On 25/01/2015 11:00 PM, Ian Wells wrote:
Lots of open questions in here, because I think we need a long conversation
on the subject.
On 23 January 2015 at 15:51, Kevin Benton <blak...@gmail.com> wrote:
It seems like a change to using internal RPC interfaces would be pretty
unstable at this point.
Can we start by identifying the shortcomings of the HTTP interface and see
if we can address them before making the jump to using an interface which
has been internal to Neutron so far?
I think the protocol being used is a distraction from the actual
Firstly, you'd have to explain to me why HTTP is so much slower than RPC.
If HTTP is incredibly slow, can be be sped up? If RPC is moving the data
around using the same calls, what changes? Secondly, the problem seems
more that we make too many roundtrips - which would be the same over RPC -
and if that's true, perhaps we should be doing bulk operations - which is
I agree. If the problem is too many round trips our the interaction
being too chatty, I would expect moving towards more service oriented
APIs - where HTTP tends to be appropriate. I think we should focus on
better separation of concerns, and approaches such as bulk operations
using notifications where cross process synchronization for a task is
required. Exploring transport alternatives seems premature until after
we are satisfied that our house is in order architecture-wise.
Furthermore, I have some "off-the-cuff" concerns over claims that HTTP
is slower than RPC in our case. I'm actually used to arguing that RPC is
faster than HTTP but based on how our RPCs work, I find such an argument
counter-intuitive. Our REST API calls are direct client->server requests
with GET's returning results immediately. Our RPC calls involve AMQP and
a messaging queue server, with requests and replies encapsulated in
separate messages. If no reply is required, then the RPC *might* be
dispatched more quickly from the client side as it is simply a message
being queued. The actual servicing of the request (server side dispatch
or "upcall" in broker-parlance) happens "some time later", meaning
possibly never. If the RPC has a return value, then the client must wait
for the return reply message, which again involves an AMQP message being
constructed, published and queued, then finally consumed. At the very
least, this implies latency for dependent on the relative location and
availability of the queue server.
As an aside (meaning you might want to skip this part), one way our RPC
mechanism might be "better" than REST over HTTP calls is in the cost of
constructing and encoding of requests and replies. However, this is more
of a function of how requests are encoded and less how the are sent.
Changing how request payloads are constructed would close that gap.
Again reducing the number of requests required to "do something" would
reduce the significance of any differences here. Unless the difference
between the two methods were enormous (like double or an order of
magnitude) then reducing the number of calls to perform a task still has
more gain than switching methods. Another difference might be in how
well the "transport" implementation scales. I would consider disastrous
scaling characteristics a pretty compelling argument.
I absolutely do agree that Neutron should be doing more of the work, and
Nova less, when it comes to port binding. (And, in fact, I'd like that we
stopped considering it 'Nova-Neutron' port binding, since in theory another
service attaching stuff to the network could request a port be bound; it
just happens at the moment that it's always Nova.)
One other problem, not yet raised, is that Nova doesn't express its needs
when it asks for a port to be bound, and this is actually becoming a
problem for me right now. At the moment, Neutron knows, almost
psychically, what binding type Nova will accept, and hands it over; Nova
then deals with whatever binding type it receives (optimisitically
expecting it's one it will support, and getting shirty if it isn't). The
problem I'm seeing at the moment, and other people have mentioned, is that
certain forwarders can only bind a vhostuser port to a VM if the VM itself
has hugepages enabled. They could fall back to another binding type but at
the moment that isn't an option: Nova doesn't tell Neutron anything about
what it supports, so there's no data on which to choose. It should be
saying 'I will take these binding types in this preference order'. I
think, in fact, that asking Neutron for bindings of a certain preference
type order, would give us much more flexibility - like, for instance, not
having to know exactly which binding type to deliver to which compute node
in multi-hypervisor environments, where at the moment the choice is made in
I scanned through the etherpad and I really like Salvatore's idea of adding
a service plugin to Neutron that is designed specifically for interacting
with Nova. All of the Nova notification interactions can be handled there
and we can add new API components designed for Nova's use (e.g. syncing
data, etc). Does anyone have any objections to that approach?
I think we should be leaning the other way, actually - working out what a
generic service - think a container management service, or an edge network
service - would want to ask when it wanted to connect to a virtual network,
and making an Neutron interface that supports that properly *without* being
tailored to Nova. The requirements are similar in all cases, so it's not
clear that a generic interface would be any more complex.
Notifications on data changes in Neutron to prevent orphaning is another
example of a repeating pattern. It's probably the same for any service
that binds to Neutron, but right now Neutron has Nova-specific code in it.
Broadening the scope, it's also likely the same in Cinder, and in fact it's
also pretty similar to the problem you get when you delete a project in
Keystone and all your resources get orphaned. Is a Nova-Neutron specific
solution the right thing to do?
I have reservations. It all depends on what it is going to do. Referring
to it as nova-centric might also be a distraction. As part of scoping
out the work for refactoring the nova.network.neutronv2.API code (you
all know about that, right?), I discussed refactoring the neutronclient
to make it more usable as client library with several people. This
refactoring would prioritize nova's requirements when making changes,
but the changes would still be generally usable. Instead of approaching
this directly, I decided to start by wrapping the client in nova -
working around shortcomings and rationalizing the API somewhat. I always
saw this eventually being pushed out of nova into an independent client
library and in cases where the neutron API itself is being "worked
around", relevant changes made in the neutron API itself. In that sense
it is not unlike what Salvatore proposes but the approach is different
and ultimately not nova-specific at all.
OpenStack Development Mailing List (not for usage questions)