Berlin Buzzwords 2025
2025-06-15..2025-06-17
-- online --
[Berlin Buzzwords 2012, Berlin Buzzwords 2015, Berlin Buzzwords 2016, Berlin Buzzwords 2018, Berlin Buzzwords 2019, Berlin Buzzwords 2020, Berlin Buzzwords 2021, Berlin Buzzwords 2022, Berlin Buzzwords 2023, Berlin Buzzwords 2024, Berlin Buzzwords 2025, Berlin Buzzwords 2026]
as the organisers put it
BERLIN BUZZWORDS 15-17 June 2025 @KulturbrauereiBerlin Buzzwords is Europe's leading conference for modern data infrastructure, search and machine learning and is focused on open source software projects. Recordings BERLIN BUZZWORDS 15-17 June 2025 @Kulturbrauerei Berlin Buzzwords is Europe's leading conference for modern data infrastructure, search and machine learning and is focused on open source software [...]
| title | author | length |
|---|---|---|
| Unpacking Digital Sovereignty: How to avoid fueling the nationalist rise | Aline Blankertz | 44m |
Argues that current European “digital sovereignty” plans risk empowering nationalist and corporate concentration rather than democratic control, citing the… | ||
| Go Beyond Basic RAG with Agentic Behavior | Bilge Yücel | 40m |
Claims that extending RAG with agentic behavior lets an LLM decide when to rewrite queries, call tools, and run multi‑step retrieval, demonstrated by a Python… | ||
| Performance Tuning Apache Solr for Dense Vectors | Kevin Liang | 21m |
A 169M-document index of 1,024‑dim vectors starts at ~900–1,000 ms end‑to‑end, 10–11 CPU cores and caps near 50 QPS. Normalizing vectors and switching to dot… | ||
| Melting Icebergs: Direct access to Kafka Data via Iceberg | Tom Scott, Roman Kolesnev | 38m |
A runtime projection presents Kafka topics as Iceberg tables without moving data by batching Avro messages into offset-range “files,” building Iceberg metadata… | ||
| Cross Domain Enterprise Search - Content Diversity at Scale | Atri Sharma, Abhishek Singh | 39m |
An enterprise search platform handles more than ten content types and scales to billions of documents by using dual indexing (BM25 + vector), corpus-aware… | ||
| Advancing Multi-Modal Search Capabilities in Search Pipeline | Dhrubo Saha | 26m |
Demonstrates ML inference processors inserted at ingest/search/response stages of an OpenSearch pipeline that generate embeddings, rewrite queries, and rerank… | ||
| Flink Jobs as Agents – Stream Processing for Agentic AI | Steffen Hoellinger | 44m |
Claims that Flink jobs become autonomous agents by reading Kafka topics, invoking foundation models, and writing actions to output topics are demonstrated with… | ||
| Text Search on Images with Quantized ColPali | Sonam Pankaj | 20m |
ColPali is a late‑interaction vision‑LLM pipeline fine‑tuned on Pali Gemma that embeds whole page images so you can skip OCR, chunking and parsing; relevance… | ||
| Siphon : Modern Data Stack with SF-CH & Iceberg | Ved Prakash | 22m |
Siphon reads Postgres logical replication from a read replica, decodes CDC in Go, and pushes events via NATS JetStream to consumers that write to Snowflake… | ||
| Reproducibility in Embedding Benchmarks | Isaac Chung | 25m |
MTEB pins dataset and checkpoint revisions and version-controls code, but reproducibility breaks from compute costs, prompt sensitivity, and implementation… | ||
| All the DataOps, all the paradigms | Lars Albertsson | 47m |
Choices between Hadoop, warehouses, lakes, streams and meshes change ops cost and iteration speed by orders of magnitude: a 2,000-node dependency graph… | ||
| Analysing Public Kafka Data from NASA Satellites | Frank Munz | 39m |
Shows how NASA’s GCN public Kafka streams feed a Databricks Lakehouse using Apache Kafka, Apache Spark (Spark 4 declarative pipelines), Unity Catalog and… | ||
| How to train a fast LLM for coding tasks | Ivan Dolgov | 34m |
JetBrains trains a purpose-built in‑IDE code completer called Mellum, a ~4B parameter model fine-tuned for fill‑in‑the‑middle code completion and context… | ||
| Visual Literacy: Complex Document Retrieval with VLMs | Isabelle Mohr, Saba Sturua | 37m |
Claims that vision‑language embeddings close the “visual information gap” for charts, tables and complex layouts by using a 3B‑parameter Quen 2.5VL backbone… | ||
| Why Chatbots Still Fail: The Hidden Pitfalls of RAG | Lewin von Saldern | 33m |
Claims that RAG fails because of model choice or prompts miss the main culprit: bad chunking. A medical chatbot built with a vector DB and off‑the‑shelf… | ||
| Mastering real-time anomaly detection with open source tools | Olena Kutsenko | 39m |
Builds a streaming pipeline that trains and retrains time‑series models on historical data in Apache Iceberg, ingests live events via Apache Kafka, processes… | ||
| Streamlining Search Quality: Search Relevance Workbench | Stavros Macrakis, Eric Pugh | 44m |
Search Relevance Workbench (SRW) is an open source, engine-agnostic suite that collects click streams using the User Behavior Insights standard, ingests… | ||
| What you see is what you mean; intent based ecommerce search | Dennis Berger, Marco Petris, Volker Carlguth | 41m |
Otto integrates an LLM upfront to turn a query into intent clusters that drive retrieval, reducing backend fan‑out from 50→5 queries and preserving context… | ||
| Anatomy of Table-Level Locks in PostgreSQL | Gülçin Yıldırım Jelínek | 39m |
Explains how PostgreSQL’s MVCC avoids row-level blocking but DDL takes table locks, with access exclusive conflicts everything illustrated by ALTER TABLE… | ||
| Performance & Fault Tolerance: Building a Modern Database | Guy Shtub | 41m |
ScyllaDB claims x10 throughput over Cassandra by being written in C++ and running shard-per-core so each core owns CPU IO and storage, avoiding GC pauses that… | ||
| Which GPU for Local LLMs? | Radu Gheorghe, Rafał Kuć | 22m |
Benchmarks run with llama.cpp compare consumer and datacenter GPUs, showing H100 tops generation at ~300 tokens/s while an RTX 4090 often beats H100 on prompt… | ||
| Mixture of Encoders: A Vector-Native Approach to Search | Filip Makraduli, Ben Gutkovich | 19m |
Mixture of Encoders maps text, numbers, categories, images and temporal signals into a single Mixture of Encoders embedding so queries and metadata search… | ||
| Zero to Scale: Telemetry pipeline with Apache Cassandra | Shikhar Srivastava, Nomin-Erdene Oyun | 19m |
Processes send tens of billions of messages a day (≈40 TB, peaks >1M/min) and a telemetry pipeline is built using ZeroMQ proxy to funnel thousands of local… | ||
| State of native access in Apache Lucene | Uwe Schindler | 43m |
Lucene 10 adopts Java 21 and uses Project Panama to get true off-heap access via MemorySegment and fast foreign calls; the memory-mapped directory now issues… | ||
| AI-Powered Search Results Navigation with LLMs & JSON Schema | Ilaria Petreti, Anna Ruggero, Edward Lambe | 41m |
An AI Filter Assistant for SDMX data is implemented with Python FastAPI, OpenAI GPT-4.1 (comparison vs GPT‑4 and GPT‑4o‑mini), and Apache Solr; it builds… | ||
| Self-hosting AI LLMs - a beginners guide | Nick Burch | 20m |
Finds models on Hugging Face (not GitHub), targets about 1–7B parameters for laptops, and uses 4-5 bit quantization to cut model size ~4x with little quality… | ||
| Accelerating QuestDB: Lessons from a 6x Performance Boost | Javier Ramirez | 40m |
QuestDB (Apache‑2.0) moves from being 33× slower on ClickBench to roughly 3× and reports up to a 6x performance boost on some queries by adding SIMD-style… | ||
| Building a knowledge graph for climate policy | Harrison Pim, Fred O'Loughlin | 42m |
An open-source climate policy knowledge graph maps ~1,500 curated concepts across 2M pages and 45M passages and is already powering a search used by ~350K… | ||
| Shipping Lucene 10.0, 25 years in the making | Adrien Grand, Luca Cavanna | 37m |
Apache Lucene 10.0 lands after a monthslong push that surfaces specific wins and snafus: sparse indexing and measurable CPU/IO improvements, minimum Java… | ||
| Quiet on Set: Building an On-Air Sign with Open Source Tech | Danica Fine | 43m |
Builds a working on‑air sign: a Raspberry Pi plus a 64x64 RGB LED matrix and a matrix HAT that requires soldering two connections, driven within a second of… | ||
| Going Local-First: A Primer | Miloš Sutanovac | 24m |
Local‑first starts from a 2019 paper and promises fast no spinners UIs, multi‑device continuity, offline‑first behavior and user ownership of data. Practical… | ||
| The Dark Secrets of Stream Processing | Yingjun Wu | 40m |
RisingWave is a Postgres‑compatible distributed SQL streaming DB that adopts Backfilling Takes Forever, Consistency causes system stalls, and S3 costs spike… | ||
| gamma_flow: Denoise, classify and disentangle spectral data! | Viola Rädle, Raphael Franke | 18m |
gamma_flow is an open-source Python package that performs real‑time analysis of gamma spectra by using a supervised non‑negative matrix factorization to map… | ||
| Precision farming powered by K3s and TensorRT | Wieneke Keller, Sebastian Lenartowicz | 31m |
Aurea Imaging builds a TreeScout: a rugged Jetson edge box running Python microservices orchestrated with K3s on Jetson, converting PyTorch YOLO models to… | ||
| "Do What I Mean": The History of AI and Program Synthesis | William Benton | 41m |
A concise lineage from 50 years of program synthesis to today's LLM coding shows concrete tools and results: FlashFill and successor FlashMeta use backward… | ||
| More Than Just The Tip Of The Iceberg | Michal Gancarski | 01h 09m |
Hands-on workshop that sets up a local Apache Iceberg environment with Docker Compose stack, MinIO object storage, LakeKeeper REST catalog, PySpark, pyiceberg… | ||
| miniCOIL: Sparse Neural Retrieval Done Right | Evgeniya Sukhodolskaya | 35m |
miniCOIL adds semantic weights to BM25 by reserving consecutive indices per token to store 4-dimensional meaning vectors and uses a one-layer dimensionality… | ||
| Evolution of Uber's Search Platform | Yupeng Fu | 38m |
Search at Uber moves from Elasticsearch to a custom Lucene-based engine (called Sea) with a live concurrent index plus snapshot and batch base-index rebuilt by… | ||
| FlinkCDC: Streamlining your data analytics pipelines | Muhammet Orazov | 42m |
FlinkCDC is an Apache Flink component that reads DB binlogs (MySQL/Postgres via an embedded Debezium client or native MongoDB readers), applies an incremental… | ||
| Data Quality Management: The Good, The Bad, and The Messy | Jan Meskens | 41m |
Shows concrete failures like undelivered packages from wrong addresses and wrong salaries from bad timesheets then traces causes to pipeline bugs, missing… | ||
| From Culture to Open Source: Build Value-driven Communities | Marion Nehring, Jessie de Groot | 41m |
Company Weav8 links product adoption to a people-first culture, arguing that kindness and collaboration drive developer engagement; practices include regular… | ||
| When StatefulSets are not enough | Manish Gill | 35m |
Kubernetes StatefulSets block safe vertical resizing and zero-downtime autoscaling for ClickHouse Cloud, so the team builds a make-before-break flow by… | ||
| Best Practices for Running Databases on Kubernetes | Peter Zaitsev | 45m |
Argues Kubernetes is ready for databases if you apply concrete practices: use Percona Operators to automate replication, failover, backups and lifecycle; store… | ||
| From Search to Insight: Leveraging OpenSearch for Scalable, AI-Driven Search Experiences | Saurabh Singh | 39m |
OpenSearch 3.0 delivers concrete changes: Lucene 10 and JVM 21 yield about 20% faster queries than 2.x and 10× throughput versus 1.x, while ANN/HNSW vector… | ||
| Qumat: Apache Mahout Quantum Compute | Trevor Grant, Andrew Musselman | 34m |
Mahout adds a new Qumat layer that treats quantum circuits as complex-valued matrix chains so algorithms written once run on multiple backends (qiskit, cirq… | ||
| How I Sidestepped ‘Being Glue’ | Fatima Taj | 31m |
Migration projects run as 95/5 AB rollout experiments using Jupyter notebooks owned by stakeholder teams, with debugging and parity checks stretching one to… | ||
| Contexts & Machines: How Document Parsing Shapes RAG results | Alessio Vertemati, Andrea Ponti | 22m |
Experiment parses five ~100‑page technical reports, generates ~200 LLM Q&A pairs with RAAS, and compares pypdf, PDFact, unstructured and llama pars combined… | ||
| Exploring reranking depth in modern search pipelines | Athanasios Papaoikonomou | 40m |
Experiments use BM25 + Elasticsearch index and multiple cross‑encoder rerankers evaluated on BEIR and similar datasets with NDCG@10, measuring rerank depth K… | ||
| Flavors of PostgreSQL® and you: how to choose a Postgres | Celeste Horgan | 37m |
Compares concrete systems: Google AlloyDB Omni (forked Postgres with a columnar engine and extended vector indexes, Docker image that’s a black box and thin… | ||
| What’s New in the OpenSearch Project and Ecosystem | Dotan Horovits | 36m |
OpenSearch 3.0 ships with Lucene 10, JDK 21 and Node 14 minimums, moves to foundation governance, and claims performance gains (about 20% aggregate vs 2.x and… | ||
| Harnessing AI to strengthen trustworthy information | Lucian Precup, Giovanna Monti | 43m |
Demonstrates a Lucene‑based A2 search stack and an integrated RAG assistant that returns sentence‑level citations, plus experiments with Wayback Machine… | ||
| Airflow 3 - the new beginning | Jarek Potiuk | 40m |
Airflow 3 is released (current 3.0.2) after ~8–9 months and a 20–50 person effort, claiming to fix long-standing pain points while preserving backward… | ||
| A decade of lessons in Open Source licensing | Justin Mclean | 39m |
Over a decade and more than 1,000 releases (actually 1,647 threads voted), the talk gives concrete patterns: common breakages are missing or incorrect license… | ||
| End-to-End Semantic Search with Apache Solr 9.8 LLM Module | Alessandro Benedetti | 45m |
Apache Solr 9.8 adds an LLM module that lets Solr call external embedding services (OpenAI, Cohere, Mistral, HuggingFace via LangChain4j) to encode text at… | ||
| Apache Iceberg ingestion with Apache NiFi | Lester Martin | 39m |
Shows NiFi flows that pull from Kafka or files, convert records to Avro, write to S3 and register files in Iceberg via Trino with Starburst Galaxy used for… | ||
| Observability for All! | Carly Richmond | 40m |
Shows how to correlate browser and backend signals by instrumenting a demo record-store app with OpenTelemetry JS SDK and the Elastic RUM agent, sending traces… | ||
| Breaking Search For Fun and Profit | Charlie Hull | 20m |
Shows concrete breakages on real sites: London School of Economics search returns nothing for "accommodation" while a working student-accommodation site… | ||
| Hybrid search on hybrid models, at scale | Radu Pop, Pietro Mele | 49m |
Extensible hybrid search runs on Elasticsearch multi-index architecture with OpenVINO inference, Spring Boot microservices and Kubernetes; Whisper… | ||
| How [not] to evaluate your RAG | Roman Grebennikov | 40m |
A production RAG for a fintech support-agent assistant is built on a self-hosted Mixi/nixie search backend, with documents converted to Markdown and a… | ||
| Vespa.ai’s Personalized Search: Advanced Ranking & Tensor framework | Piotr Kobziakowski | 40m |
Vespa demonstrates a working e‑commerce demo built in days that combines LLMs (lava, GPT‑4o style prompts) to generate item descriptions and sparse tensor… | ||
| Delay accounting: an underrated feature of the Linux kernel | Nikolay Sivko | 19m |
Kernel exposes per-task wait-time counters (delay accounting) that measure how long a runnable task sits in the run queue, accessible via netlink socket to… | ||
| nothing matches | ||