---
title: "Architecture"
description: "834 posts about Architecture, summarised, each linking to the original."
---

# Architecture
> 834 posts about Architecture, summarised, each linking to the original.

## Articles

### [Integrating the Rust Delta Kernel into ClickHouse](https://yomu.fyi/post/integrating-the-rust-delta-kernel-into-clickhouse.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Melvyn Peignon, Kseniia Sumarokova, Raúl Marín
- Published: May 22, 2026
- Type: Problem & solution

ClickHouse initially built native protocol support to read Delta Lake tables, but maintaining pace with the format's evolving specification created high complexity and fragmented features. To resolve this, engineers integrated the Rust Delta Kernel to offload protocol handling while retaining ClickHouse's optimized Parquet data reading. The Delta Kernel parses JSON transaction logs, resolves snapshots, manages transactional write metadata, and exposes schema evolution mapping alongside Change Data Feed streams. ClickHouse also contributed upstream enhancements to the Rust kernel, adding dynamic runtime logging and asynchronous metadata processing via FFI handles. While ClickHouse can attach to existing Delta tables, the integration cannot yet initialize empty Delta Lake tables.


### [Making User-Sequence Data More Cost-Efficient, Faster, and Easier to Use](https://yomu.fyi/post/making-user-sequence-data-more-cost-efficient-faster-and-easier-to-use.md)
- Company: [Pinterest](https://yomu.fyi/company/pinterest.md)
- Author: Pinterest Engineering
- Published: May 21, 2026
- Type: Problem & solution

User sequences serve as core primitives powering machine learning models across ranking, retrieval, and recommendation surfaces at Pinterest. Managing these sequences previously caused high infrastructure costs, operational fragility, and data drift between offline training pipelines and online serving systems. In response, engineering teams redesigned the platform around a single definition executed across multiple runtimes, including real-time streaming, batch backfills, and online inference. The updated architecture pairs a unified enrichment engine with columnar time-partitioned storage and standardized validation tiers. Consequently, the redesign achieved lower storage and network costs, faster signal onboarding via configuration, and improved engagement metrics across major recommendation surfaces.


### [How Partition Access Visualizations Reduced our Data Lake S3 Cost by 33%](https://yomu.fyi/post/how-partition-access-visualizations-reduced-our-data-lake-s3-cost-by-3.md)
- Company: [Yelp](https://yomu.fyi/company/yelp.md)
- Author: Nick Del Nano, Data Streaming
- Published: May 21, 2026
- Type: Problem & solution

Data teams at Yelp struggled to identify table consumers and track how datasets were accessed across their petabyte-scale data lake. To solve this, the team built a usage attribution system using Amazon S3 server access logs to visualize time-based partition keys against access timestamps and IAM roles. These visualizations revealed distinctive patterns for daily batch jobs, backfills, and ad hoc queries, guiding migrations of thousands of tables to Apache Iceberg format. Furthermore, Yelp implemented a Default Access Retention strategy gated by IAM bucket policies to prevent accidental queries from resetting S3 Intelligent Tiering transitions. Combining these granular usage insights with targeted S3 storage classes reduced Yelp's S3 storage costs by 33 percent.


### [Scaling Airbnb’s identity graph with a unified knowledge graph infrastructure](https://yomu.fyi/post/scaling-airbnb-s-identity-graph-with-a-unified-knowledge-graph-infrast.md)
- Company: [Airbnb](https://yomu.fyi/company/airbnb.md)
- Author: Lucen Zhao
- Published: May 19, 2026

Airbnb migrated its massive identity graph from a third-party graph database vendor to a unified, in-house knowledge graph platform to solve long-tail latency and scaling bottlenecks. Built on JanusGraph with AWS DynamoDB for persistence and OpenSearch for indexing, the new architecture decouples storage operations from graph traversal logic. The migration improved P99 query latency, eliminated routine instance reboots, and supported a tenfold increase in write throughput.


### [Scaling Real-Time Traffic Forecasting with a Graph-Aware Transformer](https://yomu.fyi/post/scaling-real-time-traffic-forecasting-with-a-graph-aware-transformer.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: May 19, 2026
- Type: Explainer

DeepETT is Uber’s real-time traffic forecasting system for estimating traversal times on roughly 100 million road segments, supporting routing and arrival-time estimates. It replaces a legacy blend of recent observations and historical patterns with a deep-learning model designed for rapidly changing conditions, sparse roads, and long trips. To keep deployment predictable at global scale, the system uses fixed-size computation over pre-aggregated segment, graph, region, hourly, weekly, and event views rather than a dynamic graph, while predicting segment ETTs independently of routing. Training targets segment-level mean squared error, while trip-level mean absolute error is evaluated offline after routing sums segment forecasts; calibration and continual real-time correction address downstream error. In a global A/B test, long-trip arrival-time accuracy improved 6%, variance explained rose 19%, navigation defects fell 2.73%, and mobility gross bookings increased by an estimated $100 million annually, with forecasts exceeding 2 million per second.


### [MemEx: A Programmable Scratchpad for LLM Agents](https://yomu.fyi/post/memex-a-programmable-scratchpad-for-llm-agents.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: The Databricks AI Research Team
- Published: May 19, 2026
- Type: Explainer

MemEx addresses the context-window bottleneck in LLM agents, where large tool outputs persist in the shared prompt even when later steps need only a small portion. It gives the agent a persistent, typed Python kernel: tool results remain as objects in scope, code can filter, summarize, and compose them, and only printed values are materialized as context tokens. MemEx preserves existing Tool Calling schemas while adding typed submit(), persistent state, backend-agnostic execution, and asynchronous spawn\_agent() sub-agents, enabling multi-step workflows and parallel trajectory analysis. Across nine-model evaluations on OfficeQA Pro and Enterprise Structured Retrieval, it was reported as better and cheaper than parallel structured Tool Calling; examples include roughly 10-point OfficeQA gains for some models and Qwen structured-retrieval accuracy rising from 18% to 36% and 20% to 38% at lower token cost.


### [How ChatFeatured migrated from PlanetScale Postgres to Postgres Managed by ClickHouse to power AI brand discovery](https://yomu.fyi/post/how-chatfeatured-migrated-from-planetscale-postgres-to-postgres-manage.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: ClickHouse
- Published: May 18, 2026
- Type: Problem & solution

ChatFeatured, a Toronto-based startup, helps brands track and improve how AI models discover and recommend them, including by analyzing cited sources and guiding content production. As prompt ingestion grew to about 1,000 prompts daily across six AI models, managed Postgres instances became slow, CPU reached 90% during ingestion windows, and PlanetScale hit an IOPS limit; analytics queries remained the deeper constraint. The team migrated from PlanetScale to Postgres managed by ClickHouse using the Postgres-to-Postgres ClickPipes connector, which handled initial load, schema transfer, continuous sync, and a five-minute maintenance window; the migration took 30 minutes overall. This setup retains Postgres for transactional workloads while providing a built-in path to ClickHouse for analytics, and a 30-day LifestyleRx query fell from about two and a half minutes to less than a second, with 20x compression reported for CDN log data.


### [Making LLMs faster without sacrificing accuracy](https://yomu.fyi/post/making-llms-faster-without-sacrificing-accuracy.md)
- Company: amazon.science
- Author: Tao Yu, Youngsuk Park
- Published: May 15, 2026
- Type: Problem & solution

Standard scaling laws optimize parameter count and training data budgets to minimize loss but ignore internal Transformer architecture configurations. To address throughput disparities among equally sized models, researchers developed a scaling framework incorporating hidden size, the multilayer perceptron to attention parameter ratio, and grouped-query attention. The method calibrates a correction surface over standard Chinchilla scaling laws to independently optimize separable architectural parameters for target accuracy. Evaluated across more than 200 models up to three billion parameters, the framework produced the Panda model family for maximum accuracy and the Surefire family for Pareto efficiency. Surefire models matched or exceeded LLaMA-3.2 accuracy while increasing inference throughput by 12% to 47% across vLLM and SGLang serving systems on modern GPUs.


### [Operator: A look under the hood](https://yomu.fyi/post/operator-a-look-under-the-hood.md)
- Company: [Intercom](https://yomu.fyi/company/intercom.md)
- Author: Jack Ryan
- Published: May 15, 2026
- Type: Explainer

Operator is a customer-operations Agent built to help teams understand, manage, and improve customer experience across workspaces. The post distinguishes a production Agent from a prototype that simply gives an LLM API access and a system prompt, describing more than 50 purpose-built tools and 10 skills for data retrieval, reporting, investigation, and operational tasks. Its intelligence layer combines semantic search tuned against millions of support conversations, workspace-specific attribute awareness, and encoded reasoning, while the action layer supports configuration and content changes through reviewable diffs requiring explicit approval. A hybrid conversational and graphical interface presents charts, dashboards, and structured proposals within the existing platform. The post argues that read-only custom tools can be built in weeks, but production reliability, safe actions, multi-tenant isolation, graceful degradation, and ongoing maintenance require substantially greater investment.


### [ClickHouse vs Prometheus for High Cardinality, Part 2: Cardinality in ClickHouse](https://yomu.fyi/post/clickhouse-vs-prometheus-for-high-cardinality-part-2-cardinality-in-cl.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Rory Crispin, Dale McDiarmid
- Published: May 15, 2026
- Type: Explainer

Traditional time-series databases like Prometheus experience write amplification, memory overhead, and operational challenges when high-cardinality label combinations churn rapidly. ClickHouse alters these economics by abandoning dedicated series objects in favor of a wide events model, storing telemetry as rows in columnar tables with dynamic Map attributes and metric measurements. Dynamic attributes can be handled using sharded Map types alongside materialized columns and text indexes for frequently queried labels. This design trades ingestion overhead for query-time processing, relying on columnar compression, partition pruning, and parallel scans to compute aggregations on demand. ClickHouse avoids the ingest-side bottlenecks of high cardinality, though Prometheus remains preferable for moderate cardinality, native PromQL semantics, and single-series lookups.


### [Beyond Prediction: Solving the Multiple Knapsack Problem at Scale: How Uber Optimizes Incentives](https://yomu.fyi/post/beyond-prediction-solving-the-multiple-knapsack-problem-at-scale-how-u.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: May 14, 2026
- Type: Explainer

Uber’s Tarot platform frames incentive allocation as a Multiple Knapsack Problem (MKP), replacing FIFO contention handling and reactive budget management. It combines segmentation, budget pacing, uplift predictions, multi-objective value scoring, and a plug-and-play optimizer coordinated by an orchestrator. For each run, Tarot forms user-lever opportunities, predicts incremental ROI and cost, reconciles available budget, and selects assignments under team, business, and concurrency constraints before publishing them to domain services. The initial HiGHS linear programming approach took more than 24 hours to find a feasible solution for 100,000 users and one lever, while CP-SAT solved the same dataset in minutes and handled millions of users across multiple levers within stated SLA timeframes. The result is a production architecture that moves beyond heuristics by combining constraint programming with budget pacing, while future work targets automated exploration and model feedback.


### [From latency to instant: Modernizing GitHub Issues navigation performance](https://yomu.fyi/post/from-latency-to-instant-modernizing-github-issues-navigation-performan.md)
- Company: [Github](https://yomu.fyi/company/github.md)
- Author: Natalie Guevara
- Published: May 14, 2026
- Type: Problem & solution

GitHub Issues addressed navigation latency by shifting workloads to the client using a local-first, stale-while-revalidate architecture. To evaluate perceived delays, the team tracked user transitions through Highest Priority Content thresholds, aiming for instant rendering in under 200 milliseconds. The core implementation added a persistent client-side cache using IndexedDB, paired with a synchronous in-memory tier to serve hot issue payloads without asynchronous overhead. A selective preheating mechanism resolves missing cache entries for high-intent links without overburdening backend capacity with redundant requests. Following broad rollout, the proportion of instant React soft navigations increased from 4% to approximately 22%, yielding an overall cache-hit ratio of around 33%.


### [ClickHouse vs Prometheus for High Cardinality, Part 1: Understanding the Problem](https://yomu.fyi/post/clickhouse-vs-prometheus-for-high-cardinality-part-1-understanding-the.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Rory Crispin, Dale McDiarmid
- Published: May 14, 2026
- Type: Explainer

Observability systems frequently encounter high cardinality when metrics contain numerous unique label combinations across multidimensional cloud infrastructure. In Prometheus, every unique label combination instantiates an independent time series that incurs metadata overhead, requires in-memory memSeries registration, and continually expands inverted index posting lists. The ingestion engine stores samples in two-hour Head blocks using compressed 120-sample chunks with XOR-based encodings, but rapid series creation and elevated scrape frequencies multiply chunk allocations and memory consumption. When executing read operations, broad aggregations and regular expressions require intersecting expansive posting lists and decoding entire chunks sequentially because the engine cannot push value predicates directly into compressed storage blocks. Consequently, managing high-cardinality workloads in series-oriented architectures forces operational trade-offs between system stability and granular visibility into ephemeral components.


### [Unlocking asynchronicity in continuous batching](https://yomu.fyi/post/unlocking-asynchronicity-in-continuous-batching.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Rémi Ouazan Reboul, Pedro Cuenca, Aritra Roy Gosthipaty
- Published: May 14, 2026
- Type: Tutorial

Continuous batching improves GPU utilization with tightly packed batches, but its synchronous loop makes the CPU and GPU take turns, leaving the GPU idle while the CPU prepares each next batch. Profiling 8K-token generation with a batch size of 32 and an 8B model recorded 300.6 seconds, with the GPU idle for 24.0% of runtime. The post builds asynchronous batching by overlapping CPU preparation with GPU computation, using non-default CUDA streams, non-blocking transfers, and synchronization events to coordinate dependencies. In the same experiment, GPU activity rose to 99.4% and total generation time fell to 234.5 seconds, a 22% speedup, without new kernels or model changes while maintaining model accuracy.


### [Postgres FDW: Pushdown is a negotiation](https://yomu.fyi/post/postgres-fdw-pushdown-is-a-negotiation.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Kaushik Iska, David Wheeler, Philip Dubé
- Published: May 14, 2026
- Type: Explainer

Postgres Foreign Data Wrapper extensions allow PostgreSQL to query external datastores like ClickHouse by delegating execution to remote engines. Engineering pg\_clickhouse centers on pushdown decisions that determine whether SQL expressions run remotely or stream raw data back across the wire. Because pushdown depends on planner callbacks, deparser translations, and ClickHouse semantic compatibility, a single untranslated clause can block upper-level query pushdown entirely. Resolving pushdown barriers enables complex analytic queries to return hundreds of rows in milliseconds rather than pulling tens of millions of rows for local processing. Ultimately, pushdown development requires an iterative negotiation across differing SQL grammars, occasionally requiring the revocation of translations that fail to maintain strict equivalence.


### [ClickStack SQL Charting and Alerting](https://yomu.fyi/post/clickstack-sql-charting-and-alerting.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Drew Davis, Dale McDiarmid
- Published: May 13, 2026
- Type: Announcement

ClickStack introduced SQL-based visualizations and SQL-based alerting to allow users to build charts and alerts using arbitrary ClickHouse SQL queries. Traditional query builders trade analytical flexibility for simplicity, preventing teams from easily expressing multi-stage computations like rolling baselines, anomaly detection, or dynamic SLO calculations. To maintain interactivity within dashboards, the system supports ClickHouse query parameters and macros that inject active time ranges, interval sizes, dashboard filters, and source tables. Furthermore, SQL-based alerting evaluates the numeric columns output by these queries, such as returning binary flags when values exceed statistical boundaries. By shifting computational logic directly into SQL, teams can implement complex operational checks without relying on static external thresholds.


### [An Engineer’s Guide to Better AI Skills: Implementing a Testing Process to Optimize Agent…](https://yomu.fyi/post/an-engineer-s-guide-to-better-ai-skills-implementing-a-testing-process.md)
- Company: [Pinterest](https://yomu.fyi/company/pinterest.md)
- Author: Pinterest Engineering
- Published: May 12, 2026
- Type: Problem & solution

Software engineers deploying AI agents often encounter unreliable skill invocation for repository-specific architectures, such as Pinterest's iOS framework. To quantify and resolve this issue, a Bash test harness piped positive and negative prompts into Pin-agent and Claude Code while parsing streamed JSON logs for execution signatures. Initial vanilla benchmarks yielded baseline accuracies of 73% for GPT 5.2-codex and 62% for Opus 4.5 Claude. Optimizations such as expanding YAML frontmatter descriptions with architectural context, utilizing capitalized imperative directives, and adding reference tables in AGENTS.md raised invocation rates. Both agents reliably loaded skills when prompts contained explicit invocations, demonstrating that empirical testing paired with descriptive prompting stabilizes agent skill adoption.


### [ClickHouse Cloud: Fast, Updatable Lookups with the Join Table Engine](https://yomu.fyi/post/clickhouse-cloud-fast-updatable-lookups-with-the-join-table-engine.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Hellmar Becker
- Published: May 12, 2026
- Type: Explainer

ClickHouse Cloud provides an in-memory Join table engine for precomputed hash-map lookups and joins, aimed at enriching append-only fact data with dimensions that change occasionally. With Join(ANY, LEFT, key), inserting another row for a key acts as an upsert, while joinGet and matching JOIN clauses use the in-memory structure. Open-source Join tables are not distributed and persist each INSERT batch as separate .bin files without automatic compaction, but ClickHouse Cloud transparently uses SharedJoin with MergeTree for ALL or ReplacingMergeTree for ANY. In the demonstrated enrichment query, duplicate inserts for id 1 leave value 42 visible, while the underlying table retains duplicates until merging; the in-memory load applies FINAL and always selects the latest entry, making ANY LEFT suitable for frequently updated dimensions.


### [Fighting Tool Sprawl: The Case for AI Tool Registries](https://yomu.fyi/post/fighting-tool-sprawl-the-case-for-ai-tool-registries.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Author: Peter Richards
- Published: May 11, 2026
- Type: Problem & solution

Rapid enterprise AI agent adoption without centralized infrastructure introduces widespread tool sprawl, duplicated engineering effort, and severe operational opacity. Individual application teams frequently build ad hoc, undocumented integrations, leaving security teams unable to discover, audit, or approve active agent capabilities. Internal enterprise tool registries address this coordination failure by establishing an organization-specific catalog tailored to internal security policies, data, and compliance requirements. A production-grade registry delivers core discovery and version tracking functionality while surfacing security certification metadata and enabling consistent identity-scoped access control across teams. Organizations establishing centralized tool infrastructure can accelerate agent onboarding, implement deny-by-default governance policies, and mitigate technical debt before uncoordinated deployments compound across the enterprise.


### [Powering self-driving vehicle analytics at Avride with ClickHouse Cloud](https://yomu.fyi/post/powering-self-driving-vehicle-analytics-at-avride-with-clickhouse-clou.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: ClickHouse
- Published: May 11, 2026
- Type: Problem & solution

Avride’s expanding autonomous passenger vehicle and delivery robot fleet generated thousands of data points per minute, overwhelming its legacy Apache Iceberg and Parquet metadata indexing architecture. Iceberg’s optimistic concurrency model struggled with parallel writes, pushing ingestion latencies to hours or days, inflating query times to 10–20 seconds, and requiring raw data duplication across storage locations. To resolve these bottlenecks, Avride built a custom indexing and metrics warehouse powered by ClickHouse Cloud on AWS backed by Amazon S3 object storage. The new architecture dropped ingestion delays to seconds, reduced worst-case lookup latencies to under one second, eliminated the need for bespoke caching layers, and removed payload duplication. Beyond operational analytics, the unified ClickHouse platform now supports internal ride-replay debugging tools, simulation workflows, and high-cardinality vehicle C++ performance trace profiling.


[Newer posts](https://yomu.fyi/topic/architecture/page/18.md) · [Older posts](https://yomu.fyi/topic/architecture/page/20.md)