads, with queries requesting worker counts through settings including make_distributed_plan and distributed_plan_workers_num. Workers are leased for at least 60 seconds, reused during the lease, and released when it expires; concurrent queries share the pool, which grows to the largest request. In a stated benchmark, On-Demand Compute was faster for the tested workload and at 25 concurrent queries delivered similar performance while using 50% less compute, though results vary by query shape.

• Databricks

  Improving Lakebase Postgres compute cache

  Lakebase Postgres uses disaggregated storage backed by object storage, so compute-side caching must deliver low latency and high throughput while avoiding inefficient memory use. Its original design paired a conservatively sized PostgreSQL shared buffer with a local file cache (LFC), but the roughly 1 GB shared-buffer cap pushed many hits to slower NVMe storage and could create double buffering. For fixed-size computes with CU >= 80, the new configuration disables the LFC and sets shared buffers to 75% of DRAM, using explicit 2 MB HugeTLB pages to control page-table and TLB overhead. Benchmark tests found huge pages reduced tail read latency by up to ~40% and CPU use by up to ~30%, while production examples reported doubled throughput, 5× fewer storage reads, or 5× lower CPU use. Work is underway to dynamically resize shared buffers and huge pages for autoscaling computes.

• Red Hat

  From token consumer to token provider: Building your org's AI API

  Enterprise AI usage can become costly and fragmented when teams rely on separate provider accounts, personal keys, and independently operated inference stacks, leaving spending, workloads, and governance difficult to see. The post presents Model-as-a-Service (MaaS) in Red Hat OpenShift AI 3.4 as a shared service in which a platform team deploys approved models and exposes them through a standard, OpenAI-compatible API. Administrators can use the model catalog, distributed inference through llm-d on NVIDIA L40s, scoped API keys, rate limits, and dashboards for token use, latency, errors, GPU utilization, and queue length. Developers obtain self-service endpoints and can connect tools such as Claude Code, OpenCode, or Continue with existing clients. The architecture combines OpenShift, vLLM, an AI gateway, Prometheus, Grafana, and Jaeger to centralize inference, governance, and cost attribution across local and external model providers.

• Cloudflare

  1.1.1.1 now supports post-quantum DNSSEC, all 2,420 bytes of it

  1.1.1.1 now validates DNSSEC signatures made with ML-DSA-44, a NIST-standardized post-quantum algorithm, as an initial step toward post-quantum DNSSEC. The operational problem is that each signature is 2,420 bytes—almost 38 times an ECDSA P-256 signature and larger than common DNS-over-UDP limits—while zones must continue publishing conventional signatures for older resolvers. Cloudflare’s resolver handles this by testing larger responses at Internet scale, relying on truncation and transport retries rather than fragmented UDP, and using authenticated DS records to require a valid post-quantum path when a zone advertises one, preventing downgrade to conventional validation. The change provides operational experience for an ecosystem migration that requires coordination across DNS participants; Cloudflare says it plans full post-quantum security by 2029.

• Jetbrains

  Join Us at the Zephyr Project Meetup in Amsterdam

  The Zephyr community will meet in person at JetBrains Amsterdam on September 15, 2026, for a free evening focused on the open-source real-time operating system for resource-constrained and connected devices. The meetup brings together contributors, embedded developers, and others to exchange practical experience and discuss the ecosystem through talks on developer tooling, firmware testing, AI-assisted development, migration, and education. One session will show how AFL and the Renode simulator can fuzz Zephyr firmware without physical hardware, while another presents AI-assisted setup from environment configuration through compiling, flashing, and running an application on a board. The agenda also covers JetBrains’ collaboration with Zephyr and a migration from bare-metal and FreeRTOS code to a shared codebase targeting Renesas, STM32, and Nordic hardware. Registration is required, seats are limited, and presentations will be in English and Dutch.

• NVIDIA

  From Wafer-Out to First Token: Codifying Supply Chain Expertise with Nemotron and Palantir Foundry

  NVIDIA’s supply chain must compress time-to-rack, the interval from wafer-out to an assembled system on a data-center floor, amid millions of parts, thousands of suppliers, and changing component availability. The team built a Digital Supply Chain Intelligence command center in Palantir Foundry, whose Ontology connects materials, sites, commitments, capacity, allocations, outputs, and qualitative signals. NVIDIA cuOpt formulates weekly critical-material allocation as a mixed-integer linear program that minimizes Time of Ownership and exposes binding constraints, while planners can test tradeoffs around each solution. Historical back-testing showed that solver inputs missed expert context such as partner emails, weather, geopolitics, and supplier debriefs, so the workflow captures decisions, rationales, expected results, and outcomes for post-training. Post-trained Nemotron 3.5 Lightning reached 86.7% allocation-decision accuracy on the development benchmark, exceeding Nemotron 3 Ultra’s 55.5% and base Lightning’s 17.5%, though future production risk forecasting remained difficult.

• MongoDB

  Analyze MongoDB Atlas Data in OpenAI ChatGPT Work

  MongoDB Atlas is now a connected data source in OpenAI’s Data Plugin, powered by the MongoDB Atlas Managed MCP Server. This integration lets analysts and business teams query real-time operational data in ChatGPT Work without leaving the conversation, such as monitoring conversion changes after a deployment, finding orders stuck in payment pending, comparing sales hours across regions, and reviewing customer activity before support calls. Insights can be turned into live charts and dashboards within the conversation. Access respects existing cluster controls via OAuth sign-in, with organization owners able to enable or restrict AI client access to read-only, and every query runs with the permissions of the connected user. Teams that trust MongoDB Atlas as their fully managed, real-time system of record can now question their data directly without waiting for ETL jobs.

• Cursor

  Cursor Projects

  Cursor is launching Projects, a beta feature for taking on larger bodies of work such as features, migrations, and full applications. Each Project preserves context over months, runs on its own cloud computer through Cloud Agents, and coordinates parallel subagents rather than having the coordinator write code itself. Shared files synchronize across agents’ cloud and local machines, storing research, artifacts, codebase knowledge, and preferred working practices so later agents can use them. Projects can also subscribe to Slack channels, schedules, or pull requests and delegate work when signals arrive, while local agents handle testing that requires the user’s machine. Projects are accessible from Cursor’s left-hand navigation and are rolling out to all users in beta starting today.

• Together AI

  Introducing preemptible compute: the same compute, half the price

  Together GPU Clusters is introducing preemptible compute in public preview for Kubernetes clusters in all regions, offering interruptible GPU capacity at a flat 50% of the on-demand rate. Preemptible nodes use the same NVIDIA accelerated compute, can be reclaimed when unused capacity is needed, and receive a maximum five-minute drain window for checkpointing and graceful exit. Clusters retain the requested preemptible target and automatically refill as capacity returns, while usage is metered every one to two minutes and billed sub-hourly. The feature suits resumable, retryable, or requeueable work such as experiments, batch inference, and temporary bursts, but critical services, uncheckpointed multi-day runs, and strict-SLO serving should remain on standard nodes; users must schedule eligible pods with the together.ai/compute-class=preemptible label.

• Together AI

  To Infinity and Beyond: ThunderKittens Now on NVIDIA Vera Rubin NVL72!

  Together’s kernels team ported ThunderKittens to NVIDIA Vera Rubin NVL72 and investigated why existing Blackwell GEMMs underuse the platform. The work targets NVFP4 and FP8 GEMMs: although Rubin preserves Blackwell’s programming model, the original kernels reach only about 42.1% and 44.4% of roofline because tensor cores consume operands twice as fast as the old pipeline supplies them. The team integrates Rubin’s 64-byte K step, exclusive tensor memory extending to 576 columns, oversized shared memory up to 328 KiB, B-side collector buffers, and earlier A-release signaling into its NVFP4 kernel. The resulting kernel exceeds 22 PFLOPS and is described as competitive with cuBLAS and CuTe DSL, while the post also compares Rubin’s specifications with NVIDIA HGX B200.