ASAIK, he wrote the original memcached in perl but what everyone is using now is a rewrite of memcached in C by Anatoly Vorobey [1]. Not that it diminishes his efforts or anything, just pointing it out.
I also then ported the C++ memcached (called memcacheg) to Go, for profiling Go vs C++, and named that one memcachego. So I've written it about 4 times.
Kidding aside, anything you learned during the rewrites that caused noticeable changes to the design, or it's still the same thing, just transliterated?
How were the results of the profiling? Anything interesting come out?
The others sound like they are mostly used by Google.
- thundering herds suck, so it manages how the cache is filled when a key is missing
- repeatedly retrieving super hot values will eventually exhaust the network pipe of whichever server held, so groupcache replicates hot values to more than one cache server
Since this was built for dl.google.com, the 'value' might be big (say, a chunk of the Chrome binary) and the request rates for the top cache keys probably get pretty high. So, even if memcached is doing good work for lots of folks, the distinct features of groupcache probably address real needs in the app.
And, yeah, groupcache and the dl.google.com rewrite are Go partly because Brad likes Go, but both projects had reasons to exist quite apart from deploying Go for Go's sake.