Systems can get large enough that it would take a few seconds for GC to complete thereby causing unacceptable latency (especially if you need to meet strict SLA

Systems can get large enough that it would take a few seconds for GC to complete thereby causing unacceptable latency (especially if you need to meet strict SLAs). Consider a caching datastructure that needs to be constantly updated with fresh data. Objects in the cache remain there long enough to get into the final generation only to be pulled off some time later. Without memory management (e.g. object pooling for one), you'd quickly fill up your long lived generation causing the system to repeatedly want to reclaim.