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
Traditional logging solutions like Seq, ELK, and cloud platforms often face scaling bottlenecks, including slower queries, higher operational costs, and complex pipeline management. To address this, an ASP.NET service can write structured logs directly to ClickHouse using Serilog and the Serilog.Sinks.ClickHouse library. The sink transmits data over HTTP and provides a fluent C# builder to configure table schemas, typed JSON columns, full-text indexes, and MergeTree engine parameters. ClickHouse stores log data in a columnar format that delivers 10 to 20 times compression while enabling direct SQL queries, JSON property filtering, and real-time aggregations through materialized views. Furthermore, the sink includes an in-memory buffer to handle transient network drops with configurable batching, retry mechanisms, and callback hooks for failure alerting.
The tutorial presents a pipeline for adapting a general-purpose embedding model when it misses fine-grained distinctions in domain documents and labeled query–document data is unavailable. It generates synthetic question–answer pairs with NeMo Data Designer, filters them for quality, mines hard negatives with a 95% positive-score margin ceiling, and unrolls multi-hop queries for contrastive training. The recipe fine-tunes Llama-Nemotron-Embed-1B-v2, evaluates base and adapted models with BEIR metrics, and exports the result to ONNX or TensorRT. On the reported datasets, it improved Recall@10 and NDCG@10 by more than 10%; Atlassian increased Recall@60 from 0.751 to 0.951 on JIRA data. The complete workflow can run in roughly 2–3 hours for about 500 documents, using a single 80GB Ampere-or-newer GPU.
Cursor has made Composer 2 available as a coding model in Cursor. It is priced at $0.50/M input and $2.50/M output tokens, with a faster variant priced at $1.50/M input and $7.50/M output tokens. The release reports higher scores across CursorBench, Terminal-Bench 2.0, and SWE-bench Multilingual, with Composer 2 scoring 61.3, 61.7, and 73.7, respectively. Its training combines a first continued pretraining run with reinforcement learning on long-horizon coding tasks, allowing it to handle challenges requiring hundreds of actions. Composer 2's standard pricing and included usage pool on individual plans are presented alongside the faster option, which is set as the default.
The March 2026 ClickHouse newsletter details major platform updates, community deployments, and architectural practices for high-throughput analytics. Recent product announcements include ClickHouse 26.2 featuring time-based block flushing for streaming feeds, general availability of native inverted indexes for full-text search, and the release of chDB 4 with a Pandas-compatible DataStore API. Community implementations show Cloudflare managing 1.61 quadrillion daily events, while Trigger.dev isolated multi-tenant SQL execution using a domain-specific query language that compiles safe ClickHouse queries. In data optimization techniques, Hookdeck accelerated webhook payload search from over 30 seconds to under 400 milliseconds by partitioning payload hashes into typed bucket columns and scanning iterative time windows. Additional operational guidance emphasizes setting the ttl_only_drop_parts parameter and aligning table partitioning with time units so background merges drop entire data parts cleanly without triggering mutation rewrites.
ClickHouse delivers high performance for telemetry data, but unoptimized queries on high-cardinality observability workloads can bypass pruning, inflate intermediate state, and exhaust cluster resources. ClickStack addresses this challenge by embedding engine-specific optimization best practices directly into how queries are generated, structured, and executed. For search workflows, the platform executes progressive queries across bounded time windows using optimize_read_in_order to return initial rows rapidly without scanning entire historical ranges. ClickStack also segments large chart aggregations into parallel bucket-aligned queries, automatically redirects map attribute filters to pre-materialized columns, and extends lazy materialization thresholds to minimize runtime input-output overhead. These query formulation strategies are being packaged into specialized analytical endpoints to power notebooks and external tools without requiring manual SQL tuning or exposing raw database interfaces.
Dropbox Dash optimized its LLM-as-a-judge relevance scoring system using DSPy to migrate from expensive proprietary models to cheaper open-weight alternatives. By establishing automated feedback loops based on human agreement and strict JSON format validation, the team systematically generated robust prompts for new models. This reduced human-score disagreement by 45% and slashed model adaptation time from weeks to days while enabling 10x to 100x more data labeling.
Holotron-12B is a multimodal computer-use model released by H Company and available on Hugging Face. Post-trained from NVIDIA’s Nemotron-Nano-12B-v2-VL-BF16 on proprietary data, it targets agents that must perceive, decide, and act in interactive environments while handling long contexts with multiple images. Its hybrid State-Space Model and attention architecture stores a constant state per layer per generated sequence instead of attention’s per-token key and value activations, reducing memory demands for long-context inference. In a WebVoyager evaluation on one H100 with vLLM, it reached 8.9k tokens/s at concurrency 100 versus 5.1k for Holo2-8B, and its benchmark score rose from 35.1% to 80.5% after training on localization and navigation data.
Socialpruf, a social media performance tracking platform ingesting hundreds of posts per second, experienced analytics latency, connection dropouts, and high network transfer costs while running Postgres on Neon alongside ClickHouse Cloud. To resolve these issues, the team migrated their transactional system of record to NVMe-backed Postgres managed by ClickHouse, colocating transactional and analytical workloads. Although standard Postgres logical replication failed on their 0.5 TB dataset, they achieved continuous synchronization using PeerDB before conducting a production cutover. The migration eliminated connection dropouts, reduced network transfer fees, and improved overall Postgres query performance by approximately 30 percent. To manage thousands of concurrent connections generated by ingestion workers, Socialpruf implemented ClickHouse's bundled, horizontally scalable PgBouncer instances.
Composer is trained for long-horizon coding tasks using self-summarization, allowing reinforcement-learning trajectories to exceed the model’s context window. When a fixed token trigger is reached, Composer receives a synthetic summarization query, uses scratch space to produce a condensed context, and resumes with the summary plus conversation state; training rewards every token in the chained rollout with the final outcome. Against a tuned prompt-based compaction baseline, self-summary reduced compaction-related error by 50% on CursorBench at both 80k and 40k triggers, while summaries averaged about 1,000 tokens instead of more than 5,000 and reused the KV cache. An early Composer checkpoint solved Terminal-Bench 2.0’s make-doom-for-mips after 170 turns, summarizing more than 100,000 tokens into roughly 1,000 tokens of retained information.
The official ClickHouse Python client, clickhouse-connect, originally addressed asynchronous execution by wrapping its synchronous HTTP client in a thread pool executor. However, this executor model experienced thread exhaustion, GIL contention, and severe tail latency variance under heavy concurrent workloads. To resolve these scalability bottlenecks, the developers redesigned the client around a half-sync/half-async pattern utilizing aiohttp for async network I/O. A custom bounded queue bridges streaming socket reads on the event loop with synchronous, CPU-intensive binary parsing running inside background threads. Benchmarks demonstrate that this architecture stabilizes tail latencies, delivering an average P95 latency of 556 milliseconds compared to 869 milliseconds for the legacy executor client.
ClickHouse version 26.2 introduces twenty-five new features, forty-three performance optimizations, and one hundred eighty-three bug fixes across the database engine. This release graduates both the text-index and the QBit vector embedding data type to production-ready status for search and retrieval workloads. For streaming ingestion pipelines where row arrival rates are low, the engine adds time-based block flushing configurations to write data parts before hitting default row or byte thresholds. The update also embeds the ClickStack observability user interface directly into the standard binary distribution, accessible through the local HTTP server port. Further enhancements include automatic min-max indexing for temporal columns in MergeTree tables, interactive time-based one-time password authentication in the client, and a high-performance prime number generator function.
LLMs can generate syntactically valid ClickHouse schemas whose advanced choices are poorly matched to a workload. The post examines recurring pitfalls including query-driven partitioning, per-column codecs, premature projections, materialized-view sprawl, JSON configurations, and forced deduplication with OPTIMIZE TABLE ... FINAL. It explains that partitions primarily support data management, incremental materialized views add work to every insert, and ReplacingMergeTree deduplicates during background merges rather than at insertion. The recommended method is to begin with a simple MergeTree or ReplacingMergeTree, a well-chosen ORDER BY, default compression, and no optional structures, then measure production behavior before adding complexity or seeking human expertise for high-scale, interconnected designs.
ClickStack introduced several platform updates and query performance enhancements across both open-source and Cloud environments for February 2026. ClickHouse Cloud OpenAPI now exposes ClickStack resources including dashboards, alerts, sources, and webhooks to support automated configuration-as-code workflows. In addition, ClickHouse 26.2 embeds the ClickStack UI directly into the database binary with a 4.1 MB footprint for local exploration and diagnostics. Visualizations now support native SQL mode for tables, allowing users to join disparate datasets like logs and traces across injected dynamic time variables. Finally, query execution speeds improve through streaming skip indexes, index pruning on disjunctions, and an expanded 100,000-row lazy materialization limit alongside new threshold alerting on number charts.
Grab enabled R8 optimization for its Android app after repeated attempts were blocked by debugging difficulty, organization-wide testing demands, and slow build feedback. The effort targeted a roughly 10-year-old codebase exceeding 9 million lines and combined AI-assisted debugging, a pragmatic testing strategy, and infrastructure investment. Its QA build was configured to match production R8 processing. Beyond existing shrinking, optimization added techniques such as method inlining, class merging, constant folding, dead-code elimination, and devirtualization. After about 10 weeks with one engineer coordinating across teams, the project reported around 25% lower ANR rates, a 16% reduction in reference-device download size, and nearly 27% faster startup; the startup result combined R8 with corrected Baseline and Startup Profiles.
Shopify describes how high-cardinality, deeply nested GraphQL queries exposed execution overhead beyond data loading, with field resolvers assembling responses consuming much of request time. Its GraphQL Cardinal engine replaces conventional depth-first traversal with breadth-first execution, resolving each selected field once across an aggregated set of objects and reducing repeated field-level work. For large list queries, Shopify reports up to 15x faster execution, 90% less memory use, and reductions of many seconds in P50 latency. The migration included a tuned interpreter, cheaper tracer behavior, shadow verification, benchmarks, and ongoing conversion of tens of thousands of legacy resolvers, while future work includes asynchronous patterns and lower-level C bindings.
chDB 4 introduces Data Store, a native component that allows users to write Pandas-like code that executes on the ClickHouse OLAP SQL engine. Standard Pandas workflows suffer from eager execution, excessive memory materialization, and lack of automatic query optimizations. To solve this, Data Store implements a four-layer architecture that lazily records operations as a chain of LazyOp objects and compiles them into optimized ClickHouse SQL plans upon evaluation triggers. When encountering operations unsupported by ClickHouse, the QueryPlanner splits the pipeline into segments, routing tasks between ClickHouse and Pandas while using Python's memoryview to minimize data copying overhead. This release also integrates natively into Hex notebooks.
GitHub produces billions of events across issues, pull requests, and comments, but turning this vast stream into a real-time trends analyzer presents significant technical challenges. GitTrends is an open-source demo application designed to search and compare technology keywords across nearly ten billion GitHub events. Built to showcase the full-text search capabilities recently released in ClickHouse, the application indexes raw text without pre-computed answers or intermediate data transformation. ClickHouse executes full-text searches and aggregations within a single engine in one pass, eliminating cross-system joins and data movement. The demo includes live comparisons toggling between full-text search, bloom filters, and full table scans to illustrate query performance at scale.
ClickHouse has reached general availability for its native Full-text Search feature, designed to accelerate token-based text filtering across large datasets. The capability relies on deterministic inverted indexes rather than probabilistic Bloom filter skip indexes, mapping individual tokens directly to row numbers. It supports plain strings, string arrays, and map keys or values while allowing configurable pre-processing and tokenization using standard SQL expressions. The system does not implement relevance scoring or store positional data for ranking, focusing instead on rapid filtering paired with large-scale analytical aggregations. Testing shows significant reductions in scanned granules alongside substantial latency improvements, though text indexes require more storage than Bloom filters.
Hugging Face introduces Storage Buckets as a mutable, S3-like object storage layer on the Hub for intermediate machine-learning artifacts such as checkpoints, optimizer states, processed shards, logs, and traces. Unlike versioned Models and Datasets repositories, Buckets support frequent writes, overwrites, directory synchronization, and cleanup while retaining standard Hub permissions and public or private visibility. Built on Xet, they chunk and deduplicate content across files, reducing repeated bytes, bandwidth, transfer time, and, for Enterprise customers, billed storage. The CLI and Python APIs support creation, syncing, dry runs, plans, inspection, individual file operations, and programmatic workflows; HfFileSystem and fsspec also allow compatible libraries to read and write through hf:// paths. Pre-warming can place hot data near AWS or GCP compute, while roadmap work aims to connect Buckets directly with versioned repos.
At-scale reinforcement-learning training is constrained by long, variable data-generation workloads that leave training GPUs idle, especially for reasoning and agentic models. The surveyed architecture separates inference and training onto distinct GPU pools, connects them through a rollout buffer, and synchronizes weights asynchronously so both stages can proceed concurrently. Sixteen open-source libraries are compared across orchestration, buffering, weight synchronization, staleness, partial rollouts, LoRA, and distributed backends; Ray appears in 8 of 16 orchestration implementations, while NCCL broadcast is the default weight-transfer method. The proposed TRL async trainer favors a lightweight bounded queue with per-token model_version metadata, packed NCCL transfers, and experimental support for prefix-resume or abort-and-retry handling of rollouts interrupted by weight updates.
The best engineering reads, once a week.
