Scale By the Bay 2023
2023-11-13..2023-11-15
Oakland, United States
[Scale By the Bay 2017, Scale By The Bay 2018, Scale By The Bay 2019, Scale By the Bay 2020, Scale By the Bay 2021, Scale By the Bay 2023]
| title | author | length |
|---|---|---|
| Highlights | 03m | |
Scale By the Bay is a community-focused, full-stack AI conference with many small rooms and well-curated tracks that keep up with current trends. Sessions… | ||
| Highlights of our venue! | 01m | |
ai.bythebay.io Nov 2025, Oakland, full-stack AI conference We have secured the most atmospheric venue for the Scale by the Bay Conference. Prepare to be… | ||
| Scaling Embeddings in Real-Time | Zander Matheson | 30m |
An end-to-end pipeline that turns a live Hacker News feed into up-to-the-minute retrieval context, implemented with a Bytewax Python-native stream processor… | ||
| Typed and compositional microservices in Unison | Paul Chiusano | 31m |
Unison and Unison Cloud let code be deployed and invoked like local functions: calling deploy : (A -> B) -> ServiceHash AB sends content and dependencies as… | ||
| An introduction to Open Accelerated Discovery | Karl Wehden | 29m |
OpenAD is an open source framework that wires deep search, generative design, simulation, retrosynthesis and robotic prototyping into a single API/CLI/notebook… | ||
| Keynote: AI, Quantum, and AI + Quantum | Anthony Annunziuata | 23m |
Generative AI already proposes molecule candidates and replaces some HPC simulations, and a production synthesis automation service connects model predictions… | ||
| Creating Context: Loading your life into a vector database | Ben Church | 31m |
Builds a retrieval‑augmented pipeline that takes a natural question, asks an LLM to generate a GraphQL query, executes it against a vectorized dataset, and… | ||
| Lesson learned from orchestrating large-scale GenAI, ML & data on k8s | Oleg Avdeëv, Riley Hun | 32m |
Metaflow on Kubernetes is production-ready but requires operational work: the team ports Metaflow on Kubernetes to support MPI-style parallelism, PyTorch and… | ||
| Open Source AI Coding Assistants: The Future of Software Development | Beyang Liu | 30m |
Sourcegraph builds an open-core AI coding assistant called Cody that anchors LLM outputs to a parsed code Knowledge Graph for much lower hallucination rates… | ||
| Kubernetes and MLOps for Scalable and Reproducible Generative AI | Annie Talvasto | 31m |
Explains how Kubernetes plus MLOps tools actually run generative AI pipelines, showing Kubeflow components used for interactive Jupyter notebooks, Kubeflow… | ||
| Hydro: A Stack for Faster Streaming Applications | Shadaj Laddad | 31m |
Hydroflow is a low-level streaming IR that compiles to Rust and runs each dataflow as a single-threaded program while composing across machines for… | ||
| Model transparency for AI/ML security | Mihai Maruseac | 19m |
Claims that model supply chain compromise rises alongside a >700% YoY jump in supply‑chain attacks and that real malicious LLMs already exist, and demonstrates… | ||
| Pinot, Why Are You So Fast? | Tim Berglund | 33m |
Apache Pinot claims ingest >1M events/sec directly from Kafka and serves interactive UI queries in single to tens of milliseconds by coupling columnar segments… | ||
| Why Haskell is a Terrible Choice for Startups (And Why We Picked It Anyway) | Avi Press | 31m |
Production Haskell backend serves about 200 million downloads a month and grew to 350,000 lines of Haskell, but costs show up as long GHC compile and… | ||
| Gateway to the Future: A Brief History of Kubernetes Ingress | Viktor Gamov | 30m |
Covers how Kubernetes routes external traffic from port‑forwarding and Service type=LoadBalancer to a single Ingress IP, shows demo wiring DNS with… | ||
| OSS vs. Open Source Software: What's Different? What's the Same? | Dean Wampler | 30m |
Running engineering for IBM's Accelerated Discovery Platform shows open-source science prioritizes reproducibility of results and treats models and datasets as… | ||
| Too Far Too Furious: Multi-Region Cloud Applications | Denis Magda | 31m |
Demo builds a small pizza microservices system (Kitchen + Tracker, Docker Compose locally, REST APIs) and deploys it to Google Cloud VMs in US East, Frankfurt… | ||
| Lessons Learned: Porting a Streaming Data Pipeline from Scala to Rust. | Evan Chan | 29m |
Conviva ports a mission‑critical streaming pipeline (handling ~5 trillion events/day) from Scala/Akka to Rust and reports concrete wins: a prototype using… | ||
| Taming the Beast - Better Bazel in IntelliJ | Justin Kaeser | 31m |
JetBrains describes a new BSP-based IntelliJ plugin that integrates Bazel monorepos with the IDE and is in an open beta for JVM and some Python projects. It… | ||
| Unlocking the Power of Data-Centric AI. | Fabiana Clemente, Gonçalo Martins Ribeiro | 26m |
Argues that improving data, not models, is the lever for better ML and cites ChinChilla as evidence that better data beats bigger models; recommends… | ||
| Harnessing the Power of Dual Query Engines. | Lei Gao, Jeff Xu | 27m |
Workday built a dual-query platform combining Trino/Presto for interactive SQL and Spark/Livy for ETL and ML, fronted by a custom data access service exposing… | ||
| Keys to Data-Centric AI: Asking AI to Improve its own Dataset. | Jonas Mueller | 33m |
Claims that using a baseline model plus simple algorithms finds and fixes dataset problems so models improve without code changes; open source clean laab… | ||
| Boost LLMs with Event Streaming and Retrieval Augmented Generation | Mary Grygleski | 35m |
Claim: Retrieval-Augmented Generation (RAG) cuts hallucinations and is far cheaper than re‑training foundation models while letting LLMs use up‑to‑date data… | ||
| Pragmatic, scalable, and ergonomic CI/CD using Github Actions and K8s. | Alex Klibisz | 30m |
Describes migrating Tesla Energy CI/CD to GitHub Actions with autoscaling Kubernetes runners using actions-runner-controller, standardizing CI/CD with reusable… | ||
| How to make ML teams 80% more productive. | Emmanuel Turlay | 27m |
Claims an 80% productivity gain by automating an end‑to‑end continuous learning loop: petabyte‑scale data processed with Terra (Apache Beam + Google Dataflow)… | ||
| Revolutionizing Text Transformation. | Wangda Tan, Yingjun Wu | 29m |
Claims an LLM-based pipeline that goes beyond standard SQL to produce text-to-streaming-SQL and supports RisingWave for window joins, plus Python/dataframe and… | ||
| Building private and secure applications easily. | Ryan Orendorff | 34m |
The Sunscreen FHE compiler compiles Rust into programs that run on encrypted data, auto-generating BFV parameters and a runtime that produces keys and… | ||
| Open-Source AI: Community is the Way. | Alexy Khrabrov | 31m |
Chair of the new Generative AI Commons at Linux Foundation for AI & Data announces an open-badging plan (bronze/silver/gold) tied to weights, training scripts… | ||
| Oblivious Privacy in a Statically Typed World | Sinan Pehlivanoglu | 27m |
A Scala compiler plugin called Harpocrates enforces privacy policies at access points rather than via heavy IFC annotations by removing raw data constructors… | ||
| PyTorch 2.1 - New Developments | Supriya Rao | 31m |
PyTorch 2.1 shows concrete speedups by combining torch.compile, new SDPA kernels, export-based quantization and sparsity: benchmarked pipelines report up to… | ||
| Open Source in Space. | Ivan Perez | 32m |
Two open-source tools, Copilot and Ogma, implement runtime monitoring for aerospace: Copilot generates hard real-time C with constant memory and time from a… | ||
| Practical Pipelines: A Houseplant Alerting System with ksqlDB. | Danica Fine | 31m |
Build uses a Raspberry Pi and cheap I2C capacitive moisture sensors feeding Avro-encoded readings into a managed Apache Kafka cluster, with ksqlDB doing… | ||
| Levanter: Legible, Scalable, Reproducible Foundation Models with JAX | David Hall | 24m |
Levanter is a JAX framework built for legible, scalable, reproducible foundation-model training using Haliax named tensors to eliminate shape-broadcast bugs… | ||
| How to help yourself and ML community make PyTorch better. | Nikita Shulga | 22m |
Practical steps for fixing or extending PyTorch: create a small reproducible case, run the built-in collect environment to capture Python, hardware and… | ||
| Recipe for Building a Discoverable and Governed Data Platform. | Sherin Thomas | 27m |
Claims a working recipe: uses DataHub to catalog Snowflake, Looker, Terraform, Airflow, Kinesis and custom configs, and builds column-level lineage that… | ||
| Data management for enterprise-grade foundation models. | Guang-Jie Ren | 29m |
IBM describes an enterprise pipeline that ingests web snapshots (eg Common Crawl ~1 TB per snapshot, 30–40 snapshots), performs large-scale deduplication… | ||
| Why do LLMs hallucinate? | Ofer Mendelevitch | 32m |
Explains hallucination as two concrete failure modes: missing facts inside a model’s implicit knowledge graph or wrong/noisy source text, and cites RLHF used… | ||
| Bridging Python and Lakehouse worlds with Arrow Flight and DuckDB. | Jim Dowling | 28m |
A new service replaces JDBC+Spark with Arrow Flight and DuckDB to stream Parquet/Hudi data as Arrow buffers directly into Pandas/Polars, avoiding column-to-row… | ||
| Apache Iceberg. | Dipankar Mazumdar | 34m |
Apache Iceberg replaces Hive’s directory-level tracking with file‑level tracking, storing table state in a catalog plus metadata files, manifest lists and… | ||
| Cross-company data exchange for the cloud. | Pardis Noorzad | 23m |
Current enterprise exchanges use ad hoc CSV/SFTP drops, shared credentials, APIs, SDKs, Snowflake/Redshift sharing, CDC and reverse ETL, and marketplaces… | ||
| Streaming Graph Search with Quine. | Ryan Wright | 32m |
Quine is an open source, event‑triggered streaming graph database that keeps standing query patterns live and pushes updates as data arrives; a demo shows live… | ||
| Aging: Evolving Software, Tools, and Ourselves. | Jeremy Smith | 29m |
Defines technical debt as anything that should change but hasn't and argues it accumulates both from internal hacks and external change, creating… | ||
| Your Pipeline as a Function. | Pete Hunt | 28m |
A minimal Python pipeline models a jaffle shop using duckdb, S3 and Dagster concepts: pure-ish functions return DataFrames, memoization to S3 serializes cached… | ||
| Challenges at the intersection of ML. | Louis Brandy | 28m |
Spam-fighting work at Facebook and Rockset shows that real-time ML needs both low query latency and low data latency, so classifiers run synchronously and must… | ||
| Can AI Write Its Own Story? Unveiling the Power of Self-Documenting AI. | Manasi Vartak | 28m |
Claims that GPT can auto-document models get tested end to end: a weekend prototype becomes a production feature after about 3 months, after running 2–3 dozen… | ||
| Ten years of building open source standards. | Julien Le Dem | 35m |
Story of three concrete projects and what worked: a columnar on-disk Parquet file format started 2012 with Twitter and Cloudera merging designs, enters Apache… | ||
| Better Code Design with Types and Concepts. | Tikhon Jelvis | 30m |
Argues that static types work as a design tool not just bug catchers, using Haskell and Python examples to show types give fast feedback and shape concepts… | ||
| Implementing a Multi-Tenant Generative AI Chatbot. | Chris Fregly | 32m |
Live demo shows a working multi-tenant RAG chatbot deployed to EKS that uses LangChain + a FastAPI “rag API”, Amazon Bedrock for models (Anthropic Claude for… | ||
| Keynote: Things Databases Don't Do (but should). | Gwen Shapira | 34m |
Lists concrete features databases still don't handle: built-in soft deletes with retention, native tenant awareness for isolation/restore/geo placement… | ||
| Fireside chat | Chris Lattner | 38m |
History of LLVM Clang Swift and MLIR leads to a new stack: a compiler-first language built on MLIR that keeps python compatible syntax while adding optional… | ||
| Keynote: LLMs, what are they good for? | Leah Mcguire | 29m |
Builds a Query Helper that maps natural-language questions to the correct tables in a 58-table schema, using LLMs for semantic search and prompt engineering… | ||
| Developer Productivity – DIY (with ChatGPT). | Baruch Sadogursky | 32m |
Live demo shows ChatGPT and custom GPT builders misproducing Gradle config but also giving actionable steps: enabling parallel test forks and local/remote… | ||
| Federated Learning from simulation to production | Isaac Yang, Yuan-Ting Hsieh | 30m |
NVFLARE is an Apache‑2 open source SDK that turns PyTorch/TensorFlow/Lightning training into federated jobs with 1–3 lines of code to patch trainers or add a… | ||
| Latency goes sub-second in Apache Spark Structured Streaming. | Boyang Jerry Peng | 32m |
Project Lightspeed implements offset management and moves progress tracking, state checkpointing and log purging off the critical path by making them… | ||
| PANEL - Streaming. | Oli Makhasoeva, Zander Matheson, Jerry Peng, Tim Berglund, Dean Wampler, Mary Grygleski | 01h 14m |
Panel surveys production streaming: why it matters, how teams trade latency, cost and correctness, and which tools actually ship. Examples include using a… | ||
| In Process Analytical Data Management with DuckDB | Alex Monahan | 27m |
DuckDB is an in‑process analytical SQL engine that runs inside Python, R, JS or the browser (WASM) and reads Pandas/Arrow/Parquet in place, making data… | ||
| Cost and Performance Optimization of LLM Inferencing. | Chris Matteson | 24m |
Claims that avoid buying big GPUs: use prompt engineering and caching to cut calls to expensive models, plus cheap-model-first cascading and vector similarity… | ||
| Puzzles and Problem Solving with Scala Tooling. | Stewart Stewart | 29m |
Reports using Scala 3 worksheets and Metals as a live, editable REPL: hundreds of hours of recorded puzzle sessions surface recurring mistakes and teach the… | ||
| Full Stack Scala Web Application. | Pathikrit Bhowmick | 29m |
Builds a working full‑stack Scala web app using Scala.js for the frontend, Scala CSS for styling, uPickle for JSON, SBT fastOptJS for live dev and CDN… | ||
| Navigating a Mined World: Repositories, Registries and Artifact Storages. | Melissa McKay | 32m |
Shows concrete risk metrics and fixes: Log4Shell lives in 70,000 direct and 174,000 transitive Java projects, Equifax cleanup tops $1.4B and average breach… | ||
| Efficiently serving LLMs at scale. | Nick Hill | 26m |
Deploys a 176B Bloom (≈350 GB of weights) across multi-GPU nodes using DeepSpeed tensor parallelism, a Rust front end with Python shards from Hugging Face… | ||
| Open Source Warehouse Native A/B Testing. | Graham McNicoll | 26m |
A claim that companies can run production A/B testing directly on their data warehouse using GrowthBook is demonstrated with a Docker container that connects… | ||
| What problems should enterprises actually trust an LLM to solve? | Nate Slater | 31m |
Flip AI builds a Devops LLM for root-cause analysis that interrogates observability stacks (CloudWatch, Datadog, Splunk) and returns hypotheses with… | ||
| Streaming Feature Extraction for a ML Pipeline using Memphis.dev and Bytewax. | RJ Nowling | 20m |
Shows a working pipeline that uses Memphis.dev as durable event queues and Bytewax for Python stream processing to do online feature extraction instead of… | ||
| Building scalable distributed systems powered by Dapr OSS. | Kendall Roden | 31m |
Dapr is an open source distributed application runtime that runs as a sidecar and exposes V1 APIs for pub/sub, state, secrets and workflows so application code… | ||
| Five Years of Cloud Native Rust. | Alex Leong | 25m |
Linkerd adopts Rust for the proxy data plane (chosen in 2018) and reports concrete wins: Linkerd data plane in Rust yields latencies near application baseline… | ||
| DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines | Omar Khattab | 26m |
DSPy provides a compiler that turns declarative signatures and parameterized modules into self-improving pipelines that automatically generate or fine-tune… | ||
| PANEL - Data, Code, and Life with LLMs. | Alexy Khrabrov, James Cham, Pete Skomoroch, Ryan Orban, Manasi Vartak, Miguel Bernadin, Eric Peter | 01h 00m |
One-hour debate argues whether new billion‑dollar startups in five years come from open or closed models, citing concrete evidence: Llama 2’s released weights… | ||
| Streamline Your Development Workflow | Salar Rahmanian | 27m |
Demos a concrete workflow: uses Nix package manager, nix-shell and nix flakes to create reproducible shells, shows a recorded demo that installs Python and… | ||
| Parallel Processors: Past & Future Connections Between LLMs and OS Kernels. | Charles Frye | 33m |
Argues that Transformers are a new kind of computer — a differentiable natural language processor whose core is content‑addressable vector memory rather than… | ||
| Open-Source OLAP on Open-Source Data Warehouses. | Shea Hawkins | 18m |
Combines Iceberg on S3 as an ACID, cheap partitioned data lake with OLAP engines to get subsecond analytics by caching partitions; Delta Lake benchmarks run… | ||
| Using LangChain and LlamaIndex to customize LLMs... | Bassem Yacoube | 29m |
Demos show a working RAG pipeline: ingest a 15‑page Sherlock Holmes PDF, chunk and embed text, store vectors (FAISS/Chroma), and answer queries with a… | ||
| Gaming at Scale: What we learned from building our own multiplayer environment | Kerim Satirli | 30m |
Engineers build a multiplayer Minecraft environment that uses Nomad to run Node.js bot containers, Terraform to deploy across clouds and regions, and Datadog… | ||
| Hamilton:Natively bringing SWE best practices to python data transformations | Stefan Krawczyk | 31m |
Hamilton is an open‑source Python micro orchestration framework that turns column transforms into declarative functions, builds a DAG from function names and… | ||
| Scale By The Bay is back! | Alexi krabrov, Olly mursoever | 02m |
Annual in-person conference returns Nov 13–15, Lake Merritt with a clear agenda: four tracks — software engineering, Cloud Technologies, data pipelines for… | ||
| nothing matches | ||