Doubling API concurrency cut load time 14% but introduced call failures, dev reverts change

A developer building Holder Atlas, a token-holder geography tool, found that generating a single world map required roughly 110 sequential Nansen API calls averaging nearly two seconds each, resulting in a cold load time of around 40 seconds. To speed this up, they doubled the concurrency pool from 4 to 8 workers and raised the rate limit from 5 to 10 requests per second during a benchmark test on four tokens in September 2026. The wider configuration improved the median cold load time by only 14 percent, from 40.3 seconds to 34.5 seconds, while the 95th-percentile time barely changed. More critically, the higher concurrency produced three failed API calls that the narrower setup had avoided, causing one token's cached result to differ from its cold-run output. The developer reverted to the original 4-wide, 5 rps defaults, concluding that per-call latency — not worker count — is the true bottleneck.