Shouldn't we set overcommit_memory=1?

[gastonmorixe]

Guys,

There are some official recommended settings here http://redis.io/topics/admin

One should be
RUN sysctl vm.overcommit_memory=1

What do you think?

Thanks

[yosifkit]

This was discussed in #14. But I am open to good reasons.

[tianon]

It wouldn't be possible to set this via a RUN line in the Dockerfile -- that won't carry into the image.

Also, I think this sets it system-wide, not just for the container, so I'd be wary of setting this in the entrypoint too (which would be the only place it could possibly even work).

[gastonmorixe]

so, is this a limitation of linux containers?

[Starefossen]

+1 for RUN sysctl vm.overcommit_memory=1

[archr]

+1 RUN sysctl vm.overcommit_memory=1

[tianon]

so, is this a limitation of linux containers?

Yes. Some sysctl options are scoped to the container, but a large number are read from the host. On top of that, actually changing their values requires write access to /proc/sys, which is impossible by default since /proc is read-only inside Docker. I'm +1 for documenting that folks using this image ought to set this value on their host, but we can't do anything about it inside the container.

[sebicas]

Maybe related: boot2docker/boot2docker#1083

[tianon]

See also moby/moby#4717 -- there isn't anything we can do from the Docker perspective except document how to set this value on the host (which would go into https://github.com/docker-library/docs/blob/master/redis/content.md; a PR would be awesome).

[pataquets]

moby/moby#19265 just merged. Can this solve the issue?

[tianon]

Very cool -- that won't make it into 1.11 (being released today), but we could document for users of 1.12+ how they can use --sysctl on their docker run invocations in order to accomplish this. 👍

[justincormack]

I don't think vm.overcommit_memory is a namespaced sysctl, so that will not resolve this issue - it still has to be set globally on the host.

[tianon]

@justincormack aww, fair enough -- I suppose that means there's not really much we can do here besides documenting that it's recommended and that it'd have to happen on the Docker host, not in the container 😞

[bkuhl]

This isn't a "run a container to solve a problem" kind of thing, but I appreciate there being containers running as part of a stack to help serve as documentation for what's been done to the host. That said, I believe some folks may find https://github.com/bkuhl/redis-overcommit-on-host useful.

[yosifkit]

Thanks for the note @bkuhl.

Closing since this not something that can be changed by the image, especially since we don't want to have users running the redis container with --priviledged.