GitHub stars
A star is a bookmark, not a vote
On GitHub, starring a project is the rough equivalent of bookmarking it: a small public signal that someone found it worth remembering. It is the only feedback most open source gets, so it ends up carrying more weight than it can really bear — in README badges, in funding decisions, in whether a maintainer feels like continuing.
I write DuckDB extensions, and across 102 repositories they've collected 1,869 stars since 2014. This page is an attempt to ask whether that number still means anything, by looking at the one thing a running total can never show: the rate.
Cumulative counts are flattering by construction — they only go up, so they always look like success. Arrivals per month can fall while the total climbs, and here they are falling.
Report compiled September 18, 2026.
GitHub stars per month
Bars are new stars arriving each month. The grey line is the running total, drawn on its own scale.
Watch them disagree. The line rises smoothly the whole way across, which is the version of this story that gets put in a slide deck. The bars peak and then decline. Both are true; only one is informative.
What a new project earns in its first 90 days
The previous chart has an innocent explanation: a portfolio that stops growing stops attracting attention. This one removes that excuse. For every repository, count the stars it earned in its first 90 days, then group by when it launched — each bar is the median for that half-year.
It controls for portfolio size, and asks the question that matters to anyone deciding what to build next: does releasing a new thing still reach anybody? A falling line means the constraint has moved from making things to being found.
One family of repositories is excluded — I published a protocol as a hundred-odd separate repositories over a couple of months, and counting each as an independent launch would drop every median to zero for bookkeeping reasons rather than reception. The count under each bar shows how many launches it rests on; the recent half-years are thin, and the most recent will shift as those projects age past 90 days.
Which projects are still moving
Each project's last 90 days set against the 90 before that. Together the charts above describe the portfolio; this one shows that the average hides real divergence — some projects are genuinely picking up while others have gone quiet, and the aggregate rate is the sum of both.
| Repository | Total | 90d | Prior 90d | Trend |
|---|---|---|---|---|
| airport | 349 | 8 | 14 | cooling |
| quackscale | 24 | 6 | 18 | cooling |
| cronjob | 55 | 5 | 3 | accelerating |
| datasketches | 52 | 5 | 4 | accelerating |
| a5 | 17 | 5 | 0 | accelerating |
| stochastic | 28 | 4 | 7 | cooling |
| textplot | 28 | 4 | 2 | accelerating |
| redis | 17 | 4 | 1 | accelerating |
| httpserver | 285 | 3 | 9 | cooling |
| clickhouse-sql | 94 | 3 | 4 | cooling |
| crypto | 31 | 3 | 1 | accelerating |
| adbc_scanner | 21 | 3 | 7 | cooling |
| marisa | 16 | 3 | 2 | accelerating |
| bitfilters | 10 | 3 | 1 | accelerating |
| vgi-matchrecognize | 3 | 3 | 0 | accelerating |
| openprompt | 62 | 2 | 2 | steady |
| radio | 44 | 2 | 6 | cooling |
| fuzzycomplete | 30 | 2 | 2 | steady |
| evalexpr_rhai | 27 | 2 | 0 | accelerating |
| clickhouse-native | 22 | 2 | 2 | steady |
| rapidfuzz | 19 | 2 | 2 | steady |
| vgi-rpc-python | 14 | 2 | 4 | cooling |
| hashfuncs | 14 | 2 | 0 | accelerating |
| inflector | 10 | 2 | 1 | accelerating |
| jsonata | 8 | 2 | 3 | cooling |
| json_schema | 5 | 2 | 1 | accelerating |
| vgi-rpc-java | 2 | 2 | 0 | accelerating |
| vgi-fixedformat | 2 | 2 | 0 | accelerating |
| shellfs | 96 | 1 | 3 | cooling |
| lindel | 67 | 1 | 6 | cooling |
| tributary | 57 | 1 | 2 | cooling |
| duckdb-cron-extension | 29 | 1 | 0 | accelerating |
| geosilo | 25 | 1 | 24 | cooling |
| webmacro | 16 | 1 | 0 | accelerating |
| quickjs | 15 | 1 | 3 | cooling |
| pcap | 14 | 1 | 1 | steady |
| tera | 9 | 1 | 1 | steady |
| tsid | 7 | 1 | 0 | accelerating |
| vgi-rpc-typescript | 2 | 1 | 0 | accelerating |
| vgi-quant | 2 | 1 | 1 | steady |
| vgi-java | 1 | 1 | 0 | accelerating |
| vgi-rust | 1 | 1 | 0 | accelerating |
| vgi-go | 1 | 1 | 0 | accelerating |
| vgi-rpc-rust | 1 | 1 | 0 | accelerating |
| vgi-rpc-website | 1 | 1 | 0 | accelerating |
| vgi-charset | 1 | 1 | 0 | accelerating |
| vgi-asn1 | 1 | 1 | 0 | accelerating |
| vgi-tiktoken | 1 | 1 | 0 | accelerating |
| vgi-iso20022 | 1 | 1 | 0 | accelerating |
| cupola | 1 | 1 | 0 | accelerating |
| vgi-barcode | 1 | 1 | 0 | accelerating |
| vgi-units | 1 | 1 | 0 | accelerating |
| vgi-symbols | 1 | 1 | 0 | accelerating |
| vgi-compress | 1 | 1 | 0 | accelerating |
| vgi-xslt | 1 | 1 | 0 | accelerating |
| vgi-scholar | 1 | 1 | 0 | accelerating |
| vgi-pdf | 1 | 1 | 0 | accelerating |
| vgi-java-introduction-docs | 1 | 1 | 0 | accelerating |
| vgi-tantivy | 1 | 1 | 0 | accelerating |
| vgi-code | 1 | 1 | 0 | accelerating |
| vgi-image | 1 | 1 | 0 | accelerating |
| vgi-tika | 1 | 1 | 0 | accelerating |
| vgi-poi | 1 | 1 | 0 | accelerating |
| vgi-odata | 1 | 1 | 0 | accelerating |
| vgi-media | 1 | 1 | 0 | accelerating |
| vgi-ical | 1 | 1 | 0 | accelerating |
| airport-docs | 1 | 1 | 0 | accelerating |
| httpclient | 80 | 0 | 3 | dormant |
| python-flight-server | 17 | 0 | 3 | dormant |
| vgi-xgboost | 1 | 0 | 1 | dormant |
| vgi-crontimes | 1 | 0 | 1 | dormant |
How this is measured
GitHub records when each star was given, so this is a reconstruction of every individual star rather than a sample or an estimate. Those timestamps are reduced to per-repository daily counts when the page is built — 1,868 of 1,869 reported stars, last rebuilt September 18, 2026. The top two repositories hold 34% of all stars.
Two limits worth stating. Stars that were later removed are invisible to this method — the API returns only current stargazers, which accounts for the small gap between the reconstructed and reported totals. And GitHub exposes no link between a star and where the visitor came from, so nothing here attributes a star to a referrer; anything that claims to is guessing.