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
Google DeepMind’s Gemma 4 family is available on Hugging Face as Apache 2-licensed, multimodal models intended for use across agents, inference engines, fine-tuning libraries, and on-device deployments. The family accepts image and text inputs across five sizes, while E2B, E4B, and 12B Unified also support audio; context windows range from 128k to 256k tokens. Its design combines alternating local and global attention, dual RoPE configurations, Per-Layer Embeddings, shared KV caches, and model-specific vision and audio handling, including an encoder-free unified 12B variant. The post documents integrations for Transformers, llama.cpp, MLX, WebGPU, Rust, and other tools, alongside fine-tuning examples and demos. Reported instruction-tuned results include a 1452 estimated text-only LMArena score for 31B and benchmark results spanning reasoning, coding, vision, audio, and long-context tasks.
PostgresBench is a public, reproducible benchmarking suite created to compare transactional performance across managed PostgreSQL providers. The benchmark uses the standard pgbench tool with a TPC-B-like write-intensive workload executing across 256 clients and 16 threads for ten-minute durations. Evaluators tested five services—including Postgres managed by ClickHouse, AWS Aurora, AWS RDS, Neon, and Crunchy Bridge—under 100 GB and 500 GB scale factors using default configurations in the us-east-2 region. Across tested configurations, Postgres managed by ClickHouse delivered higher transactions per second and lower P99 latencies due to physically co-located NVMe primary storage. These results demonstrate that remote network storage layers, such as EBS volumes, introduce significant latency overhead during sustained write and fsync operations.
Granular operations across cloud computing, payment platforms, and artificial intelligence workloads produce high-volume event streams that traditional business intelligence tools struggle to unify and aggregate. Operational BI platform Rill addresses this challenge by combining declarative data loading through the dlt framework with ClickHouse's high-performance analytical aggregation engine. Data transformations, SQL-based measure expressions, and dashboard configurations are defined entirely as code using SQL and YAML. This declarative, metrics-first architecture allows developers to validate analytics locally, track changes in Git, and eliminate intermediate caching layers when querying live data. By querying ClickHouse directly and linking AI-generated insights back to underlying queries, the system provides traceable, real-time operational analytics for enterprise environments.
Falcon Perception is a 0.6B-parameter early-fusion Transformer for open-vocabulary grounding and segmentation from natural-language prompts. It processes image patches, text, and task tokens in one sequence, using hybrid attention so image tokens receive bidirectional context while text and task tokens decode causally. Its Chain-of-Perception interface predicts each instance as coordinate, size, and segmentation tokens, with specialized Fourier-feature and dot-product heads for dense outputs. The post also introduces PBench, which profiles attributes, OCR-guided identification, spatial understanding, relations, and crowded scenes. Falcon Perception reaches 68.0 Macro-F1 on SA-Co versus 62.3 for SAM 3, while presence calibration remains weaker at MCC 0.64 versus 0.82; the 0.3B Falcon OCR model scores 80.3 on olmOCR and 88.6 on OmniDocBench.
Granite 4.0 3B Vision is a compact vision-language model designed to extract information from enterprise documents, including tables, charts, forms, and semantically meaningful key-value pairs. It combines a million-scale ChartNet dataset with a code-guided synthesis pipeline, DeepStack Injection for routing semantic and high-resolution visual features to different layers, and a LoRA adapter architecture built on Granite 4.0 Micro. The modular design supports standalone image processing, text-only fallback, and integration with Docling for multi-page document pipelines. Reported evaluations show 86.4% on Chart2Summary, leading the evaluated models, second place on Chart2CSV at 62.1%, leading scores on the listed table benchmarks, and 85.5% zero-shot exact-match accuracy on VAREX.
OpenMed describes an end-to-end protein engineering pipeline that connects ESMFold structure prediction, ProteinMPNN sequence design, and custom mRNA codon optimization. The codon-language-model study compares CodonBERT, ModernBERT, and several CodonRoBERTa variants using 250,000 E. coli coding sequences, a 69-token codon vocabulary, masked-language modeling, and training on four 80GB A100 GPUs. CodonRoBERTa-large-v2 achieved perplexity 4.10, CAI Spearman correlation 0.404, and 7.7% synonymous recovery, while ModernBERT-base reached perplexity 26.24 and correlation 0.070. The work then trained a multi-species suite on 381,283 sequences spanning 25 species, producing four production models in 55 GPU-hours and documenting planned extensions such as larger training data and additional prediction heads.
Orbax and MaxText introduce continuous checkpointing to balance training reliability and performance when hardware failures or preemption events interrupt long-running jobs. Fixed checkpoint intervals create a trade-off: infrequent saves risk substantial lost work, while frequent saves can block or bottleneck training, particularly over unstable networks. MaxText enables asynchronous checkpointing that attempts a new save only after the previous background request is fulfilled; its continuous mode overrides checkpoint_period, and max_num_checkpoints_to_keep can limit storage use. A llama-3.1-70B benchmark on two v5p-128 slices found smaller P50 checkpoint intervals, but average training step time increased because of more frequent device-to-host transfers; the analysis estimated resource conservation against mean-time-between-failure assumptions. Orbax also supports minimum intervals and customizable preservation policies, while co-locating storage with training is recommended because cross-metro bandwidth can reduce checkpointing speed and reliability.
TRL v1.0 marks a transition from research codebase to a library intended for production systems, with clearer stability expectations. It implements more than 75 post-training methods, but its design prioritizes trying, comparing, and using them over coverage alone. The library responds to a field in which PPO, preference optimization methods such as DPO, ORPO, and KTO, and RLVR-style approaches such as GRPO repeatedly changed which components of the training stack appear essential. TRL addresses that volatility with a stable core under semantic versioning, an experimental surface for faster-moving methods, and deliberately limited abstractions that favor explicit implementations over generic hierarchies. v1.0 also follows distributed breaking changes across the 0.x releases, so migration from the last 0.x version is described as minimal, while future work targets asynchronous GRPO, scaling, and actionable training diagnostics.
Lyft requires real-time predictions of whether a rider will request a ride after viewing price and ETA details to balance supply and demand and personalize user experiences. Traditional machine learning models overfit on high-cardinality, sparse context intersections, whereas deep neural networks introduce unacceptable inference latency during live user sessions. To address this, engineers created a hierarchical Bayesian tree framework that organizes session data into increasingly granular context partitions. Each node houses a parametric model trained top-down, applying Gaussian Bayesian priors and L2 penalties relative to parent parameters to smooth predictions in data-rare child segments. Using simple parametric models at each node also permits monotonic constraints, guaranteeing logically consistent predictions aligned with domain expectations.
Padlet needed a real-time analytics system to provide teachers with immediate insights into student engagement across 40 million monthly unique users. To avoid the operational overhead of managing database clusters with a small engineering team, the company adopted ClickHouse Cloud and ClickPipes for managed data ingestion. Active user sessions emit 30-second heartbeat events that stream alongside page views into ClickHouse. Materialized views pre-aggregate engagement duration and view counts by date, while native HyperLogLog sketches process unique visitor calculations efficiently across arbitrary time ranges. In production, this architecture ingested roughly eight billion events in a single month and delivered analytical query results with a 45-millisecond median latency.
ClickHouse delivers rapid analytical queries across massive datasets, but suboptimal schemas and queries can significantly reduce performance and storage efficiency. Aligning table definitions with ClickHouse storage and execution mechanics resolves these bottlenecks. Key practices include placing low-cardinality columns first in the ORDER BY clause to enable granule pruning, and selecting optimal data types while avoiding Nullable wrappers to cut memory overhead. Developers should treat partitioning as a data management tool for TTL drops or merge engines rather than a generic speed booster, preventing fragmented parts from slowing query execution. Recent improvements like the Analyzer query planner have also made joins viable, though dictionaries remain faster for static lookups.
ClickHouse Cloud released new monitoring capabilities and dashboards to give platform administrators deeper visibility into deployment health and server behavior. A newly introduced Overview page consolidates critical signals into a unified view for rapid health checks. Additionally, an updated Infrastructure page tracks service scaling over time, presenting CPU and memory utilization with new aggregation options alongside modals explaining automatic scaling decisions. The platform now sends automated email and Slack notifications when deployments encounter risks such as excessive parts, failed mutations, or query concurrency saturation. These additions allow administrators to proactively identify operational bottlenecks and configure scaling limits without spending excessive time searching for telemetry data.
Dropbox reduced its core server monorepo size from 87GB to 20GB after approaching GitHub's 100GB limit and suffering from hour-long clone times. The excessive bloat was traced to Git's delta compression heuristics interacting poorly with internationalization directory paths, generating inefficient diffs across different languages. By collaborating with GitHub to run an aggressive server-side repack with tuned window and depth settings, Dropbox cut clone times to under 15 minutes.
Lyft describes a framework for estimating long-term effects of pricing and incentive decisions in its two-sided marketplace, where direct and market-mediated consequences can diverge. The method first uses residualized observational regressions to estimate how policy changes alter distributions of negative experiences such as wait times, surge, cancellations, idleness, and hourly earnings, validating those responses with switch-back experiments. It then applies Augmented Inverse Probability Weighting to estimate how these experiences affect future outcomes, summarizes the relationship with a surrogacy index, and checks calibration using user-split experiments. Finally, Lyft combines mediated and separately estimated direct effects and uses region-split experiments for end-to-end validation, selecting treated and control regions with a forward-selection algorithm inspired by FDiD. The framework is intended to support scenario planning, budget allocation, and continuous calibration of long-run marketplace forecasts.
ClickHouse announced expanded data lake capabilities, enabling direct querying, accelerated ingestion into MergeTree, and writes across Apache Iceberg and Delta Lake tables. Open lake architectures provide storage portability across cloud providers and catalogs, but standard query engines often struggle with latency for real-time workloads. Over two years of development, ClickHouse engineered a native Parquet reader with page-level parallelism, integrated the Delta Rust Kernel, and introduced support for schema evolution, time travel, and DML operations. The platform functions across AWS Glue, Unity Catalog, Polaris, and Iceberg REST Catalog on storage systems including Amazon S3, Google Cloud Storage, and Azure Blob Storage. These enhancements allow organizations to execute federated, high-concurrency SQL queries across multiple catalogs or write processed analytical results back to their lakes without vendor lock-in.
Enterprise security defenders face shrinking exploitation windows alongside massive volumes of vulnerability findings across disconnected tools. To overcome query bottlenecks from traditional relational databases, Cogent Security designed a lakehouse serving architecture using Apache Iceberg and per-customer ClickHouse instances. The architecture relies on query-driven denormalized tables, custom compression codecs, and pre-computed projections to serve hundreds of millions of rows with sub-second latency. On top of this data layer, Cogent implemented an agentic text-to-SQL system that iteratively queries live data, references an in-house Ontology Service for semantic context, and applies majority-voted evaluations. This combination improved text-to-SQL accuracy from 40% to 94% while providing sub-three-second response times across 98.7% of queries at 100 million rows.
EVA is an end-to-end framework for evaluating conversational voice agents across task accuracy and spoken conversational experience. It simulates multi-turn live-audio conversations with a user simulator, the evaluated voice agent, deterministic tool execution, validation, and metrics derived from recordings, transcripts, and tool-call logs. The framework reports EVA-A for Accuracy and EVA-X for Experience, combines code-based and LLM- or LALM-as-Judge metrics, and uses pass@3 and pass^3 across three trials to capture peak performance and consistency. Its initial release contains 50 reproducible English-language airline scenarios and 15 tools covering rebooking, itinerary changes, cancellations, standby, and vouchers. Results for 20 cascade and audio-native systems show a consistent tradeoff: agents stronger at task completion tend to provide worse experiences, while a substantial gap between pass@3 and pass^3 indicates inconsistent success.
ClickHouse Cloud required a vertical auto-scaling mechanism that preserves rapid scale-ups while avoiding prolonged over-provisioning after traffic drops. Operating on a single 30-hour lookback window forced scale-downs to take up to 30 hours, whereas an earlier fixed-factor CPU algorithm caused severe oscillations. The engineering team deployed a dual-window recommender pairing a 3-hour window for responsive scale-downs with a 30-hour window for single-step scale-ups. Alongside a trend-based hunting check to resolve conflicting window recommendations, the team introduced target-tracking CPU scaling using geometric mean watermarks and 10-minute median smoothing. This dual-window architecture reduced scale-down latency to 3 hours, prevented scaling oscillations, and significantly reduced infrastructure costs for variable database workloads.
ClickHouse redesigned its native full-text index to address latency bottlenecks caused by random read patterns on remote object storage. Instead of using a Finite State Transducer dictionary, the new architecture organizes indexed tokens into fixed-size sorted blocks compressed with front-coding. Each data part separates index data across a small in-memory sparse index file, a sequential dictionary file, and a posting list file. This layout allows the query engine to pinpoint token locations with minimal random I/O and resolve multiple full-text functions directly from index files. Consequently, ClickHouse Cloud maintains predictable search latency across shared object storage while matching local disk performance and scaling through parallel query execution.
Agentic coding has revived grep as a critical way to gather source-code context, but ripgrep can stall on large monorepos because it scans every file. The proposed remedy is a local inverted index of overlapping character trigrams, with regular expressions parsed into required and alternative trigram constraints that narrow candidate files before exact matching. The index stores posting lists in one disk file and a sorted hash-to-offset lookup table in another, memory-mapping only the table for binary-search lookups; hash collisions can broaden results but cannot produce incorrect ones. Freshness is prioritized because agents often search for text they have just written, and the client index layers user and agent changes over a Git commit. The reported conclusion is that instant local search creates qualitative workflow improvements, especially for bug investigation and large Enterprise repositories.
The best engineering reads, once a week.
