Setup jemalloc in default Dockerfile to optimize memory allocation

[northeastprince]

Ruby's use of malloc can create memory fragmentation problems, especially when using multiple threads like Puma does. Switching to an allocator that uses different patterns to avoid fragmentation can decrease memory usage by a substantial margin.

[byroot]

I agree we should do this, as it's really a no brainer in term of performance.

However as is this PR break a few stuff and need some more work. IIRC the jemalloc package comes with a command that does the proper LD_PRELOAD, we should just use that.

[gjtorikian]

Apologies for the drive-by comment, but I think the "better way" would be to patch Ruby directly to use libjemalloc, as mastodon shows: https://github.com/mastodon/mastodon/blob/1666b1955992e16f4605b414c6563ca25b3a3f18/Dockerfile#L108-L113

LD_PRELOAD affects every binary, not just Ruby, which may not be desirable. More details here.

Happy to make a PR for this change if it is desired!

[byroot]

I'm aware of this technique, but I don't want to patch a binary that comes from a previous layer, and the LD_PRELOAD leaking isn't a big concern except for Google Chrome, which is an extremely niche use in production, and is easily worked around by add ENV.delete("LD_PRELOAD") in config/boot.rb or something.

[gjtorikian]

Gotchu!

[joshuay03]

Minor typo:

Suggested change

	# Enable jemalloc for reduced memory usage and reduce latency.
	# Enable jemalloc for reduced memory usage and reduced latency.

Or just removing is more succint imo:

Suggested change

	# Enable jemalloc for reduced memory usage and reduce latency.
	# Enable jemalloc for reduced memory usage and latency.

[byroot]

1ca8c8e

[suwyn]

@byroot @northeastprince I'm not seeing jemalloc being used in the process map. Can you double check?

Compare the output of

pmap -X 1 | grep jemalloc

on the rails server process vs the output using

exec env LD_PRELOAD="$LD_PRELOAD" "${@}"

in the entrypoint script

[byroot]

Yeah, I forgot to add export. Fixed in 9940dc8. Thanks for noticing.