I think the below code from memcache/apr_memcache.c::apr_memcache_find_server_hash_default can cause a cache inconsistency. If the server determined by the provided hash value is dead we just choose the next one. If we do this when writing and the previous dead server comes back with its cache (cannot happen if it got restarted, but can happen if there was a temporary network connectivity issue) the next read operation might read a stale cache entry from this resurrected dead server.
do { ms = mc->live_servers[h % mc->ntotal]; if(ms->status == APR_MC_SERVER_LIVE) { break; } else { if (curtime == 0) { curtime = apr_time_now(); } #if APR_HAS_THREADS apr_thread_mutex_lock(ms->lock); #endif /* Try the dead server, every 5 seconds */ if (curtime - ms->btime > apr_time_from_sec(5)) { ms->btime = curtime; if (mc_version_ping(ms) == APR_SUCCESS) { make_server_live(mc, ms); #if APR_HAS_THREADS apr_thread_mutex_unlock(ms->lock); #endif break; } } #if APR_HAS_THREADS apr_thread_mutex_unlock(ms->lock); #endif } h++; i++; } while(i < mc->ntotal); Regards RĂ¼diger