| | Debugging Airflow Dags shouldn’t feel like guesswork. This free eBook walks through a practical debugging process for Dags, tasks, connections, and external systems, with an emphasis on finding root causes instead of chasing symptoms.
*Message from this week's sponsor, Astronomer |
| | |
| | 🔗 Note: if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-580.html |
| | |
| | In many real-world datasets, numbers are much more likely to begin with 1 than 9. This interactive explainer shows why, then tests the pattern across Fibonacci numbers, populations, GDPs, mountain heights, and other datasets. The raw data, collection scripts, and API queries are included, making the whole exploration reproducible.
Vatsal Bakshi |
| | |
| | Run the same SQL query over the same rows and you should get the same answer. Unless you’re summing floating-point values. In this post, Radim Marek shows how parallel aggregation can produce slightly different totals, why disabling parallelism only gives you a consistently wrong answer, and what to do when the numbers actually need to reconcile.
boringSQL | Radim Marek |
| | |
| | AI chats feel like conversations, but they’re better understood as editable transcripts that are handed to a fresh model on every turn. Once you see them that way, a lot of useful tactics start to make sense, from cleaner prompts and better critiques to multi-model workflows you can use in real work.
Lon Riesberg |
| | |
| | Seasonal adjustment is often the easy part. First you have to find the data. Peter Ellis builds a monthly tourism series for Samoa from a mixture of Excel files and PDF reports, tests the extraction carefully, then uses R’s seasonal, fable, and feasts packages to separate the underlying trend from seasonal noise.
free range statistics | Peter Ellis |
| | |
| | Recovered Factory is a public workshop for data journalism, where messy datasets, open methodology, reusable tools, and honest notes about what worked and what didn’t all sit in plain view. Their dispatches are especially relevant for people who care about practical data work, from PDF pipelines and geospatial stacks to MCP-powered data exploration.
// sponsored |
| | |
| | | | This open textbook on anomaly detection covers everything from simple outliers to unusual patterns in time series, networks, images, and text. It also explains why some popular techniques can be misleading. It's a work in progress, but there's a lot here to start with.
Rob J Hyndman |
| | |
| | | Typst is becoming a serious alternative for technical publishing, but charts have often needed to be created somewhere else. Gribouille brings a ggplot2-style Grammar of Graphics directly into Typst, with layers, scales, facets, smoothers, themes, and a surprisingly extensive example gallery. It’s young, but already quite capable.
Mickaël Canouil |
| | |
| | AI agents can write chart specifications, but the results are often brittle, verbose, or awkwardly designed. Flint takes a different approach. Agents describe the chart using a compact semantic language, and Flint handles details like scales, labels, spacing, and layout before compiling it to Vega-Lite, ECharts, or Chart.js.
Microsoft Research |
| | |
| | | Foundation models are moving beyond language, images, and time series. Google’s TabFM applies the same zero-shot idea to tabular data, using examples from a new dataset as context rather than training a model from scratch. It handles classification and regression, works through a scikit-learn-style interface, and comes with open code and pretrained weights.
Google Research | Weihao Kong and Abhimanyu Das |
| | |
| | | This is one of those projects that sounds like a joke until you see how thoroughly it’s been built. DOOMQL runs an entire Doom-like game inside SQLite, complete with enemies, collisions, combat, rendering, and a live view of the SQL behind each action.
GitHub | doomql |
| | |
| | | This week's issue sponsored by: | | |
|
