4 ms·
"The old system responded to requests somewhere between two and 10 long seconds; the new one takes 30 milliseconds, on average." Wow, 30 milliseconds. Subtract
by camel_gopher 11y ago
"The old system responded to requests somewhere between two and 10 long seconds; the new one takes 30 milliseconds, on average."
Wow, 30 milliseconds. Subtract out the RTT, and you have app request times on the order of 5ms. Outstanding. And impossible.
- cdelsolar 11y agoWhy impossible?
- ChuckMcM 11y agoWhen Blekko was serving web traffic if it came from the primary index and was cached (a repeat request) it was served up in under a millisecond. I learned at Google to never under estimate what sort of performance you could get reading the result out of the memory of a nearby computer. 10,000 computers with 96GB each on them can easily keep 640TB of data ready to return very very quickly. Its even faster if you have your own zero copy TCP/IP layer (right from client memory into the network interface card). For an example of scale, 10,000 nodes with 96GB each costs less than $30M dollars. The government spent billions on the healthcare.gov project.