Loading…
Browse by type
Problem & solution reads
Problem & solution posts on Yomu, summarised newest first.
- 903 posts
- 20 companies
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.
Problem & solution posts
Boosting multimodal inference performance by >10% with a single Python dictionary
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.
Gala supercharges analytics performance with ClickHouse on AWS
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.
The federal data paradox: Rich in data, poor in access
Federal agencies have invested substantially in data infrastructure, but program directors, policy analysts, oversight officials, and budget examiners often still rely on technical intermediaries to answer operational questions. The post frames this as the unresolved “last mile” of federal data modernization: data lakes, APIs, dashboards, evidence-based policymaking mandates, and agency CDO functions have advanced infrastructure without making it usable by most decision-makers. It presents Databricks Genie as a natural-language interface that lets staff query agency data in plain language, including questions requiring joins across disbursement, eligibility, and geographic data, while retaining existing access controls and policies. Genie runs on Unity Catalog with role-based access controls, audit logging, and data lineage; it also supports federated cross-agency queries and records queries, answers, and sources, which the post associates with oversight, FOIA readiness, and accountability.
Optimizing ML Workload Network Efficiency (Part I): Feature Trimmer
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.
Driving Budapest forward: How BKK uses Databricks to transform city mobility
BKK, Budapest’s unified transport authority, manages public transit, shared mobility, infrastructure, and traffic systems for a city of 1.7 million people. Its legacy on-premises warehouse and fragmented Excel- and PowerBI-based reporting struggled with growing, varied data, limiting efficient access and analysis. BKK phased a migration to Azure Databricks, starting with mobility datasets and carefully modeling and validating data from vehicle GPS, passenger sensors, boarding schedules, and other systems. The Databricks Lakehouse lets analysts process large geospatial datasets in shared SQL, Python, or R notebooks, supporting minute-level shared-mobility tracking, route-performance analysis, airport-bus forecasting, and dynamic scheduling. These capabilities provide faster insights and more responsive decisions, while governed access and cost visibility support broader collaboration and BKK’s vision for a digital twin of Budapest’s mobility system.
Goodbye limitations, hello data: How Qonto is rethinking observability with ClickHouse Cloud
Qonto struggled with slow queries, system instability, and aggressive data sampling when handling distributed traces in Grafana Tempo. To enable wide-event querying without strict cardinality limits, the team migrated their telemetry storage to ClickHouse Cloud. Telemetry from applications, Kubernetes, and GitHub routes through OpenTelemetry collectors over AWS PrivateLink into ClickHouse, while Grafana remains the primary dashboard interface. By compressing 231 TB of uncompressed trace attributes down to 376 GB, Qonto achieved substantial storage savings and integrated an AI incident companion using the ClickHouse Model Context Protocol. The new architecture eliminated cardinality policing, allowed product and support teams to self-serve incident investigations, and enabled real-time streaming pre-aggregations using Apache Flink.
Shipping faster isn't learning faster
Product organizations can ship features in days while taking weeks to understand their behavioral impact, because fragmented analytics stacks depend on analysts, BI expertise, or SQL skills. The post argues that this architectural bottleneck slows the insight-to-ship cycle, causing roadmap decisions to rely on instinct, anecdotes, and lagging indicators. It presents Databricks AI/BI Genie as a conversational interface to event-level behavioral data, with experiment integration, cohort analysis, and product-specific growth-metric definitions. According to the post, product leaders can ask questions in plain language without filing analyst requests, while governed data access supports faster follow-up and feature-impact decisions. It reports that Genie users across 3,300+ Databricks customers cited a 49% productivity gain, 41% faster speed to market, and 5x faster ad-hoc analysis, though these are reported customer results.
Why your OEE dashboard is lying to you
Manufacturing OEE dashboards can look healthy while omitting recurring downtime that supervisors recognize, because relevant facts are split across SCADA, MES logs, maintenance tickets, and shift reports. OEE combines Availability, Performance, and Quality, but extracting those inputs often requires SQL or an analyst, delaying root-cause analysis after throughput drops. Databricks Genie is presented as a conversational AI layer over a unified data platform that leaves MES and SCADA in place while allowing business leaders to ask plain-language questions such as OEE by line against maintenance windows. Its semantic-layer awareness maps terms to actual fields, while governed access and logged, source-cited answers support different visibility levels and traceability. The proposed shift is from static reporting to faster operational questioning about production status, forecast risk, line performance, and quality signals.
The turbine that tried to tell you it was failing
Gas turbines produce millions of daily readings—vibration, temperature, pressure, flow rates, and electrical output—but maintenance teams may learn about warning signals only when an unplanned outage occurs. The post argues that predictive maintenance has struggled less because of model capability than because decision-makers lack fluid access to model findings and operational context. Databricks Genie provides a conversational interface to sensor, maintenance, dispatch, cost, and regulatory data, allowing questions about elevated vibration against maintenance baselines and timing maintenance around outage cycles. It is presented as decision support, not automation, with time-series analysis, maintenance history, integrated costs, and compliance context helping asset managers investigate fleets and act with more confidence.
Predicting readmissions isn't enough. Acting in time is.
Readmission risk models can identify patients likely to return within 30 days, but prediction alone does not ensure timely intervention. In large health systems, risk information may remain in population-health dashboards or EHR flags without reaching the care coordinator with enough context to create an effective post-discharge plan. The post presents Databricks Genie as a natural-language interface for governed clinical and outcomes data, allowing leaders to query readmission rates and compare periods while combining EHR, operational, financial, risk-score, intervention, and outcome information. Genie operates within Unity Catalog governance, with access controls, audit logging, and de-identification policies, and is described as supporting EHR integration and clinical taxonomy awareness; the intended result is to shorten the path from prediction to action.
Clinical trials run longer than they have to. That's a patient problem.
Clinical trial operations can lose weeks because site-level performance data is reviewed with a 2–4 week lag: enrollment velocity is monthly, while protocol deviation rates are assessed quarterly. The post presents Databricks Genie as a natural-language interface to unified trial data, allowing clinical operations leaders to query enrollment, screen-failure, protocol-deviation, query-response, and data-entry metrics across sites. Its stated capabilities include automatic site comparison, integration across CTMS, EDC, safety databases, and site-performance data, plus protocol-aware reasoning and traceability to source records for GCP documentation. The example query identifies Phase II oncology sites with screen-failure rates above 40% over 60 days and compares enrollment pace with activation targets, positioning earlier detection as a way to reduce timeline impact and speed treatment access.
Network quality is a revenue problem, not a technical one
Telecommunications network operations centers monitor extensive performance telemetry, but the source describes a gap between technical network events and the customers, contracts, and revenue they affect. A degraded tower serving 12,000 postpaid customers averaging $85 per month presents a different business problem from one affecting 12,000 prepaid customers with high churn propensity, while enterprise SLA exposure can remain hidden from the NOC. Databricks Genie connects network elements, serving areas, customer records, contract terms, SLA thresholds, and churn-risk scores so leaders can query commercial impact in seconds. It also supports proactive alerts based on defined commercial thresholds, enabling restoration and investment decisions to account for customer impact rather than technical severity alone.
When predicting the next hit requires more than intuition
Content investment decisions in entertainment commit substantial funds despite limited information, because leaders often combine executive instinct, competitive benchmarks, and historical performance data. Although companies hold viewing and subscriber signals such as episode completion, skip patterns, genre performance by demographic, acquisition, and retention, those insights are frequently buried in dashboards and spreadsheets. The post presents Databricks Genie as a natural-language interface that lets content leaders query governed performance data directly, including comparisons linking content types with 90-day retention, subscriber demographics, lifetime value, acquisition source, and historical greenlights. It argues that current, conversational access can complement creative judgment by reducing dependence on delayed analyst recommendations, while external market data can add competitive context.
Rethinking SQL ETL for modern data platforms
SQL ETL is often distributed across a warehouse, transformation framework, orchestrator, monitoring, lineage, and data-quality systems, creating operational fragmentation as data teams scale. The post proposes a unified platform model that combines execution, orchestration, observability, governance, and optimization while supporting analytics engineers, warehouse engineers, and analysts through dbt, scripts and stored procedures, Materialized Views, declarative pipelines, and no-code tools. Databricks implements this model with shared execution, governance, and observability, using serverless infrastructure and AI-driven optimization alongside open table formats and ANSI SQL. The stated benefits include reduced operational overhead, support for batch and real-time workloads, and incremental modernization rather than disruptive rewrites. Reported examples include HP’s 32% cloud savings and 36% lower combined job runtime after moving to serverless compute, and Adobe reducing query execution from eight minutes to three seconds with Materialized Views.
Evaluating AI at Scale: How Thumbtack Approaches Reliability, Safety, and Quality in GenAI
Thumbtack evaluates generative AI features across customer and service professional experiences to prevent unsupported claims, tone inconsistencies, and safety issues. The engineering organization transitioned from decentralized evaluation efforts to a dedicated cross-functional Evals team that provides shared infrastructure while preserving product team autonomy. System architectures combine rule-based checks, AI-as-a-judge scorers powered by DeepEval, trust and safety assessments, and crowdsourced human reviews. Teams leverage three parallel orchestration paths, including an MLflow-tracked stack, nightly Databricks batch jobs integrated with data warehouses, and multi-layer human-in-the-loop pipelines. Future milestones target centralized LLM judge registries, full task journey evaluations, automated judge-writing workflows, and cost-aware evaluation orchestration.
uAct - Unified Action Platform
Uber built uAct to consolidate action items and approval requests generated by its many internal applications, addressing notifications that could be lost among email and follow-ups. It combines real-time fetching with scheduled offline ingestion through Apache Camel, stores requests, and merges them in received and sent views. Users can act through deep links or within uAct, where the Action Dispatcher authenticates and authorizes actions and the Audit Trail Provider records them for compliance. Its onion-architecture backend uses orchestrators, application-specific mappers, MySQL for requests and CacheFront for preferences, while parallelized paginated calls and strict timeouts support performance as integrations grow. The platform has been implemented across Uber and also provides configurable views, hiding, manager filters, observability, privacy controls, and planned Slack notifications, reporting, nudging, and additional integrations.
AI is the answer to the sales growth-without-headcount problem
Sales leaders have treated pipeline growth as constrained by headcount, making a 30% increase appear to require 30% more people. The post presents Intercom’s use of Fin, its Customer Agent, for inbound sales as a separate pipeline source rather than merely an efficiency layer for SDRs, with its own metrics, targets, and owner. Fin engages inbound prospects instantly, qualifies and routes high-intent visitors, and keeps lower-intent prospects warm, while SDRs focus on phone qualification, relationship-building, and helping trial users realize value. Drawing on customer-service implementation, the operating model includes dedicated AI ownership and continuous optimization against meetings, pipeline, and revenue. The conclusion is that AI can raise the pipeline ceiling above current headcount capacity, although the constraint shifts to how well the Agent understands the product, customers, and demand qualification.
Formatting an entire 25 million line codebase overnight: the rubyfmt story
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.
From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest
Pinterest built a dedicated shopping candidate generation model to optimize for lower-funnel purchase conversions rather than relying strictly on engagement retrieval models. Because offsite conversion events are sparse and delayed, engineers trained a single multi-surface two-tower model using conversion data supplemented by duration-weighted click engagement and unengaged ad impressions as hard negatives. The two-tower retrieval architecture incorporates a parallel DCN v2 and MLP layer structure that allows simultaneous learning from the same inputs without information bottlenecks. Later revisions transitioned the system from a multi-head loss structure to a unified single-head multi-task architecture paired with an advertiser-level objective. Deployed across more than 600 million monthly active users, the system increased shopping conversion volume by 2.3% and boosted return on ad spend by 3.1% for United States shopping campaigns.
Powering KPMG UK Audit's AI future with Databricks
KPMG UK is evolving its Audit data platform with Databricks to support AI-enabled analytics while preserving governance, rigour, and professional standards. The program converges structured data, advanced analytics, and AI on a unified cloud-native Lakehouse, with Databricks SQL replacing core SQL Server workloads and Delta underpinning the platform. Lakebridge assessed migration complexity, while Databricks-hosted large language models including Claude Sonnet and Genie Code helped convert T-SQL, refactor stored procedures, modularize queries, and suggest Delta-based optimizations under engineer review. This reduced refactoring time by around 60% and enabled modernization of more than 400 scripts and stored procedures in roughly three months. Databricks SQL Serverless provides elastic compute for spiky workloads, while Genie offers traceable, version-controlled SQL and Delta Sharing supports governed data exchange.
Top topics
- Architecture830
- Performance813
- LLMs655
- Machine Learning619
- Open Source562
- AI Agents546
- AI459
- Developer Experience447
Weekly digest
The best engineering reads, once a week.