While debugging mallocs are very useful they tend not to be as efficient as system mallocs. I don't know anything about jemalloc, is it only used via dynamic loading patching if enabled? And how will this work on OS/X platforms (where I think that the system malloc has much of this built into – see mallocinfo)?

jemalloc is actually the main malloc of FreeBSD, so it is a system malloc.

So on FreeBSD we don't need it. I'm worried that on some hardware the vendor may have "value added" a custom malloc that knows about e.g. their memory backplane.

This is a good explanation for why jemalloc is useful for memory profiling Debugging Java Native Memory Leaks

The library is only used if you set the LD_PRELOAD environment variable and the profiling only occurs if you set the MALLOC_CONF environment variable appropriately. Otherwise, it uses the system malloc.

So, if you want to use the memory profiling, do this before starting the program. It should generate a report every time a GB of memory is allocated.

export LD_PRELOAD=/usr/local/lib/libjemalloc.so

export MALLOC_CONF='prof:true,prof_prefix:/qserv/run/var/log/jeprof.out,lg_prof_interval:31,lg_prof_sample:22'

Please note in title above, the proposed dependency add is only for Qserv

I see no reason why this RFC can't be adopted. Is the implementation ticket DM-10971? (for switching qserv).

jemalloc was packaged and published in ticket DM-11214. This has allowed us to use jemalloc for the intended debug purposes in branch builds. It remains to be seen whether we'll cut Qserv over to jemalloc in general, but that decision and work is not under the umbrella of this RFC.