Can We Double HTTP Client Throughput?
summary
New HTTP client undici claims a concrete speedup: using keep‑alive and pipelining avoids TCP head‑of‑line blocking so measured end‑to‑end throughput doubles and can peak ~3x versus node core http in benchmarks (500 parallel requests, 10 ms target, 50 socket limit). Practical fixes include always using an agent, switching to LIFO agent scheduling to cut ECONRESET/timeouts, and enabling pipelining where datacenter links are reliable; the library is implemented from scratch and shipped.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:50A little bit about me
- 3:21Microservices!
- 3:57The link between microservices
- 4:53TCP 3-way handshake
- 5:58TCP Congestion Window
- 8:52Scenario
- 9:28Benchmark results
- 10:00HTTP/1.1 pipelining
- 11:04Beware of Head-of-Line Blocking
- 11:57Event loop primer
- 13:34Timeouts and ECONNRESETS
related talks
26mCan we Double the Node.js HTTP Client Throughput?
48mBetter performance with HTTP/2
49mHTTP: History & Performance
26mHow Fixing a Bug in NodeJS Accidentally Doubled Performance
48mBetter performances on the server side with HTTP/2
01h 00mHTTP/2 - The future of the web today
50mHigh-performance web apps using HTTP/2
43mPerformance in the World of HTTP 2.0