News for hackers. Stuff that matters.

nuggetwheat

For most of these systems, the answer will depend on how much hardware (disks and RAM) you're throwing at the problem. The RAM-to-disk ratio is a big factor. If your entire data set fits in memory, the latency will be measured in microseconds. If none of your data fits in memory and each request requires a disk seek, the average latency will be closer 10 milliseconds (assuming there are enough disks to keep up with the requests). Take a look at average latency section in the following benchmark comparing Hypertable with HBase: http://www.hypertable.com/pub/perfeval/test1/ Hypertable average latency goes from about 1.5 milliseconds down to about .5 milliseconds. These numbers are average, however, some requests hit disk and were close to 10 milliseconds, but many requests were served out of the cache. The read request pattern also makes a big difference. Read patterns with locality (e.g. Zipfian distribution) will have much a much better cache hit percentage. -Doug