So I think it is a fair point that if you did: juju deploy application --constraits mem=4GB and then did something like: juju add-unit application --to lxd:XXX
That those constraints would end up interpreted differently. And also that: juju add-unit application --to YYY would similarly just ignore the constraints. That said, Juju does *not* do scheduling today, aside from explicit placement, and put-on-another-machine. And providing people with the functionality so that they have a way to express *something* about constraining their containers instead of doing a "juju deploy" and then having to do a "juju run" and know all the details of how to apply that. It may be perfectly fine for someone to do "juju add-unit -n 8 --to 5 --constraints 2GB" even though they only have 8GB on that machine. As their expectation is that any one container might try to peak at 2GB, but the expected steady state is <1GB and they just don't want runaway allocation in one container to starve all the others. I know we're also mulling over how we might propose syntax that would let you declare "this must have direct underlay access" vs "it's ok if this is behind NAT", if we can find something nice there it might also work for memory, etc. John =:-> On Fri, Jan 13, 2017 at 10:14 PM, Nate Finch <nate.fi...@canonical.com> wrote: > I just feel like we're entering a minefield that our application and CLI > aren't really built to handle. I think we *should* handle it, but it needs > to be well planned out, instead of just doing a tiny piece at a time and > only figuring out later if we did the right thing. > > There's a few problems I can see: > > 1.) you can have 10 lxd containers with memory limit of 2GB on a machine > with 4GB of RAM. Deploying 10 applications to those containers that each > have a constraint of mem=2GB will not work as you expect. We could add > extra bookkeeping for this, and warn you that you appear to be > oversubscribing the host, but that's more work. > > 2.) What happens if you try to deploy a container without a memory limit > on a host that already has a container on it? > > For example: > 4GB host > 2GB lxd container > try to deploy a new service in a container on this machine. > Do we warn? We have no clue how much RAM the service will use. Maybe > it'll be fine, maybe it won't. > > 3.) Our CLI doesn't really work well with constraints on containers: > > juju deploy mysql --constraints mem=2G --to lxd > > Does this deploy a machine with 2GB of ram and a container with a 2GB ram > limit on it? Or does it deploy a machine with 2GB of ram and a container > with no limit on it? It has to be one or the other, and currently we have > no way of indicating which we want to do, and no way to do the other one > without using multiple commands. > > This is a more likely use case, creating a bigger machine that can hold > multiple containers: > juju add-machine --constraints mem=4GB > // adds machine, let's say 5 > // create a container on machine 5 with 2GB memory limit > juju deploy mysql --constraints mem=2GB --to lxd:5 > > At least in this case, the deploy command is clear, there's only one thing > they can possibly mean. Usually, the placement directive would override > the constraint, but in this case, it does what you would want... but it is > a littler weird that --to lxd:5 uses the constraint, but --to 5 ignores it. > > Note that you can't just write a simple script to do the above, because > the machine number is variable, so you have to parse our output and then > use that for the next command. It's still scriptable, obviously, but it's > more complicated script than just two lines of bash. > > Also note that using this second method, you can't deploy more than one > unit at a time, unless you want multiple units on containers on the same > machine (which I think would be pretty odd). > > > > On Fri, Jan 13, 2017 at 3:48 AM Rick Harding <rick.hard...@canonical.com> > wrote: > > In the end, you say you want an instance with 2gb of ram and if the cloud > has an instance with that exact limit it is in fact an exact limit. The key > things here is the clouds don't have infinite malleable instance types like > containers (this works for kvm and for lxd). So I'm not sure the mis-match > is as far apart as it seems. root disk means give me a disk this big, if > you ask for 2 core as long as you can match an instance type with 2 cores > it's exactly the max you get. > > It seems part of this can be more adjusting the language from "minimum" to > something closer to "requested X" where request gives it more of a "I want > X" without the min/max boundaries. > > > > On Fri, Jan 13, 2017 at 3:14 AM John Meinel <j...@arbash-meinel.com> > wrote: > > So we could make it so that constraints are actually 'exactly' for LXD, > which would then conform to both minimum and maximum, and would still be > actually useful for people deploying to containers. We could certainly > probe the host machine and say "you asked for 48 cores, and the host > machine doesn't have it". > > However, note that explicit placement also takes precedence over > constraints anyway. If you do: > juju deploy mysql --constraints mem=4G > today, and then do: > juju add-unit --to 2 > We don't apply the constraint limitations to that specific unit. Arguably > we should at *least* be warning that the constraints for the overall > application don't appear to be valid for this instance. > > I guess I'd rather see constraints still set limits for containers, > because people really want that functionality, and that we warn any time > you do a direct placement and the constraints aren't satisfied. (but warn > isn't failing the attempt) > > John > =:-> > > On Fri, Jan 13, 2017 at 10:09 AM, Stuart Bishop < > stuart.bis...@canonical.com> wrote: > > On 13 January 2017 at 02:20, Nate Finch <nate.fi...@canonical.com> wrote: > > I'm implementing constraints for lxd containers and provider... and > stumbled on an impedance mismatch that I don't know how to handle. > > > > I'm not really sure how to resolve this problem. Maybe it's not a > problem. Maybe constraints just have a different meaning for containers? > You have to specify the machine number you're deploying to for any > deployment past the first anyway, so you're already manually choosing the > machine, at which point, constraints don't really make sense anyway. > > > I don't think Juju can handle this. Either constraints have different > meanings with different cloud providers, or lxd needs to accept minimum > constraints (along with any other cloud providers with this behavior). > > If you decide constraints need to consistently mean minimum, then I'd > argue it is best to not pass them to current-gen lxd at all. Enforcing that > containers are restricted to the minimum viable resources declared in a > bundle does not seem helpful, and Juju does not have enough information to > choose suitable maximums (and if it did, would not know if they would > remain suitable tomorrow). > > -- > Stuart Bishop <stuart.bis...@canonical.com> > > -- > Juju-dev mailing list > Juju-dev@lists.ubuntu.com > Modify settings or unsubscribe at: https://lists.ubuntu.com/mailm > an/listinfo/juju-dev > > > -- > Juju-dev mailing list > Juju-dev@lists.ubuntu.com > Modify settings or unsubscribe at: https://lists.ubuntu.com/mailm > an/listinfo/juju-dev > > -- > Juju-dev mailing list > Juju-dev@lists.ubuntu.com > Modify settings or unsubscribe at: https://lists.ubuntu.com/mailm > an/listinfo/juju-dev > >
-- Juju-dev mailing list Juju-dev@lists.ubuntu.com Modify settings or unsubscribe at: https://lists.ubuntu.com/mailman/listinfo/juju-dev
