The GeekNarrator

Performance is getting harder to ignore, and Tomás argues that AI is making the problem urgent in ways most teams are not ready for. He explains why AI-generated code, exploding system surface area, and rising infrastructure costs are creating a perfect storm, and why Perfloop exists to turn performance engineering into a rigorous, automated science.Tomás, the creator of Perfloop, breaks down how the product maps your codebase and telemetry into a living model of your system, then hunts for performance anti-patterns with a verification harness that puts every hypothesis through measured, repeatable experiments. Instead of relying on a freestyle coding agent, Perfloop uses structured discovery, deterministic checks, and adversarial verification to produce results you can actually trust.Chapters:00:00 The performance problem behind Perfloop00:39 Why performance work is a battle03:59 How AI agents raise the stakes08:55 Quality control at the scale of agentic development10:34 How Perfloop maps systems and finds opportunities14:32 Using telemetry to prioritize hot paths18:13 Perfloop’s open-source performance wins22:18 How Perfloop measures and verifies improvements25:11 Common performance patterns and bottlenecks28:47 The “apex hunt” for bigger architectural gains30:58 Initiatives and setting high-level performance goals35:21 Connecting code optimizations to customer impact39:59 Why independent verification matters44:36 Choosing the right AI models for each task49:45 What a harness does in Perfloop53:59 The possibility of pre-merge performance reviews57:24 How AI may change the role of performance engineers63:54 Why performance requires measurementImportant links:Patterns: https://app.perfloop.ai/t/oss/patternsPerfloop: https://app.perfloop.ai/t/oss/startFor memberships: join this channel as a member here:https://www.youtube.com/channel/UC_mGuY4g0mggeUGM6V1osdA/joinDon't forget to like, share, and subscribe for more insights!=============================================================================Like building stuff? Try out CodeCrafters and build amazing real world systems like Redis, Kafka, Sqlite. Use the link below to signup and get 40% off on paid subscription.https://app.codecrafters.io/join?via=geeknarrator=============================================================================Database internals series: https://youtu.be/yV_Zp0Mi3xsPopular playlists:Realtime streaming systems: https://www.youtube.com/playlist?list=PLL7QpTxsA4se-mAKKoVOs3VcaP71X_LA-Software Engineering: https://www.youtube.com/playlist?list=PLL7QpTxsA4sf6By03bot5BhKoMgxDUU17Distributed systems and databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4sfLDUnjBJXJGFhhz94jDd_dModern databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4scSeZAsCUXijtnfW5ARlrsNStay Curios! Keep Learning!

Andrey Breslav (creator of Kotlin and founder of CodeSpeak) discusses why, in the AI era, the hard problem isn’t generating code but clearly deciding what software should do. He explains how “specs” range from formal documents to informal prompts, and why CodeSpeak is moving from long docs to a structured, versioned set of interconnected requirements extracted from everyday agent chats. The conversation covers how clearer requirements reduce ambiguity (including using output stability/semantic entropy), help teams communicate, surface inconsistencies early, and ease review bottlenecks as AI increases code volume. They compare prompt-driven work to spec-driven workflows, talk through examples like a modified tic-tac-toe game, and explore testing (regression, conformance, acceptance), constraints and temporary decisions, modularity and discoverability of requirements, and how roles and languages may evolve while humans stay in control.Chapters:00:00 Introduction04:16 What Counts as a Spec07:49 Human vs Machine Specs10:16 Codespeak Structured Requirements12:14 Writing Better Requirements17:51 Right Abstraction Level22:25 Why Prompts Break at Scale29:58 Keeping Requirements in Sync33:30 Testing in Spec Driven Dev39:01 Beyond Tests and Trust41:11 Abstraction Tradeoffs43:49 Trusting Generated Code44:34 Safe Zones and Checks46:58 Shared Spec Artifacts49:24 Modular Specs and Grounding52:28 Why Teams Adopt Specs55:11 Review Bottlenecks and Balance01:00:21 Capturing Constraints as Intent01:02:29 Do Languages Still Matter01:07:48 Roles in the AI Era01:10:42 CodeSpeak Workflow SummaryImportant Links:- https://codespeak.dev/For memberships: join this channel as a member here:https://www.youtube.com/channel/UC_mGuY4g0mggeUGM6V1osdA/joinDon't forget to like, share, and subscribe for more insights!=============================================================================Like building stuff? Try out CodeCrafters and build amazing real world systems like Redis, Kafka, Sqlite. Use the link below to signup and get 40% off on paid subscription.https://app.codecrafters.io/join?via=geeknarrator=============================================================================Database internals series: https://youtu.be/yV_Zp0Mi3xsPopular playlists:Realtime streaming systems: https://www.youtube.com/playlist?list=PLL7QpTxsA4se-mAKKoVOs3VcaP71X_LA-Software Engineering: https://www.youtube.com/playlist?list=PLL7QpTxsA4sf6By03bot5BhKoMgxDUU17Distributed systems and databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4sfLDUnjBJXJGFhhz94jDd_dModern databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4scSeZAsCUXijtnfW5ARlrsNStay Curios! Keep Learning!

The episode explores why modern databases keep reinventing the same distributed-systems machinery and argues that a major part of database cost is the operational tax of running replication-heavy systems. Our guest, Almog Gavra, co-founder of Responsive, explains how his team pivoted from operating Kafka Streams as a service to building SlateDB and the “Open Data” manifesto: an object-storage-native LSM foundation that can power multiple database types (vector, time series, logs, key-value) with shared tuning knobs and failure modes. They discuss why distributed-systems complexity is often harder than query engines, how LSM trees provide a tunable tradeoff between read/write/space amplification, caching layers and cost transparency, separating readers/writers, stateless ingest, single-writer availability and fencing via S3 compare-and-set, offloading compaction, and how the architecture enables near-free snapshots. They also cover when this approach doesn’t fit: OLTP that can stay on Postgres and ultra-low-latency workloads where cold object-store misses are unacceptable.Chapters:00:00 Introduction08:36 Open Data Manifesto18:34 Specialized vs General25:10 SlateDB Architecture32:51 LSM Trees as Tuning Dial38:58 Tuning Without Overload39:46 Cost Aware Config Knobs41:51 Latency Cost Durability Tradeoffs46:46 Caching Strategies And Layers50:23 Split Readers And Writers52:43 Single Writer Versus Multi Writer55:16 Scaling And Partitioning Writes58:58 Failure Modes And Fencing01:05:23 Compaction As Separate Worker01:09:28 Snapshots And Garbage Collection01:10:25 When Open Data Is Not FitImportant links and references:OpenData: http://github.com/opendata-oss/opendataOpenData manifesto: https://www.opendata.dev/blog/manifestoReach out to Almog: https://www.linkedin.com/in/agavra/ or https://x.com/almoggavraDostovesky paper on LSM: https://nivdayan.github.io/dostoevsky.pdfLatency/Cost/Durability Triad: https://materializedview.io/p/cloud-storage-triad-latency-cost-durabilitySlateDB: https://github.com/slatedb/slatedb"how SSTs work": https://www.bitsxpages.com/p/sorted-string-tables-sst-from-firstFor memberships: join this channel as a member here:https://www.youtube.com/channel/UC_mGuY4g0mggeUGM6V1osdA/joinDon't forget to like, share, and subscribe for more insights!=============================================================================Like building stuff? Try out CodeCrafters and build amazing real world systems like Redis, Kafka, Sqlite. Use the link below to signup and get 40% off on paid subscription.https://app.codecrafters.io/join?via=geeknarrator=============================================================================Database internals series: https://youtu.be/yV_Zp0Mi3xsPopular playlists:Realtime streaming systems: https://www.youtube.com/playlist?list=PLL7QpTxsA4se-mAKKoVOs3VcaP71X_LA-Software Engineering: https://www.youtube.com/playlist?list=PLL7QpTxsA4sf6By03bot5BhKoMgxDUU17Distributed systems and databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4sfLDUnjBJXJGFhhz94jDd_dModern databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4scSeZAsCUXijtnfW5ARlrsNStay Curios! Keep Learning!

Assertions vs Types: Design by Contract, Deterministic Simulation Testing, and Safety vs Availability (TigerBeetle Vol. 3)In this episode (volume 3), Kai talks with Alex about assertions, type systems, and design by contract, arguing it’s not “types vs assertions” but using both: types for cheap, structural guarantees and assertions where types become too costly or obscure logic. Alex defines assertions (in Zig) as a function that crashes the program on false, explains why disabling assertions in production is dangerous, and ties reliable assertion use to deterministic simulation/generative testing to exercise error paths. We discuss aiming for very high assertion density, repeating weak and strong assertions across call sites and callees to form an interlocking “net,” and note you don’t need special language features for DBC. We also cover safety vs availability tradeoffs, recovery/isolation boundaries (“let it crash” with recovery), TigerBeetle’s approach to correctness, a real cache/hash-table bug caught by an assertion, handling poison-pill failures via fix-forward and frequent releases, control plane vs data plane performance tactics for assertions, and why Alex doesn’t use AI to write TigerBeetle code.Chapters:00:00 Welcome and Intro01:33 Assertions Versus Types03:05 Silver Bullets Origins08:10 When Types Get Costly11:06 What Is an Assertion?12:40 Never Disable Assertions15:30 Testing and Error Paths19:52 Simulation Testing Harness22:50 Where to Assert Everywhere27:01 Redundant Contracts Benefits33:08 No Language Features Needed38:01 Visibility and Abstractions40:47 Boundaries and Integration44:01 Safety Versus Liveness Setup44:31 Safety vs Availability Tradeoffs46:16 Let It Crash Philosophy47:13 Isolation and Recovery Boundaries48:02 TigerBeetle vs IDE Priorities53:48 Always Assertions Pattern55:46 Cascading Failures in Clusters57:57 Fix Forward and Fast Releases01:02:27 Worst TigerBeetle Bug Story01:07:00 Control Plane vs Data Plane01:09:50 Assertion Performance Tactics01:15:18 AI Limits for Safety Systems01:18:55 Closing Advice on AssertionsFor memberships: join this channel as a member here:https://www.youtube.com/channel/UC_mGuY4g0mggeUGM6V1osdA/joinDon't forget to like, share, and subscribe for more insights!=============================================================================Like building stuff? Try out CodeCrafters and build amazing real world systems like Redis, Kafka, Sqlite. Use the link below to signup and get 40% off on paid subscription.https://app.codecrafters.io/join?via=geeknarrator=============================================================================Database internals series: https://youtu.be/yV_Zp0Mi3xsPopular playlists:Realtime streaming systems: https://www.youtube.com/playlist?list=PLL7QpTxsA4se-mAKKoVOs3VcaP71X_LA-Software Engineering: https://www.youtube.com/playlist?list=PLL7QpTxsA4sf6By03bot5BhKoMgxDUU17Distributed systems and databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4sfLDUnjBJXJGFhhz94jDd_dModern databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4scSeZAsCUXijtnfW5ARlrsNStay Curios! Keep Learning!

In this episode, Catherine, Co-Founder, and CEO of Kernel, reveals the magic behind Kernel's approach using uni-kernels and micro VMs, enabling browser startups in just milliseconds. We talk about the technical challenges of agent-based web interactions, how Cash App leverages Kernel for live QA of e-commerce sites, the intricacies of handling authentication, and the future potential of attaching virtual GPUs for optimal performance. Whether you're an AI developer or fascinated by the backbone of internet automation, don't miss this insightful conversation. Chapters:00:00 Introduction to AI Agent Challenges00:35 Welcome and Episode Overview01:58 Guest Introduction and Background02:39 The Problem Statement and Kernel's Origin07:21 Infrastructure for AI Agents16:09 Kernel's Technical Architecture20:33 Use Cases and Real-World Applications24:38 Challenges and Future Directions27:20 Running on Bare Metal: Optimizing Browser Instances28:02 Challenges in Infrastructure Optimization30:34 Headful Browsers and Human Interaction30:58 Recording and Debugging Browser Sessions33:42 Preventing Misuse of Browser Automation39:55 Handling Authentication and Secure Access44:57 Bot Detection and Good Actor Automations48:56 Future Developments and GPU Integration52:41 Conclusion and Final ThoughtsImportant links:- Homepage to go get a free account (no credit card required) and just try us out: https://www.kernel.sh/ - Our chromium on Unikernels OSS repo: https://news.ycombinator.com/item?id=43705144- The blog post where in it we benchmarked ourselves against all others and ranked the fastest browser infrastructure in the world: https://www.kernel.sh/blog/fast For memberships: join this channel as a member here:https://www.youtube.com/channel/UC_mGuY4g0mggeUGM6V1osdA/joinDon't forget to like, share, and subscribe for more insights!=============================================================================Like building stuff? Try out CodeCrafters and build amazing real world systems like Redis, Kafka, Sqlite. Use the link below to signup and get 40% off on paid subscription.https://app.codecrafters.io/join?via=geeknarrator=============================================================================Database internals series: https://youtu.be/yV_Zp0Mi3xsPopular playlists:Realtime streaming systems: https://www.youtube.com/playlist?list=PLL7QpTxsA4se-mAKKoVOs3VcaP71X_LA-Software Engineering: https://www.youtube.com/playlist?list=PLL7QpTxsA4sf6By03bot5BhKoMgxDUU17Distributed systems and databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4sfLDUnjBJXJGFhhz94jDd_dModern databases: https://www.youtube.com/playlist?list=PLL7QpTxsA4scSeZAsCUXijtnfW5ARlrsNStay Curios! Keep Learning!