On Jul 28, 2:52 am, Henrik Schröder <[email protected]> wrote: > I don't know, I'm just guessing. :-) But there's really only two scenarios > that are possible: > > 1) Ajeet's test is broken, he doesn't really store 5 million unique items, > and fails to detect the cache misses. >
This should be impossible. Each key is guaranteed to be unique ( otherwise there would be a lot of other breakages. ) > 2) Stats is broken, Ajeet stores 5 million unique items, but the server > counts them incorrectly. This is possible. Is it true that ( total_items - curr_items == evictions ) ? If so, then the stats obtained in the experiment are proof enough. If this is not the case though, then something more intricate is going on. > The two explanations he offered in his original email are both false. > Spymemcached doesn't hash keys before sending them to the server, so his > client doesn't cause key collisions. And a few items are not evicted because > his eviction stat is 0, and his test doesn't show any cache misses. I never mentioned any hashing by Spymemcached. Its the hashing in memcached that I was having questions about.
