Loading…
Topic
812 posts about Performance. Every summary links to the original.
See the idea
Bare metal gives engineers direct control over GPU, network, and virtualization layers, but it shifts elasticity and lifecycle work onto the team. This trade-off map shows when that bargain pays off.
Bun packs a runtime, package manager, bundler, and test runner into one binary. Here is how those four parts fit together and where the seams still show.
A five-step how-to for deciding when an AMD Instinct GPU should stay dedicated and when SR-IOV virtual functions, VMs, or sidecar isolation can safely share it.
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
ClickHouse and Datadog announce a partnership aimed at reducing the tradeoff between retaining complete telemetry and keeping it accessible through familiar observability workflows. New Preview capabilities let organizations route logs directly to ClickHouse through Datadog Observability Pipelines and search those logs from the Datadog Log Explorer. Pipelines can parse, enrich, filter, transform, and redact data before storage, while ClickHouse Cloud’s separation of storage and compute supports long-term retention, independently scaled query resources, and workload isolation. Federated log search queries data directly in ClickHouse, avoiding duplicate copies or re-ingestion while allowing high-volume logs to remain searchable and historical investigations to use full-fidelity telemetry. The partnership combines Datadog’s collection and investigation experience with ClickHouse’s storage and analytics foundation; the native integration and federated search are available in Preview.
ClickHouse Cloud is now available through Stripe Projects, a Stripe CLI workflow that lets developers and AI agents provision real infrastructure from the command line. Running stripe projects init my-app and adding a service provisions ClickHouse or managed Postgres, returns working credentials to the environment through .env, and hands control to clickhousectl; clickhousectl greater than v0.3.0 is required. The integration supports ClickHouse for real-time analytics and managed Postgres, identified as a public beta, for transactional workloads; both services use the same CLI workflow, land in the same account, and share credentials, while Postgres uses local NVMe storage and is described as offering microsecond latency and up to 10x faster performance on I/O-heavy workloads. The feature is in developer preview, and agents can continue through clickhousectl for data ingestion with ClickPipes and service queries without manual credential handling or context switching.
amazon.science
Amazon announced the general availability of M9g and M9gd EC2 instances powered by the new Graviton5 processor. Built on a three-nanometer process, Graviton5 features 192 Neoverse V3 cores, DDR5-8800 memory support, PCIe gen6 interconnects, and 192 megabytes of level-three cache. The chip transitions from seven discrete dies to four unified chiplets connected at 420 gigabytes per second, eliminating dedicated I/O and memory controller dies. These architecture updates deliver up to 25% higher computational performance over Graviton4, with up to 35% faster performance for web applications and machine learning inference. The new instances also introduce the Nitro Isolation Engine, which uses formal verification to enforce hardware and software isolation between virtual machines.
Bugbot has been updated to run over three times faster, cost 22% less, and find 10% more bugs per review; 90% of runs now finish in under three minutes. The /review command lets users run Bugbot or Security Review before pushing code, while /review-bugbot and /review-security invoke those agents directly. Reviews run through /review sync with Bugbot on GitHub and GitLab, allowing a matching pull-request diff to skip a duplicate review, and Bugbot can now be configured to inspect only changes since its last review. The gains come from harness improvements and training Composer 2.5, which powers Bugbot; organizations opting out of that model use the next best available model, with speed and performance varying by configuration.
ClickHouse announced the public beta of ClickHouse Agents, a fully managed agentic analytics service in ClickHouse Cloud powered by Claude. It provides a no-code builder for agents grounded in live ClickHouse data, plus a managed chat experience in the Cloud console, so users need no SQL or separate hosting setup. Built on LibreChat, the service supports natural-language questions, SQL and visualizations, sandboxed Bash, Python, and JavaScript execution, shareable artifacts, skills, memories, MCP-compatible connections, and multi-agent workflows. Claude Sonnet and Haiku are pre-configured, while enterprise controls include SSO, encrypted message storage, and governed data access. ClickHouse says its internal deployment handles roughly 70% of data warehouse queries and processes more than 45 million tokens daily, confirming its view that real-time analytics is a natural context for high-volume agent workloads.
Qube Research & Technologies (QRT), a global quantitative investment manager headquartered in London, uses ClickHouse Cloud for a data platform that serves researchers and for near-real-time risk monitoring, management, and P&L. Founded in 2018 as a cloud-native firm after the spinout of Credit Suisse’s proprietary trading arm, QRT chose a cloud-based approach rather than inheriting decades of on-premise infrastructure. The platform supports two major systems, while ClickStack consolidates logs, metrics, and traces for observability; alerts, log queries, and metric aggregations run against a single ClickHouse instance using standard SQL. QRT says its previous database was constrained by writer and reader counts and total cluster load, whereas ClickHouse provides greater speed and scalability, and that a unified monitoring backend simplifies maintenance while remaining reliably fast.
Trackio's GitHub Actions CI was limited by generic hosted machines, maintenance-related delays, and lack of readily available GPU runners, while the project needed CPU checks and CUDA-based tests. The setup keeps GitHub Actions as the CI coordinator and uses a dispatcher Space plus the huggingface/jobs-actions bridge to launch ephemeral GitHub runners inside Hugging Face Jobs, selected through labels such as hf-jobs-cpu-upgrade and hf-jobs-t4-small. The guide covers duplicating and configuring the dispatcher, creating a GitHub App, supplying runner credentials and an HF token, changing runs-on, choosing Docker images, and validating runs and logs; Playwright and NVIDIA CUDA images are used for CPU and GPU examples. In Trackio, the CPU job took 1m10s versus a 1m40s GitHub ubuntu-latest baseline, while the GPU check passed in 45s and cost less than a cent at the t4-small rate; logs were available through both systems.
Palantir’s document store needs to repair or rebuild Elasticsearch indices when mappings, settings, cluster versions, or index state change, while continuing to serve live search and writes. Its reindex machinery reconstructs schema-aware indices from the strongly consistent database, rather than relying on Elasticsearch’s _reindex API, and is designed to limit manual work and resource consumption across many deployments and schemas. For an online reindex, it creates a physical shadow index with the intended mappings and settings, backfills it from the database, dual-writes current API changes to both indices, and promotes the shadow index after completion; searches remain on the primary during the rebuild. Cross-cluster migrations use a separate machine with cluster-specific metadata, configuration, index and alias evolution, and preservation of old-cluster indices to support rollback.
ClickHouse Cloud can automatically scale infrastructure up and down, but teams may need to monitor latency SLAs for specific workloads such as a frontend dashboard requiring p99 under 200 ms. The guide tags queries with log_comment, reads system.query_log through clickhousectl, and aggregates query_duration_ms over a defined window to measure the SLA. When it is breached, operators compare query volume and latency with Prometheus-compatible CPU, memory, and concurrency metrics to distinguish increased concurrency from resource contention, then choose horizontal replica scaling or vertical replica sizing. A cron-based script can scale out when p99 exceeds a threshold, while an agent can inspect evidence and apply one scaling action using clickhousectl; actions are recorded in the ClickHouse Cloud activity log.
Nemotron 3.5 Content Safety is NVIDIA’s 4B-parameter model for evaluating user prompts, optional images, and optional assistant responses together, with multilingual coverage and enterprise-specific policy enforcement. It extends Nemotron 3 by accepting natural-language custom policies, producing optional auditable reasoning traces, and supporting binary verdicts, category-labelled verdicts, or THINK mode; disabling THINK preserves the default low-latency path. Built on Google Gemma 3 4B IT with a LoRA adapter and a 128K context window, it uses the Aegis 2.0 taxonomy of 13 core categories and 10 subcategories, and targets deployment on GPUs with 8GB or more of VRAM. The accompanying multimodal, multilingual dataset includes concise reasoning traces, while reported results average 96.5% on Multilingual Aegis, 88.8% on RTP-LX, and about 85% across the evaluated benchmark set.
Real-time analytical databases must serve low-latency queries over data that is continuously inserted at high volume, rather than merely query a static dataset quickly. The article frames time-to-insight as the combination of time to ingest, time to transform and prepare data, and time to query it. It argues that batch-oriented ingestion, delayed visibility, scheduled materialized-view refreshes, and cache invalidation during writes can make production latency far worse than headline benchmarks. It recommends evaluating freshness under concurrent ingest and queries, whether one engine supports structured analytics, text search, vector similarity, and JSON, and whether index maintenance can avoid blocking ingestion. As an example, Bullet reportedly moved to ClickHouse Cloud, reducing data latency by 99.4% to under five seconds, achieving millisecond queries, and handling 1,000× more data at comparable cost.
Instructed-Retriever-1 is introduced as the retrieval-specialized model behind an update to Agent Bricks Knowledge Assistant, targeting search quality and latency. It parallelizes query and filter generation for broader candidate recall, then applies multi-pivot groupwise reranking to improve precision, using one model for both retrieval stages. The model was trained on synthetic enterprise-style environments and evaluated on realistic Knowledge Assistant workloads, where it scored 81.0 nDCG@10 for reranking and matched Claude Sonnet 4.5 retrieval quality on KARLBench. Serving uses a Mixture-of-Experts architecture, FP8 quantization, speculative decoding, and infrastructure tuning; the update reports more than 3x faster search, 2x faster answer generation, and TTFT around two seconds. End-to-end latency was consistently below 10 seconds in the offline evaluation setup.
Supabase has raised a $500M Series F at a $10B pre-money valuation, led by GIC. Existing investors participated, Stripe invested again, and Georgian and Salesforce Ventures joined as new investors; the funding will accelerate open-source and Postgres tooling, support growth, and provide employee liquidity. Supabase is also releasing Multigres v0.1 alpha, an open-source, self-hostable scalable operating system for Postgres designed to provide high availability and operational simplicity; a future release is intended to add Vitess-grade horizontal scaling. The alpha is ready to try but not production-ready, with production-readiness expected in a few months. Over the past year, database launches grew 600%, more than 60% of new databases came from AI tools, and nearly 10 million developers built on Supabase.
The post presents a redesign of hf, the Hugging Face Hub’s official CLI, for both human users and coding agents that download, upload, manage repositories, run Jobs, and administer other Hub resources. The CLI detects agents through environment variables, switches to compact, complete TSV output, and supports explicit human, agent, JSON, and quiet formats; agent mode was introduced in hf v1.9.0. Next-command hints, stderr separation, non-interactive failure messages, retry-safe operations, confirmations, and dry-run previews support multi-step automation without removing human-friendly terminal behavior. Benchmarks report equal-or-better success with roughly 1.3× to 1.8× fewer tokens than curl or the SDK on comparable tasks, and up to 6× fewer on complex workflows; the generated hf-cli skill reduced tool calls by about 30% without reducing token use or improving reliability.
ClickHouse announces Click-a-thon 2026, its first in-person hackathon in India, scheduled for Bengaluru on August 1–2, 2026. The event invites working tech professionals to form teams of two to four for a 24-hour build with four tracks—Real-Time Analytics, Observability, Data Warehousing, and Agentic AI and Analytics—and a ₹10,00,000 prize pool. Its stated purpose is to bring developers together to build practical systems under tight constraints and experiment with ClickHouse alongside modern open-source tools. The post points to ClickStack for observability, Langfuse for LLM tracing and evaluation, and LibreChat for model and agent interfaces, while describing combinations such as streaming ingestion, materialized views, projections, analytical queries, and tool-using agents over data. Applications close June 25, with shortlists planned for the first week of July; the team captain applies on behalf of the team.
Gaming platforms need sub-second session data for personalization, recommendations, content scheduling, and operational controls across millions of devices. The described pipeline uses Apache Spark Real-Time Mode with transformWithState to process Kafka events, maintain per-device session state, and emit scheduled heartbeats and timeout events without a separate streaming engine. Events are grouped by deviceId; a StatefulProcessor uses MapState keyed by session ID, while handleInputRows() handles starts and ends and handleExpiredTimer() registers and fires 30-second timers. The implementation was tested at about 500K input events per minute, 4M active sessions, and 8M heartbeat records, producing 16x output amplification. End-to-end latency reached 432 ms at p99, reported as 20x faster than micro-batch mode, and the source states that Real-Time Mode is generally available.
DharmaOCR applies Direct Preference Optimization (DPO) beyond chat alignment to reduce text degeneration in structured OCR, where outputs are either acceptable transcriptions or failure loops. In a benchmark of Brazilian Portuguese structured-document OCR, vanilla degeneration rates among open-source model families ranged from below 1% to above 33%, while supervised fine-tuning (SFT) did not consistently reach production-acceptable levels. After SFT, the pipeline generated multiple candidates for 23,726 documents, used an automated LLM judge to score them, and retained the model’s degenerate outputs as rejected examples paired with chosen outputs. DPO reduced degeneration in every tested family, by 59.4% on average and up to 87.6%, without reducing extraction quality according to the benchmark; applicability is presented as dependent on failures being categorically identifiable, reliably scoreable, and sufficiently numerous.
Ramp’s Stack is an AI-native accounting suite designed to automate repetitive book-closing work, including reconciliations, variance analysis, data entry, and schedules and accruals. To avoid overfitting to individual design partners, Ramp built a benchmark of synthetic business worlds, realistic accounting tasks, and accountant-written grading criteria, including standard and roll-forward worlds for testing memory transfer. The benchmark contains 237 tasks and 3,469 grading criteria in the analyzed slice, and supports repeated runs to compare models, prompts, tools, skills, harness changes, and memory behavior. Optimization included ablating skills, shrinking a spreadsheet skill from 14,000 to 5,000 characters, and tuning the system against end-to-end task performance rather than narrow evaluations. The resulting Stack system achieved the highest agent performance, with 4% higher accuracy and 3% better Pass@1, while remaining on the latency frontier with GPT 5.4; schedules and accruals were harder for raw models than variance analysis.
Over two years, ClickHouse focused on making joins faster for analytical SQL workloads, moving them from a performance gap to a first-class strength. The first year established the foundation with parallel hash joins, local automatic join reordering, filter pushdown, and execution improvements, producing a 4.4× speedup on the TPC-H SF100 join-heavy workload by version 25.4, compared with 22.4. The second year made optimizations effective by default, adding correlated-subquery support, lazy column replication, runtime filters, and statistics-based join reordering. Correlated queries can now run directly, while compact representations avoid copying repeated values and runtime filtering skips rows before hash-table lookups. Across the same benchmark and hardware, these changes delivered another 6× speedup, for a total 26× improvement; one six-table query fell from 3903.7 seconds to 2.7 while peak memory dropped from 99.1 GiB to 3.9 GiB.
Holo3.1 is a family of computer-use models released to improve robustness across web, desktop, and mobile environments, agent frameworks, and deployment targets. Based on Qwen, it adds native function-calling alongside structured JSON outputs, expands mobile automation, and offers 0.8B, 4B, 9B, and 35B-A3B sizes for different cost, latency, and performance needs. On AndroidWorld, the 35B-A3B model rises from 67% to 79.3%, while the 4B and 9B variants rise from 58% to 72%; function-calling and native execution achieve near-parity across the stated benchmarks, and Holo3.1 improves more than 25% over Holo3 in the Holotab harness. The release also provides 35B-A3B FP8, Q4 GGUF, and NVFP4 checkpoints for local inference: NVFP4 matches FP8 on OSWorld at about two points below BF16, reaches 1.74 times BF16 token throughput on DGX Spark, and with harness optimizations cuts average step time from 6.8 seconds to 3.3 seconds.
The best engineering reads, once a week.
