Fifteen years of data questions

Stack Overflow is a fifteen year log of what people were stuck on. I pulled the yearly question count for fifteen data tags, from SQL and Excel through pandas and Spark to dbt, Polars and DuckDB. The counts rise for a decade and then fall off a cliff. The shares are the interesting part.

What I counted

One API call per tag per calendar year: how many questions on Stack Overflow carried that tag and were asked that year. Fifteen tags, 2011 through 2025, 225 numbers. Questions that were later deleted are not in the count, and a question with two of my tags is counted twice, once in each. That is the whole method.

The counts fell for everything

Total questions across these tags grew every year from 2011 to 2020, even in 2015 when SQL alone dropped by twenty thousand. Then 2021 and 2022 slip by about a tenth each, 2023 halves, 2024 halves again, and 2025 lands at under five percent of the peak.

Every single tag follows that curve, including the ones that were still growing on every other measure. Snowflake questions fell from 3,328 in 2022 to 334 in 2025. dbt fell from 464 to 89. Nobody thinks those tools lost most of their users in three years. The site lost the questions.

This is why I think comparing raw counts across years is now useless for anything after 2022. Compare shares. A share asks a better question: of the people who still turned up to ask, what were they asking about.

What the shares say

Python is about half of everything in every year, between 46 and 58 percent, and I hid it for most of the story because it is a language tag, not a data tool. Of what remains, three things stand out.

SQL eroded and Excel did not. In 2011 SQL was 73 percent of the non-Python questions. By 2020 it was 24 percent and by 2025, 18 percent. Excel sat between 10 and 20 percent for the whole fifteen years. My reading is that SQL questions did not go away, they moved: the same person who asked how to write a join in 2013 was asking how to do it in pandas by 2019.

pandas is the story of the middle decade. Zero in 2011, past Excel in 2018, 50,179 questions in 2020, past SQL in 2021, a quarter of non-Python questions at its peak. R, for comparison, held a flat 26 to 31 percent from 2015 onward and dplyr peaked at 3.

The warehouse band is thin and it only grows. Snowflake, BigQuery and dbt together were under one percent of non-Python questions through 2016 and 6.2 percent in 2025. That share rose in 2023, 2024 and 2025, straight through the collapse. Power BI did the same, from 1.8 percent in 2022 to 5.1 percent in 2025. Spark went the other way, peaking at 7.3 percent in 2016 and settling around 4.

The newest tools

Polars had no questions before 2021 and was 3.1 percent of non-Python questions in 2025. DuckDB went from two questions in 2019 to just under one percent. Both peak in share in 2025, which is the smallest year in the data. That is exactly the pattern you would expect if the people still asking on Stack Overflow are the ones the chat assistants cannot help, because the tool is too new to be in the training data. I would not bet on either share holding once the models catch up.

What to do with this

• Stop quoting Stack Overflow counts as adoption after 2022. Quote shares, and say what the denominator is.
• When you pick tags to track, check the real tag name first. Two of my fifteen returned zeros until I did.
• Rerun the script once a year. The 2025 collapse was fast; whatever replaces the site as a signal will show up as a share change, not a count.
• Treat the warehouse and BI bands as the leading indicator. Along with Polars, DuckDB and Excel, they were among the few shares that grew every year through the fall.