Still on PTO for one day, but I'll try to answer. Sorry for the delay, was
out traveling.

2017-08-21 23:09 GMT+02:00 Matt Riedemann <>:

> I don't dabble in the RequestSpec code much, but in trying to fix bug
> 1712008 [1] I'm venturing in there and have some questions. This is mostly
> an email to Sylvain for when he gets back from vacation but I wanted to
> dump it before moving forward.
Heh :-)

> Mainly, what is the difference between RequestSpec.force_hosts/force_nodes
> and RequestSpec.requested_destination?
force_[hosts/nodes] was a former key in the legacy filter_properties
dictionary that was passed to the scheduler. When I designed the
RequestSpec object, I just provided the same names for the fields that were
related to those keys.
requested_destination is a new field that wasn't previously in the legacy
dictionaries but was just added when I wrote in Newton a new feature about
verifying a target when moving.

There is a main behavioural difference between those two RequestSpec
fields, that is because the behaviours are different between a boot request
and a move operation.
When you boot an instance, you can pass (weirdly and terribly from an API
point) a target for that instance by using the availability_zone flag. Some
request  like nova boot inst1 --availability_zone nova:foo:bar will
actually call the scheduler to just verify if host "foo" (and node "bar")
is alive *without* verifying filters.
When you move that instance by passing a target (something like "nova
live-migrate inst1 foo2") will rather call the scheduler to verify only
that host (and not the others) *by running filters*.

TBC, if you 'force' a boot, it won't be a real forced boot because it will
still call the scheduler but it will dumbly accept the proposed destination.
If you want to *propose* a destination for a move, it will ask the
scheduler to *really* verify that destination (or end up the migration if
not possible). You can still want to *force* a move, but in that case, it
just bypasses the scheduler.

> When should one be used over the other? I take it that
> requested_destination is the newest and coolest thing and we should always
> use that first, and that's what the nova-api code is using, but I also see
> the scheduler code checking force_hosts/force_nodes.
MHO is that force_hosts is just a technical hack based on an unclear API
contract about guarantting a boot request to succeed against a specific
target. I'd rather propose a new API contract for providing a destination
exactly based on the same semantics that the ones we have for move
operations, ie. :
- if a destination is proposed, run scheduler filters only against that
destination and error out the boot request if the host can't sustain the
- if a destination is proposed and a 'force' flag is used, then just pass
the RPC request to the target compute service and magically expect
everything will work.

That said, there is a corner case now we have allocations made by the
scheduler. If we go straight to the compute service, we won't create the
allocation if a force operation is made. In that case, something (that has
to be discussed in the spec mentioning the API change) has to reconcile
allocations for the compute service. Maybe it's just a technical detail,
but that has to be written in the spec tho.

> Is that all legacy compatibility code? And if so, then why don't we handle
> requested_destination in RequestSpec routines like
> reset_forced_destinations() and to_legacy_filter_properties_dict(), i.e.
> for the latter, if it's a new style RequestSpec with requested_destination
> set, but we have to backport and call to_legacy_filter_properties_dict(),
> shouldn't requested_destination be used to set force_hosts/force_nodes on
> the old style filter properties?
> Since RequestSpec.requested_destination is the thing that restricts a
> move operation to a single cell, it seems pretty important to always be
> using that field when forcing where an instance is moving to. But I'm
> confused about whether or not both requested_destination *and*
> force_hosts/force_nodes should be set since the compat code doesn't seem to
> transform the former into the latter.
I agree, that sounds important to me to have consistent behaviours between
all nova operations that require a placement decision. I have that in my
pipe for a very long time (I mean, writing the spec for modifying the API
contract on instance creation) so I push it a bit more.

> If this is all transitional code, we should really document the hell out
> of this in the RequestSpec class itself for anyone trying to write new
> client side code with it, like me.
That sounds a very good short win to me.


> [1]
> --
> Thanks,
> Matt
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe:
OpenStack Development Mailing List (not for usage questions)

Reply via email to