⭐ We shaped our new repo star history API around production evidence and maintainer feedback. Now it provides trends without returning individual identities or requiring unbounded database work. The practical engineering tradeoffs that made it happen ⬇️

From 2,500 requests to 24: Rebuilding star history without the stargazers

By Taylor Cuilty & Sarita Iyer

Article · @github

Oct 14, 2026 · 10:00 PM UTC

32
14
10
132
38,131
RelevantRecentLikes
The naive "Backend Engineers" celebrate moving from unbounded database scans to time-range filtering as absolute query optimization wizardry, completely blind to the skewed distribution tax. Paginating aggregated history by strict calendar week boundaries without pre-aggregated counts or hyperloglog hyper-sketches guarantees your read tier will choke on extreme traffic spikes when a viral repository spikes 50k stars in a single timezone block. The deterministic query analysis and localized write-heavy database allocation guard: # Realized Mechanics: Query Execution Optimization and Database Scale Guard def verify_query_execution_cost(): # Enforcing hard data processing boundaries over dynamic range aggregations target_weekly_star_density = 55000 # High-throughput viral repository surge scale max_safe_scan_threshold = 10000 is_pre_aggregation_active = False # Systems architecture must block ad-hoc groupings when query density breaches hardware thresholds if target_weekly_star_density > max_safe_scan_threshold and not is_pre_aggregation_active: raise ResourceWarning("Database Saturation: Group-by operation over dynamic sparse nodes has exceeded I/O limits.") if __name__ == "__main__": verify_query_execution_cost()
219
Designing for the question instead of preserving the old payload is the right move. One contract detail I’d make very explicit: this is the history of currently retained stars, not an immutable star/unstar event log. That distinction matters when people compare a launch window months later.
4
finally, star history that doesn’t cost a fortune to fetch 🤝 just dropped a follow
40
trending without identities sounds good, but how real are those trends?
5
interesting to see so much focus on privacy with trends
7
Aggregating server side instead of handing out stargazer lists is a quiet security win too. Those lists fed fake star farms and targeted phishing at maintainers. Is the trend data bucketed so nobody can diff two snapshots and back out who starred?
72
Trends without identities is the right call. Most people only want the curve anyway: is this repo still growing, or did it spike once and stall? Curious if the API keeps daily granularity for older repos or buckets them.
4
The useful line is that returning less data did not make it cheaper, filtering by time before counting did. Worth checking where your own slow endpoints do their filtering.
54
star history finally getting respect
55
the identities were never the data, just how you got the timestamps
41
2,500 requests down to 24 and nobody's identity leaves the building. rare when the cheaper design is also the more polite one
86
The LIMIT/OFFSET part should be required reading for anyone building a reporting page, smaller response and the database still did all the expensive grouping first. Filtering by date before you group is a boring fix and I like it more for that
62
Going from 2,500 requests to 24 by returning trends instead of identities is a good lesson on its own. Aggregating first fixed privacy and performance in one move. Did any maintainers push back on losing the per-user list?
66
the real cost was counting every star to know how many pages. a cursor never needs the total
71
Going from 2,500 requests to 24 is a big cut, but stargazers is just gone now?
38
nice that it gives trends without exposing who starred i used star counts in a study of PR reply times and more stars actually meant slower replies to outside contributors
1
19
Smaller results, less work? I explained this as a manga.
14
returning daily star counts instead of stargazer lists fixes the scraping problem and cuts 2,500 requests to 24
19
The bit about unstars changing the old curve is easy to miss. I'd want that noted on the chart before using it to judge how a launch went.
21
Load more replies