mailwal

Data Berlin #58

mucio from the Data Berlin meetup

Data Berlin #58

The only spooky queries are the one hitting you cloud budget

Oct 27

Yes, we are late. School holidays rearranged the calendar, Halloween rearranged everything else, and this issue lands a few days after it was supposed to. Sorry about that.

This weekend it is all about candy and buckets: fill a bowl, find the plastic pumpkin, and if last year's costume still fits, that counts as a plan. Meanwhile some of the kittens around here have already started their Black Friday training, which tells you more about November than any calendar does.

Before we start: our next meetup is on 3 November. The room is already full, so if you did not get a seat, do not panic — like every time we will stream it online, and the link will appear on our YouTube channel on the day.

And now the usual: what we read, what shipped, what is happening in Berlin, and who is hiring. Issue #58, from here on.

Data Berlin Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

📖Interesting Readings

A field postmortem that follows an intermittent Postgres outage down to a background read-ahead setting almost nobody tunes, and to the mailing-list thread where the risk was flagged years ago and waved through. – Postgres recovery_prefetch Considered Harmful and Causing Outages

How LinkedIn replaced the machinery behind hundreds of billions of daily events while traffic kept flowing, and why the team compared sampled output on both sides before shifting users across. – Rebuilding LinkedIn's Tracking Infrastructure: Scaling, Migrating, and Hardening a High-Volume Event Pipeline

AI-generated charts are usually static pictures; this declarative language lets an agent describe what should happen when a reader clicks a bar, and compiles it into a chart that actually responds. – Flint-interactive: A language for creating interactive data visualizations

A handful of companies are going past fine-tuning and continuing pre-training on their own corpus, and Ben Lorica collects what that looks like from the inside and what it demands of the data team. – What I learned from companies training their own models

A widely used open-source data quality tool is being absorbed into an integration platform, and this reads the deal as a bet that agents will need a trustworthy catalog more than another dashboard. – Boomi acquires Elementary | What it means for the data industry

A long conversation with Will Larson about which unwritten rules stopped being true since the 2010s, and what that has done to the path from individual contributor to CTO. – The revised rules of engineering management with Will Larson

An insider account of how IBM Watson Health absorbed a nimble healthcare data company and ground it into a cautionary tale about buying your way into artificial intelligence. – The Ghost of Watson Past: Why Enterprise AI Hype Still Eats Startups Alive

A dozen projects and two acquisitions later, the person who put a consensus log in front of SQLite in 2014 writes down the timeline the way it happened, starting before anyone called it a category. – Twelve years of distributed SQLite

Everyone bolts a confidence score onto an agent's answer and treats the number as a probability, and this unpacks what it would actually have to mean before anyone trusts it. – Asking LLM for confidence

An agent can fix the bug and quietly move something else; this is how to review a generated change for everything it touched, not just the line it was asked about. – Your AI agent fixed the bug, but what else did it change?

🚀 Shipping Now

  • DuckDB 2.0.0 LTS — The single-file analytical database ships its largest release so far: a remote wire protocol, asynchronous I/O, a native variant type, an agent mode and triggers, out of roughly 20,000 commits since 1.5.

  • Snowflake declarative feature store — Feature definitions can now live as versioned configuration files deployed to the warehouse with a plan-and-apply command, the same week the online serving layer reached general availability.

  • Databricks Genie Code tasks — A prompt can now run as one step of a Lakeflow job, with the agent reading earlier task outputs, calling tools and linking back to its own chat.

  • BigQuery document chunking — A table function splits unstructured documents into coherent passages inside SQL, ready for an embedding model without a separate parsing pipeline.

  • MotherDuck secure shares — Views attached to a share can filter rows, hide columns or mask values for each consumer, so one shared database serves several levels of access.

  • TimescaleDB 2.31.0 — Updates and deletes on the compressed columnstore unpack keys in vectorised batches, speeding up the mutations and the recompression that follows them.

☕ Upcoming Meetups & Events

💼 This Week’s Job Picks

📊 Data Analyst

📈 Analytics Engineer

⚙️ Data Engineer

🤖 AI / ML / LLM + Data Scientist

🧭 Leadership

This is the end.

If you want to see us in person, November 3rd is the date. The room is full, the stream is open, and the talks are the kind you end up arguing about on the way home.

We are also reworking the site at the moment: it is not there yet, and we would rather ship it late than ship it wrong. In the meantime, if you missed any of the past meetups, the recordings are all on our YouTube channel.

It’s time to press the send button
— Data Berlin team

Data Berlin Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

There is absolutely nothing after this line, but in case you are curious and you want to support us, feel free to click.

Absolutely nothing.

Invite your friends and earn rewards
If you enjoy Data Berlin Newsletter, share it with your friends and earn rewards when they subscribe.

Invite Friends

© 2026 mucio