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
Multimodal vision-language models pose new challenges for inference engines like SGLang. Profiling with py-spy revealed that the scheduler’s process_input_requests consumed about 13% of CPU time, with a quarter of that spent on repeated calls to torch.UntypedStorage._new_shared_cuda for CUDA IPC handle book-keeping. The solution was a simple Python dictionary cache for pool handles, eliminating redundant opens. On a Qwen2.5-VL-3B-Instruct benchmark with a single H100, throughput increased 16.2% from 22.2 to 25.7 requests per second, mean time-to-first-token dropped 13.2%, and mean time-per-output-token dropped 17.2%. The improvement is merged in SGLang v0.5.10 and applies to any multimodal model using SGLang’s CUDA IPC transport.
Blockchain gaming platform Gala struggled to ingest and query expanding volumes of telemetry and blockchain data on its Databricks-based data infrastructure, resulting in multi-minute query times and heavy engineering maintenance overhead. To solve these bottlenecks, the company migrated its data platform to ClickHouse on AWS, ingesting data through Kafka, Airbyte, Amazon S3, and Fivetran. Gala leveraged ClickHouse's indexing capabilities alongside saved queries and API endpoints to power company-wide Metabase business intelligence dashboards for non-technical teams. Following the migration completed in December 2024, data capacity expanded from 3 TB to 9 TB while query times dropped from minutes to sub-second on previously unoptimized tables. The switch also lowered initial infrastructure costs by 30 percent, with plans underway to implement ClickHouse ClickPipes for improved data processing pipelines.
Pinterest deployed a root-leaf architecture for online machine learning serving to decouple CPU-based feature preprocessing from GPU-based inference. Because the root cluster retrieved a union of features across models and fanned out the full set to every leaf, network bandwidth became a severe bottleneck. After initial tests with fbthrift LZ4 compression yielded a 20% bandwidth reduction at the cost of CPU and latency overhead, the team implemented Feature Trimmer. This mechanism uses exported model signatures as the source of truth to prune unused features on the root before transmission. By dispatching only required features per model, the system reduced network bandwidth, decreased latency, and enabled cluster rightsizing that saved over $4M annually.
Granite 4.1 is a family of dense decoder-only language models in 3B, 8B, and 30B sizes, trained on approximately 15T tokens. Its five-phase pre-training pipeline moves from broad web-scale data to progressively curated mixtures, increases math and code emphasis, and extends context from 4K to 512K tokens through staged long-context training. Supervised fine-tuning uses about 4.1M curated samples evaluated with LLM-as-Judge and rule-based checks, while reinforcement learning uses on-policy GRPO with DAPO loss. The 8B instruct model matches or surpasses the earlier Granite 4.0-H-Small 32B-A9B MoE model despite its simpler dense architecture and fewer parameters, and the Granite 4.1 models are released under Apache 2.0.
Evaluating query performance differences across ClickHouse versions previously required managing Docker containers or manually downloading historical binaries from GitHub releases. The clickhousectl command-line interface simplifies this process by downloading specific ClickHouse binaries, running multiple background server instances concurrently with automatic port assignment, and executing queries across local environments. Users can copy table schemas between local instances by piping raw output from SHOW CREATE TABLE queries and replicate datasets using ClickHouse's remote table function. In a benchmark comparing ClickHouse 25.12 and 26.3 on an S3-backed Parquet dataset, version 26.3 introduced a Parquet metadata cache that reduced repeat query times from roughly 9 seconds to 1–2 seconds. System tables like system.server_settings and system.metrics enable operators to inspect cache configurations, eviction policies, and memory consumption directly.
NVIDIA Nemotron 3 Nano Omni is introduced as an omni-modal understanding model for document analysis, multiple-image reasoning, automatic speech recognition, long audio-video understanding, agentic computer use, and general reasoning. It extends the Nemotron multimodal line from vision-language processing to text, image, video, and audio, targeting long, dense inputs such as 100-plus-page documents and mixed audio-visual evidence. Its unified encoder-projector-decoder combines a Nemotron 3 Nano 30B-A3B hybrid Mamba-Transformer MoE backbone with C-RADIOv4-H and Parakeet-TDT-0.6B-v2 encoders; dynamic visual resolution, Conv3D tubelets, and Efficient Video Sampling support multimodal context compression. The post reports leading or top results across document, video, audio-video, voice, and ASR benchmarks, alongside up to 9x higher throughput and 2.9x single-stream reasoning speed on multimodal use cases, and provides BF16, FP8, and NVFP4 checkpoints.
Stripe’s 25-million-line Ruby codebase needed a true autoformatter; before rubyfmt, no industry-wide Ruby formatter existed, and earlier tools had crashed on its files. rubyfmt began as a zero-configuration open-source project motivated by formatting disputes, Ruby’s slow startup, and the uncertainty faced by engineers arriving from other languages. To meet a 100ms target, it initially used a hand-authored C launcher and later moved formatting into Rust, with ripper parse trees first passing through JSON and eventually being deserialized directly from Ruby VALUE objects. Stripe adopted per-file opt-in, expanded coverage cautiously, and formatted the codebase on a Saturday in 2024; the excerpt reports that all 42-million-line Ruby codebase is now formatted. In 2025, migrating from ripper to Prism removed the need to link a Ruby VM and made rubyfmt smaller and dramatically faster.
ClickHouse announced several product updates and integrations designed to deepen native compatibility with Google Cloud. Migration of ClickHouse Cloud to Google Axion ARM-based processors demonstrated a 30 to 55 percent query speed increase and reduced data load times from 64 to 26 seconds on ClickBench benchmarks. An integration with Google Lakehouse Runtime Catalog enables querying Iceberg tables on Google Cloud Storage via the DataLakeCatalog engine without data movement. Furthermore, Bring Your Own Cloud reached general availability on Google Cloud Marketplace, and ClickPipes added an unordered mode using Pub/Sub notifications for GCS ingestion. Google Antigravity integrated ClickHouse through an MCP server to translate natural language into SQL and generate interactive React charts.
The article presents three web applications built around OpenAI’s open-source Privacy Filter, a PII detector that labels eight categories in one forward pass over a 128k-token context. Document Privacy Explorer extracts PDF or DOCX text and returns category-labeled spans, while Image Anonymizer combines Tesseract word boxes with character offsets to produce editable pixel redaction rectangles. SmartRedact Paste replaces detected spans with category placeholders, stores public and token-gated views, and evicts expired pastes in a daemon thread. Across the apps, gradio.Server’s @server.api routes model work through Gradio’s queue and gradio_client, while plain FastAPI routes serve custom HTML, views, and lightweight lookups. The implementations keep browser-side filtering, canvas edits, and PNG export out of server round-trips, and the paste service is about 200 lines including storage.
DeepSeek V4 is presented as an open model family aimed at long-running agentic workloads, with Pro and Flash MoE checkpoints offering 1M-token context windows. Its central architecture alternates Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), compressing KV entries by 4x and 128x respectively, while using FP8/BF16 storage and FP4 indexing. The post-training design preserves reasoning across user turns when tool calls are present, introduces a dedicated XML tool-call schema, and relies on DSec, a Rust sandbox platform for RL rollouts. At 1M tokens, Pro uses 27% of V3.2's single-token FLOPs and 10% of its KV cache, while Flash uses 10% and 7%; V4-Pro-Max reports 80.6% on SWE Verified and 0.59 MRCR 8-needle retrieval at 1M.
Recent releases of pg_clickhouse, a Postgres foreign data wrapper for ClickHouse, introduce query pushdown enhancements and streaming result handling. Version 0.1.10 adds pushdown for JSONB accessor operators and extraction functions by mapping them to ClickHouse sub-column syntax and toJSONString calls outside SELECT clauses. In version 0.2.0, date and time functions such as CURRENT_DATE, CURRENT_TIMESTAMP, and clock_timestamp() push down to ClickHouse while honoring the active Postgres session time zone and precision settings. To address memory pressure and out-of-memory risks from large foreign queries, the extension adds query result streaming to the HTTP driver with a default batch buffer of approximately 50MB. Container benchmarking against the NYC taxi dataset showed streaming reduced peak memory consumption from over 600 MiB to under 86 MiB.
ClickHouse introduced a redesigned full-text search powered by inverted indexes that feeds matching documents directly into its vectorized analytical engine for observability and log analytics workloads. To evaluate this capability, the open-source versions of ClickHouse and Elasticsearch were benchmarked on single-node AWS EC2 instances across synthetic OpenTelemetry log datasets scaling up to 50 billion rows. The benchmark evaluated nine representative query patterns combining text search with filtering, grouping, counting, and time-series aggregations under both cold and hot cache conditions. ClickHouse achieved a roughly fivefold smaller storage footprint on disk while running cold queries four to six times faster and hot queries approximately 1.7 to 2.6 times faster than Elasticsearch. Furthermore, ClickHouse ingested the 50-billion-row dataset in under four hours out of the box, whereas the tuned Elasticsearch setup required approximately five days.
Join ordering is difficult because relational optimizers choose among exponentially many plans and depend on potentially inaccurate cardinality estimates. Databricks and UPenn researchers built an LLM agent that uses structured outputs to generate valid join reorderings and executes candidates through a tool returning runtime and intermediate subplan sizes. It can test up to 50 iterations, while the evaluation tested 15 rollouts per query across 113 Join Order Benchmark queries after duplicating each dataset row ten times. With a frontier model, the best tested orders improved geometric-mean latency by 1.288 times and reduced P90 latency by 41% versus the standard Databricks optimizer. A five-way example links its gain to LIKE predicates that are exceptionally difficult for cardinality estimators.
Palantir describes the frontend work behind adding read-only translation mode to Dossier, a collaborative rich-text editor used in multilingual, restrictive environments where public translation services may be unavailable or disallowed. The design lets users choose source and target languages, view a complete document in translation, copy text, and add comments while routing requests through the in-house Language Model Service. Engineers extracted shared translation components, added a Redux state slice and per-element cache, chunked content by paragraphs and other blocks, and used Intersection Observer, batching, debouncing, and caching to translate visible content incrementally; edits invalidate affected translations and trigger retranslation. To preserve sentence-level quality, inline formatting is stripped before translation, so the current translated view is plain text. The resulting mode is available in every Dossier, translating content lazily as users scroll and updating translated blocks when collaborators edit the document.
ClickHouse Cloud transitioned its Google Cloud infrastructure from x86 N2D instances to Arm-based Google Axion C4A instances to resolve CPU throttling and performance ceilings under peak query loads. The migration leverages Axion's dedicated physical cores per vCPU, higher memory bandwidth, local SSDs, and improved thermal and clock frequency stability. In automated ClickBench testing across 100 million rows, every query in the 43-query benchmark improved by 30% to 55%, while data load times decreased by approximately half. A production analysis of 38 migrated customer services demonstrated an average 15% reduction in compute credits consumed per query across diverse production workloads. ClickHouse Cloud executed this rollout using automated validation checks, scheduled maintenance windows, and capacity coordination with Google Cloud.
Shopify built a tool-calling agent for Sidekick that converts natural-language requests into Shopify Flow workflows, addressing the cold-start problem by working backward from thousands of anonymized production workflows. The team generated synthetic queries and tool trajectories, fine-tuned Qwen3-32B, and changed Flow’s nested JSON DSL into an equivalent Python representation connected by a round-trip transpiler. Using Python improved syntactic correctness by 22 points and semantic correctness by 13 points on a 300-example benchmark, while matching production tool names, ordering, responses, and prompts proved essential to avoid degradation. Production evaluation exposed a 35% gap masked by benchmark parity, so Shopify built a weekly feedback loop that scores conversations, filters training data, identifies slice gaps, retrains, and deploys. The resulting agent serves most production traffic, runs 2.2 times faster at 68% lower cost, and outperforms closed models.
In large-scale ClickHouse Cloud deployments, replicas traditionally load entire primary and secondary indexes from object storage into working memory, creating massive redundant memory overhead as fleets expand. Index sharding resolves this bottleneck by distributing the index analysis phase across available replicas using consistent hashing over data parts. Instead of every node loading full index structures, each replica loads and analyzes only its assigned fraction of primary keys, bloom filters, text indexes, or vector indexes. The query initiator merges the resulting granule match ranges, allowing parallel replicas to stream data without any single machine retaining the whole index. This approach freed working memory for query execution and accelerated index analysis by up to 7.7x in benchmark tests on a 50 billion row table.
Cursor's desktop app faces disruptive crashes as usage and features such as subagents, instant grep, and browser use have expanded, with most failures caused by out-of-memory conditions. The team measures instability across Electron's multi-process architecture using telemetry, per-session and per-request crash rates, minidumps, stack traces, crash breadcrumbs, and dashboards that update within minutes. Its two-pronged debugging strategy combines feature-level experiments and oversize-message instrumentation with bottom-up analysis based on a crash watcher, Chrome DevTools Protocol (CDP) stacks, automated daily investigations, opt-in heap snapshots, and low-rate allocation profiling. Findings led to chunking large data, killswitches, lifecycle fixes for leaks, and process isolation for extensions, while prevention uses Bugbot rules, stress tests, garbage collection, performance tests, and automated rollbacks. OOM-per-session fell 80% from its late-February peak and OOM-per-request fell 73% since March 1.
QIMMA is an Arabic LLM leaderboard built to address fragmented evaluation, translation artifacts, weak benchmark quality checks, limited coverage, and reproducibility gaps. It treats benchmark validation as a prerequisite for reported scores rather than an afterthought. The suite consolidates 109 subsets from 14 source benchmarks into more than 52,000 samples across seven domains, with 99% native Arabic content, Arabic-adapted HumanEval+ and MBPP+ coding tasks, and six prompt templates. Before model evaluation, Qwen3-235B-A22B-Instruct and DeepSeek-V3-671B assess samples using a 10-point rubric, flagged cases receive native-speaker review, and code statements are refined without changing task identifiers, reference solutions, or tests. In the April 2026 rankings, Qwen/Qwen3.5-397B-A17B-FP8 ranks first, while Arabic-specialized models lead cultural and linguistic tasks and multilingual models lead coding.
ClickHouse announced an expanded strategic collaboration with Google Cloud introducing four core architectural milestones. First, native integration with Google Cloud Lakehouse allows direct sub-second queries across data lakes and ClickHouse deployments without ETL pipelines or data duplication. Second, the Bring Your Own Cloud deployment model lets organizations run managed ClickHouse Cloud inside their own Google Cloud Virtual Private Cloud to retain data sovereignty and network controls. Third, ClickHouse Cloud is migrating automatically to Google's custom Arm-based Axion processors, boosting query throughput and cost efficiency for analytical workloads without requiring application changes. Finally, a new ClickHouse MCP server integration connects ClickHouse with Google Antigravity to enhance developer and analyst workflows.
The best engineering reads, once a week.
