Minutes IETF103: netconf
||Minutes IETF103: netconf
DIRECT LINK TO YOUTUBE VIDEO:
Session Intro & WG Status (15 minutes)
Ignas - Status update, zerotouch should be on Telechat in December.
Kent/Ignas - Request for volunteers for shepherd writeup.
Kent Watsen (15 min)
Status and Issues on Client-Server Drafts
Mikael A: I just want to say I like that we're not discussing if we should be
able to set keepalives because I want to be able to configure it at all layer
so fully support the effort and then exactly how to do it I don't have a strong
opinion but this seems good to me so far.
Jason S: I am just trying to understand the concept of having it like turn on
keepalive at TCP layer on a server you're gonna turn on globally and then it's
gonna have keepalives for every TCP session, or are you saying that every TCP
session that's using NETCONF. What is the granularity of the control you're
Kent W: Well it'd be in a configuration data model and these are really
groupings so whatever grouping stack is, in this case NETCONF, for instance,
it's using the ssh client server grouping which itself would use TCP
client-server grouping so you would inherit some keepalive configuration for
the TCP keepalive from that grouping and then likewise some ssh keepalive
configuration from that grouping and if ever we get around to it some NETCONF
level keepalive from that grouping and it would only apply to that particular
Jason S: Okay so that's not a combination and so you'd be turning on TCP keep
lives for this NETCONF session correct?
Kent W: Not for the entire operating system
Jason S: Yeah okay thanks.
Tim C: I will say, I think the approach to having keepalives at every layer
makes sense because again last time we talked about it it's needed for that
piece of it. To answer Jason's question is that, indeed, this is within the
context of the NETCONF session. I will say that as we were doing this for some
of the stuff within the BroadbandForum work that some of the implementations
are that when you turn this on you touch those TCP keeplives it's actually for
all the sessions. That there's some limitations on some of the implementations.
Not that the approach still should be the context of the NETCONF session.
Sometimes the implementation says that the only thing that you can twiddle is
on all the sessions on some of the implementations. My concern with the
refactoring is that we've got modules coming out of the wazoo. I mean there's
just a ton of modules. I get what you're doing but there's just a lot of
modules now. If you're gonna do if every layer ... So I'm just wondering if the
organization and housekeeping gets beyond the benefit.
Kent W: And that's what I meant by anticipating exasperation. I have somewhat
worn out the welcome here in terms of the refactorings we've done and I get it.
But i also see this as being the best way to solve this problem.
Mikael A: In Linux there is a knob to turn on this system wide so you could
test that and see whatever you come up with for TCP, would fit into IETF system
for instance and know if it would make sense. A test of this module to see if
generic enough. If it is, I would say publish it, because I might want to
support this. At the system level as default settings and also at the NETCONF
server session level, so that the NETCONF server would use the socket option to
turn this on for just its sessions and I would also like in IETF systems model
for the entire system.
Kent W: Both. You are in support of this and also modifying IETF system
Mikael A: I'm saying at least spend 10 minutes on seeing if the the way you do
the TCP part is generic enough that it would be able to fit into a IETF system.
I would support an effort to do that for IETF system as well, because as far as
I know there is no system-wide settings and at least some operating systems do
turn TCP keepalives on by default.
Kent W: Okay well look into it.
Rob W: I think separating modules probably makes sense, but doesn't mean they
have to separate drafts. All of these could be bundled. Some of these modules
together are relatively small, and that may reduce some of the overhead or
process overhead here.
Mikael A: I wouldn't even need it per connection, or IP or something. I just
need a knob default on, or default off for all sessions.
Kent W: At system?
Mikael A: No, for NETCONF. TCP keepalive for NETCONF and SSH. I don't even need
that. I don't know if someone else expressed the need to do this on a per
connection level. I just wouldn't. I want like a default. Set it on and off
basically (at the system level?) that's all is what I need. If someone else
needs more then okay. Then we need to do more. I'm just saying there are a lot
of scenarios in which I think you just want to be able to turn it on and off
and you don't need to set it per destination or anything.
Kent W: Understood. Where would we put a global setting like this, is the
question. Maybe IETF system?
Mikael A: Default on and off for the NETCONF server for all sessions that come
in. I don't need it to do it per IP. I just need to say the NETCONF server
should by default keep tcp keepalive on. I don't need to say for this IP range
or something like that.
Rob W: So definitely support your suggestion saying that if you do do this,
split out to keep the minimal and in the future it needs to be expanded that
could be done in the future revision. Just trying to avoid feature creep.
Kent W: Right and actually already that's the strategy we've taken like for
instance with the SSH and TLS. We're really just focusing on the minimal
necessary to configure the crypto stack so that we can do it. But if you look
at various SSH and TLS implementations there's many more configurable options
we're not touching any of them. So more the same in that regard.
Tim C: Just to make the comment from Michael, is that we're using it for
various applications. So there's a TR 301within the Broadband Forum that does
call home that uses TCP keepalives. And so there's a the generic setting, which
may or may not work. I don't know. Not sure it works because it's on a stream
basis and we've got multiple endpoints that we have to talk to. But we've also
started looking at the augmentation because we weren't waiting when you're
gonna do right. I think we can work with you on trying to figure out, how to
get the best adaptation possible to work in different scenarios.
Kent W: If you know people could collaborate with me I'd truly believe this
could be done by 104. We would still have to figure out how to do keep lives
and I don't have an alternative proposal. This is the only proposal I have at
the moment. I don't think the effort is huge if the co-author could help me. I
do believe we could get it done really with probably just one more month worth
Alex Clemm and Reshad Rahman (10 min)
Update on YANG Push and Related Drafts
Kent W: Okay. I'll just make some comments. Thank you Reshad for helping with
the RESTCONF notif draft. It wasn't one of the original three that we had
discussed, but adding it I think was good, mostly because it actually helped us
find some other issues right in the SUBSCRIBE notifications draft. It had a
sort of a trickle cascading benefit, I think, beyond just sort of enabling us
to allow NETCONF and RESTCONF to move forward together, which in general is
good for the working group. To the working group, as shepherd, as soon as I
received these drafts, I'll also send out an email to the work group asking
those who had posted comments to just review that were hoping to see was is
there and let the shepherd know if there's anything amiss.
Alex C: Just one clarification. When is send to IESG? Once the Shepherd review
is complete? Is it serialized like that?
Kent W: Yes, for the most part. The shepherd does the writeup and then and
shepherd isn't necessarily always the chairs but the write-up occurs and then
the chairs discuss whether or not it's appropriate to submit for publication
which is equivalent or synonymous to going to the IESG. It almost happens at
the same time. But what that really means it goes to the AD, Ignas, in this
case, where he would do a AD writeup and he needs to schedule it for telechat
which he just mentioned a moment back. It could be a month out but you have to
get on the calendar. Once that occurs then there will be a number of discuss
items and the various IESG members will have comments, all of them will ballot
on your drafts. They will be discuss items and you will need to resolve all
those discuss items. This isn't really normally in the view of the working
group. It happens off the working group list. The chairs and the shepherds who
are involved in that process. That can take as long as it takes it and I've
seen it sometimes go quickly a couple weeks and other times months. Just
depending on how it goes. Then when that concludes, it goes to RFC editor and
then there are RFC editor will look into other issues and they'll be more back
Alex C: I was mostly concerned about getting it in front of the IESG. Ok. Thank
Tianran Zhou (10 min)
UDP based Publication Channel for Streaming Telemetry &
Kent W: As a contributor First just a general clarification statement what this
draft is presenting is a notif model. We have like NETCONF and RESTCONF notif
model. This would be a UDP-based transport notif.
Tianran Z: Yes
Kent W: How we might characterize this draft with all the other notif drafts.
The current notif drafts, they're really just providing dynamic
subscription-only support because we never really got around to thinking that
we would want to have configured NETCONF and RESTCONF subscription. But we do
want to have configured udp-based subscriptions and that's how this draft came
to be adopted working group supported item, and also I think having dynamics
descriptions make sense as well, so that's the clarification. Secondly, I think
we need to be clear that what we're describing here is a new protocol. This
would be a new binary protocol. We're calling it UDP.
Tianran Z: The name is UDP based publication channel
Kent W: We do have a name for but it is a UDP based protocol. It is defining
its own message header and it can contain different encodings. Is defining a
new UDP based protocol making sense relative to making use of an alternative
existing UDP based protocol and I know you discussed IPFIX and COAP. I guess
with IPFIX one of the things you mentioned was that it doesn't support
different encodings but maybe that's okay. Maybe just a single encoding would
be okay. For COAP, you mentioned that the message ID was only 65 thousand. I
think the concern there is that IoT devices and their rate of transmissions
would be very slow but a high-end router could send 65,000 messages within a
single second. It would lead to many issues. But if that's the only concern
there is an opportunity for this working group to approach the COAP working
group to ask them if there might be a possibility to extend that message
header.I'm just exploring ideas other ways that we might be able to solve the
general problem which the working group wants to solve which is a UDP-based
notification message for subscribe notifications. I think we should still
consider the solution space some more.
Henk B: We are working on the concise yang telemetry draft and we are using
COAP and the message ideas for detecting duplicates. If you expect to have a
duplicate in a 16-bit space then you need a bigger message ID. But I don't
think that is a concern. It can rewind in that scope, if you don't expect
duplicates in that dimension. The association between the requests also. The
subscription to the stream is the COAP token and that's eight bytes I don't
think that's the problem. Maybe message ID is sort of misinterpret idea I
think. I don't see that it is a problem but on the other hand you want to have
this inside system like I heard like between line cards or in this is not
leaving the the data store system component, I have the feeling, so maybe then
having a listening server. This is best RESTCONF. For COAP server is a little
bit over too much then you can just establish a UDP stream. Tt depends on the
application. If it has to go through the internet probably COAP is a good idea,
but if it is just for high volume inside a system you can basically unpack all
of the overhead tuned for Internet Protocol.
???: I was just explaining that IPFIX does leave the box. it's generated on the
line card itself and it goes out the box to a collector.
Henk B: This is of course correct but I thought IPFIX was used in a different
place. For this purpose therefore has a different scope of application. I think
because of encoding.
Kent W: Just add that conversation I've also done UDP-based logging where the
log receiver was on the subnet to the line cards and it would receive all the
logs and then do aggregation and compression and deduplication, and then send
them over the LAN. I think that's your point.
Henk B: Yeah that's my point.
Kent W: It's over over the LAN you don't really have to tag it and besides if
you miss one what would you do about it anyway.
Henk B: My last comment is anything but binary representation doesn't make much
sense. Inside, if you're talking about burdened by TCP state I think being
burdened by something else as a binary is even worse. I think it's rather
obvious not to use human readable clear text formats like JSON or XML I think
it would defeat the initial purpose of the block.
Tianran Z: I should say those requirements are from our customers so we
designed for them.
Rob S: Google. I find that the whole section of this draft to do with any kind
of reliable delivery and discussion of how you should only really deploy this
over reliable networks is under specified. Our operational experience of having
tried to put something UDP-based streaming into production is there are no
reliable delivery channels. You have bits of your network where you can't
possibly assume that all packets are going to get through or there's no
congestion, because there isn't the amount of bandwidth you can buy there is
not sufficient. The cost of having to assume that the channel is unreliable is
the law of periodic replication of the data. So you can deal with
retransmission .I would go so far as to say as soon as you have to deal with
retransmission you might as well use TCP anyway. With TCP you also get the
advantages of knowing that reliably when you sent an event it got to the other
end so you can reduce the number of times you need to stream data. We don't
think that it's actually possible to do over an unreliable channel event based
updates, because any system then can't really rely on it and with any kind of
latency. I think you should probably add some discussion to your draft as to
what the cost of doing this over UDP is and really try and figure out how
retransmission works in this model. Especially if it actually works to a line
card which is kind of the motivation here. You're assuming that there's a cash
on the line card to have any packet within some known window to be requested.
My suspicion and operational experience of having a few thousand devices that
run telemetry at this point, across number of vendors, is that you will just go
to TCP again; as soon as you have to deal with these problems which are kind of
the operational realities. I don't really think we should be pushing the
industry in a way that doesn't really work.
Tianran Z: The reliability is the part that is not the real reliability as in
TCP. It's a kind of partial liability it's a trade-off between reliability and
Rob S: Right. But the the problem is how do I build any kind of system that
relies on the data being there, so I can if I'm trying to do anything with
interface statistics and I know that there might be fidelity loss because I've
got lost packets I can't rely on it. I can't do anything event based because an
interface goes down in my network and then you don't have any way to react to
it.You don't know that the state is there. The natural requirement then is that
you end up building a polling system to make sure that you have a current
enough view to reconcile and our scaling analysis kind of shows as soon as you
do that, you're going to end up with significantly more data than you would via
TCP. This scalability argument kind of falls down. We've been pushing this
Kent W: As a contributor. What is the motivation for your UDP-based draft? Is
it the reliability? I don't think that was it, so much as the desire to enable
the line cards to send the UDP packets having the same source IP. For the other
draft that you are about to present, the multiple stream originators, the
desire is to enable that distributed source.
Rob S: So we've looked at this. I think that there's a model whereby you have
a distributed system that has different components that can each have TCP. I
think you're going to end up going that way if you ever care about reliability.
If you say this is a hundred percent unreliable then I think you can kind of
talk yourself into this UDP model but if you say I want distribution because it
gives me more scalability (point to be proven as to whether that's really
required) and then you you can still do TCP or it is a lightweight TCP protocol
to the line cards and you're kind of inventing a new protocol here, as you
pointed out. I would suggest for debug ability it's kind of a challenge if you
have N producers that are all producing with the same source IP. We have
challenges around being able to know whether you're actually in synchronization
with that system if you've got N different producers and one line card stops
producing data. You don't really know you've got no metric to be able to alert
on say if this source isn't sending data anymore.
Kent W: I think we're jumping into the next draft but I think the idea with
that draft is the is that the configuration model would allow you to configure
the UDP to the system and then the system implicitly distributes to line cards
and tells each of them. But if you do were to do TCP you'd have to be explicit
the configuration model would actually have to configure the IP address for
that line card.
Rob S: Yeah I'm suggesting a bit of configuration pain is better for the
Henk B: Again if you are expecting to have congestions, you will have UDP
datagram loss. That is a fundamental decision you have to make. Do you expect
congestion with your streams or not. If you have that expectation, which I
think is likely then you have to deal with retransmits and for that you should
not reinvent yet another (reliable)transmission mechanism for UDP for every
draft in the IETF. There is a good template for that in COAP where a reliable
message every thousands message is sent, and you can then see how many messages
you lost and that window can be retransmitted. It's a little bit like TCP but
light weight. I call it a reliable COAP. That is an alternative. My suggestion
is to approach the research group that is meeting here. There are two drafts in
development which talks about how to associate data items that are in series
and the problem of retransmits is discussed. If you have a problem that is not
solved in general and you want to solve it with your draft.
Tianran Z: I know of COAP used for IoT. Do you have an example for COAP
application that is used for routers?
???: In the DDoS protection working group we use a kind of our basic transport
protocol for the for the scenario, if you want to look.
Tianran Z: I would like to see.
Henk B: Just because something that was initially intended to be used in the
constrained environments doesn't make it unfeasible for the rest of the Internet
Kent W: As a contributor. Just a quick follow up on the discussion about
retransmissions. When I first saw the message ID with the UDP I never thought
that it would be for the purpose of knowing when to request for a
retransmission. I only thought it would be used for ordering of the packets
received by the receiver because UDP doesn't guarantee order delivery. And for
detecting gaps now when a message was dropped. Not to request but to know that
you lost a message. I never thought that there would be a desire to try to
build reliability on top of a UDP based protocol.
Rob S: I think that that's a interesting operational like mode of operation.
Like I said before with SNMP.I can poll the device and know I get some stats
back. Maybe there's some loss in them but I know at what interval I'm polling
in. In this (UDP) mode where there's no reliability, if the device just shuts
up you can't tell because you didn't get a sequence number to tell. It becomes
quite operationally difficult to not assume reliability when there isn't a
guarantee that the thing at the other end is sending data. I mean we tried
this. We looked at it as the preferred way to start with, and this a long thing
with internal collector deployment things about how you know when it can
reconnect, about how you deal with redundancy between collectors and those kind
of things. I think it just makes for more and more challenges. That is kind of
why I think the draft could do with some discussion of like how you actually
operate the system like this.
Jeff T: Just to interrupt comments. I work for a company, where we use
streaming extensively from thousands of devices. Spend quite some time looking
at UDP and TCP and I also discussed with potential customers. UDP was a no no.
It has to be reliable otherwise you need to build additional layer to ensure it
Benoît C: It interesting because we were discussing the same thing that we have
been discussing for IPFIX for 10 years. The message ID in IPFIX was just to
know about the order and just to know that you've been losing flow records. The
point is that for IPFIX it works fine because of accounting. If you lose one
packet, big deal. BTW, you expect a router to keep information records. I think
the key point is that if you rely on this mechanism for an event like Rob was
mentioning, it must be reliable. If you just going to sending monitoring
information, you can use UDP.
Rob S: Just to add to Benoit's point, It's fine I think, fundamentally for
SFlow or IPFIX to have a different nature, because we know that there's N flows
on the device where know that for n packets going through the device we know
that we only expect a sample of them therefore losing one... I don't know of
any system that's built to say with SFlow or IPFIX where with one flow sampling
I'm going to guarantee that I'll get every one of them. Whereas with telemetry
data if we're building systems that now split the control plane across the
device and off the device then we need it to be reliable just like you would
need some of this data internally to the system like links going up and down to
be reliable for routing protocols.
Benoit C: Again, it depends what you call telemetry. If it's telemetry I mean
to push high frequency all information from it's like IPFIX, your sending flow
records; even if they're not flow whatever, if you condense everything in your
telemetry so it becomes an event you can't miss it.
Alex C: Some application may require reliability while others where you're
saying you lose one record it's not a big deal. Another question is how it is
being used if you use this for periodic updates you know basically that you are
expecting updates for every period already. If you have a period missing you
would basically infer some of those things. I do agree actually that we need to
have the discussion of these operational things and the trade-offs. At the same
time I think nobody is saying that this is the be-all end-all transport for all
particular use cases. This is one use case for certain scenarios where
basically those operational scenarios that you described would be applicable.
Rob S: Again just a response that. I think there's a few challenges with those
assumptions. As soon as you say oh I'll stream everything periodically you're
going to significantly increase the data that comes from the device and by
hundreds and hundreds of times. Actually it makes this system scaleability
problem worse. You want to only send things when they change. It gives you a
significant advantage for large data sets. It also gives you a significant
advantage for interfaces that are down, on a systems with radix of a thousand
or so. Which is kind of common in today's networks. As soon as you say I'll
send things periodically you're going to end up with these scalability concerns
and you probably are now having to deal with worse scale on the device of your
UDP periodic than you would be the cost of doing TCP for reliable. Other
problem about periodic is that you don't actually know what theycollect or what
you meant to receive. If a whole line card stops sending did it get removed
from the system. It'sactually hugely difficult without lots and lots of other
accounting to know what you should have been sent during that period. The third
thing I shouldn't say is we're inventing a new protocol here to send telemetry
data let's notinvent one that we know is flawed and only works in like a small
number of cases because that is just going to complicate things.
Alex C: When you subscribe the subscriptions both can support either use case.
A user will decide whether they happy with periodic or whether on change is
actually more applicable for their particular application. If you want to have
a continuous and telemetry to do some kind of whatever whereas statistics trend
line analysis for your application. Not every use case requires on-change.
Rob S: That's true. But I guess the point is that there's some data and an
underlying you do want to sample. This doesn't mean that the system can't
support sending data periodically. There is a significant amount of data that
won't need to be sent. I would encourage people to go and look at is to look at
the data that is being pulled from devices. This is what we've done and then
look at what the proportion of it that is event based versus periodic and do
some calculation, as to what the data volume is. The scaling analysis in the
gRPC based telemetry where we can show you know significant reductions based on
this. Even though that some data is being sampled and sent periodically because
it needs to be sampled like that from underlying hardware sources.
Reshad R: Related questions. I remember there's previous message ID and that's
how the receiver knows that so many messages have been lost. But if you're not
receiving any messages how do you know that there's messages loss. re we going
to do a UDP keepalive draft?
Tianran Z: Weed this keepalive information.I don't know about this this is
something we need to consider. We also have in the other draft some mechanism
to solve this problem.
Benoit C: In IPFIX we solve that with SCTP and get a perfect solution where
actually you would have a stream which is reliable, unreliable or partially
reliable. Depending what you're sending it is monitoring it would be unreliable
you miss a couple of information, fine no big deal, you might be partially
reliable you do your best or reliable if it's event based. That's how we solve
in IPFIX. However with SCTP didn't pick up and it's an issue with line cards.
It is an operational issue and I think Rob mentioned that how do you identify
your device a router is like one IP address or it's a sum of IP addresses to
one per line cards.
Tianran Z: Sorry I do not understand your question.
Benoit C: It's not a question it's a duration that we've been looking at these
issues. Ten years ago it becomes like more an operational issue. What do you
want to solve and then you will have the solution for your protocol.
Mahesh M: Speaking as a contributor.I think the message that I'm getting from
the working group is you probably need to look at the data set to decide if you
need a UDP-based channel or if you if you need reliability. If you are gonna
build reliability into this with the sequence number why not just use TCP. As
Benoit mentioned if it's monitoring data that you're looking at losing a few
packets is not a big deal. But if you're looking at event you can't afford to
lose it. What is the cost of doing that?
Kent W: One other comments as a contributor. Just one more comment.I heard or
learned last night that the I guess it's in the COAP working group. There's a
an effort that's been going on for a couple years now to do a COAP based broker
pub/sub mechanism, but we should learn more about it and and see how it might
be usable in this space as well.
Tianran Zhou (10 min)
Subscription to Multiple Stream Originators
Kent W: Kent, as a contributor. can you go back to your previous slide, the one
that had the diagram and had the red box that said "out-of-scope". Yes. Why is
it out of scope?
Tianran Z: as I mentioned in some instance, like the carrier routers, this is
kind of the internal implementation, so I think must between the mainboard and
the line cards.
Kent W: right, okay, I think that this being out-of-scope is dependent on the
conclusion of some of the operational requirements that we were discussing a
moment ago. Going back to Rob's comment from before, a little more
configuration complexity may be warranted if we were, for instance, needing to
configure TCP instead of UDP, in which case it would have to be in scope,
because you'd have to be configuring what is the TCP interface, at least, for
each line card to use, so I think what you're saying is it's out-of-scope here
because the expectation is that, from using UDP, the routing engine can
internally communicate the line cards
Tianran Z: yes, okay
Kent W: so there's that assumption, which i think is not is still TBD, is what
Tianran Z: okay, but my concern is this part is a little bit complex, and may
vary from implementations, so I'm not sure if it can converge in this draft.
that's my concern.
Mahesh J: Mahesh, as a contributor. okay, if you could go down a couple of
slides, to where you talked about being able to reliably indicate a change in
the subscription [the Subscriptions State Change Notifications slide], it says
"all the subscription state change notification MUST be delivered. Now, when
you say "must", that means you're thinking about a reliable channel here for
delivering that change in notification?
Tianran Z: that's an interesting question. from the message layer, no, we do
not consider it must be a reliable channel so maybe in UPD case, maybe we need
Alex C: this is Alex. Just to add on or respond to that, I would not mix this
with the earlier transport discussion. The goal certainly is for the
subscriptions, per se, has always been to basically make this, well, make the
fundamental mechanism reliable, so that you can avoid having to poll things.
Now obviously, with this case, if you have a new component subscription that
was added, or something that was removed, that is an event that you would needs
to know or that you would want to know, certainly as a collector, therefore,
basically, there's something that needs to be notified. We've had some this
some internal discussion whether it should be subscription-modified or whether
there should be another type of notification, but either way, it is an event,
and suppose it should be foreseen as part of the control channel. Now, if you
want to have making this busy for the for the control part of this. Now, for
the actually telemetry stream, whether this is reliable or not, that's a
separate issue, I would separate those discussions, but this one would be
needed for reliable control channel, so to speak.
Kent W: Kent, as a contributor, maybe as a chair, I don't know. The motivation
for the working group, particularly when adopting the previous draft, as this
draft is not yet adopted, but the idea that this draft is discussing was one
that was presented at the time that we adopted the previous draft, which was
the goal to support line cards to be able to send messages themselves directly,
as opposed to trying to forward them to the routing engine, in order to for the
routing engine to send them because, from experience, we know that the internal
backplane switching fabric does not have enough bandwidth to transmit that much
data, it's just not possible, the line cards have to be able to send directly
themselves and, in fact, you know things like encryption, actually, it's
probably problematic, and this goes to the operational requirements, are we
actually thinking that for these very high logging scenarios, would the
destination be an internal receiver, something that is on the LAN, something
that itself would collect the logs in unencrypted form, do deduplication and
analysis and compression and, perhaps, even itself could convert it to binary.
we don't really need binary in the LAN, we need binary on the WAN. When we
talked to customers, their costs for bandwidth over WAN is expensive, that's
when they care, they don't care about the bandwidth on the LAN. I think we
need to understand what are the operational requirements, and what is the
problem we're trying to solve, maybe some of this would become more clear. I
still strongly support the ability for sending logs out the line cards
directly, as that's important problem solve, but the motivation for it being
binary, and the motivation for it being UDP even, I think we should go back to
asking if that's really important to solving the problem here.
Mikael A.: Mikael Abrahamsson. it struck me, the whole thing about line cards,
and it being on box, I have the use case where I might have a Wi-Fi access
point that basically doesn't have an IP address, so i'm speaking some kind of
protocol to it or I don't want it to send anything, it's going through me, but
it's a different device, and I want to like expose this in YANG so that the
Wi-Fi ,it's like I'm the NETCONF server, but I'm configuring the guy over
there, and I still want exposed to my NMS that these are two different devices.
Isn't that kind of the same problem? don't you want a more generic approach
in how to expose this in YANG and NETCONF? Because isn't this the same thing
and like if it's a line card that is like its own computer sitting in the
chassis, or if it's something else, like you're acting on behalf of that guy, I
mean, I've seen many different scenarios where you need the same concept, so
can we make it more general?
Tianran Z: yeah, I think you provide an interesting use case, maybe similar to
this IoT use case, and so we actually I think this framework is like a generic
Mikael A: my problem is that we're talking about what's in the UDP packets that
is streaming telemetry
Tianran Z: that's another draft
Mikael A: yes, I know, but it's like this here is talking about subscription,
but isn't this just configuration it's just it's not for me specifically, I'm
doing this for another guy, that he's like near, I'm controlling him, so it's
not me. Don't we need like a more generic approach? and don't we need to talk
about what the configuration is instead? isn't it like how do we do this
generically? isn't that what we should be discussing? we're talking about
subscriptions here, and how to talk to that guy, or what to actually configure,
but don't we actually need just a best common practice for exposing this entire
concept of different devices managed through one NETCONF server?
Kent W: I have a response for this, but I think Rob does as well. It is for
this question Rob? yes, please go ahead.
Rob S: Rob. I completely agree. so in the GRPC space, for both telemetry and
configuration, we added a generic way to be able to make a path be addressable
to a certain target that they entered the managing entity deals with it, and
it's used for both telemetry and for configuration. I think having case-by-case
solutions is not optimal, I think having a single way that you can say there is
this management agent that is responsible for this other domain is super useful
for many many cases
Kent W: okay, so up-leveling the problem space, great! My closing thought is,
not it's not really on this draft, but kind of to the other one as well.
Currently, the YANG Push and Friends drafts are almost out of Last Call and
into Shepherd write-up. As a working group, we've only defined support for
dynamic subscriptions, we do not yet have any support for configured
subscriptions. This draft is on the path towards enabling support for enabling
configured subscriptions, but I think that there may be other paths that we
could explore that would get us there faster, specifically HTTP-based push
mechanism is something along these lines. I don't know, maybe somebody would be
interested in putting together an ID to propose another notif draft. The nice
thing about the way that we've constructed, or deconstructed, the the YANG Push
notification drafts is that we do have all these notif mechanisms, so it's like
a Swiss Army knife. It's great that we have a UDP-based mechanism available,
certain deployments will use it for their use cases, others won't because it
doesn't match their use cases. So I think we should also consider other notif
mechanisms that would enable us to have configured subscriptions.
Tianran Z: I think the idea is become mature, and the the solution and the
scope is kind of clear, so I am I'm wondering if I can ask the working group to
adopt this document doc but
Mahesh J: I think before we get to the point of adoption I, we probably need to
address the question that Mikael raised, which is do we need to upscale this
problem definition before we get to the question of adoption. Maybe you should
consider all that first.
Tianran Z: okay, thanks.
Qin Wu (10 min)
Inline Action Capability for NETCONF
Mikael A: Mikael Abrahamsson. I don't know, I haven't read the draft, if I
understood correctly, is this only for when the will you change the
configuration but the end result operational state does not change? so you're
splitting the range into two, but the effective configuration on the line card,
or whatever, never changes? is it only for that type of configuration, or is it
also it like you split it and then you delete one and you do that in one
Qin W: uh
Mikael A: okay so you had the example one-to-five and six-to-ten, you can merge
you can merge those and into one record
Qin W: yeah
Mikael A: if this changes, nothing in in real life changes, I mean, the line
card's hardware doesn't get reprogrammed by that operation, you're changing the
configuration but the state of the device doesn't change, is it only for that
type of operation or is it also for deleting one of the VLANs in the middle
(i.e., it's for both)?
Qin W: yeah, I think we right now we really support both actually. the
motivation we can have the merge several tag into one allow you to do better
NETCONF query, you know, but actually we also support also, in some cases, you
may need to delete a some of the value from the VLAN tag ranges, so we provide
such capability in some cases, actually we need to support both and and then we
can actually, you know optimize the actual, by merge several range into the one
Mikael A: so yeah do you see this as an optimization in number of transactions,
or is it processing power on the server, or on the client?
Qin W: we don't want you add overhead to the client, actually maybe you just
need a one transaction, but this is actually transaction you know you send a
request to the server actually we are actually you know using some existing
config template to merge you the range into the one, actually all happened in
the server side, actually you reduce overhead on the client side.
Rob W: Rob Wilton, Cisco. So I'm still not convinced that this particular use
case is actually a problem. I'm not convinced there's scale issue in terms of
configuration here. Even if you split out the number of VLANs over hundreds of
interfaces, then I still think the amount of config is gonna be 10k or 20k.
Something that would be small in terms of the master consuming it. so I'm not
convinced by that aspect, that it is a problem here to be solved. In terms of
if you want to do more advanced VLAN operations, ie breaking tags or inserting
tags into into particular strings, for example then, yes, that's okay, but I
think of those maybe just be rpcs potentially on a VLAN model is how I would
implement those. So then, coming back to the general inline actions, I'm still
sort of conflicted is where this is a good thing to do, I think this more
generally is about transactions, and saying I want to give a sequence of events
to the server as one transaction, after it perform all of these things and
either succeed or fail, so that's the the guide I'd look at for this problem,
rather than just adding actions into configuration requests, but even with that
I still question whether that's a useful thing or not, I'm not convinced this
is a problem to be solved at this stage.
Qin W: but a you say that you don't know where it came, actually, we use a
edit-config as an example, if you modify the VLAN tag around the merge and a
split, actually, you may need, because you may operate on some list that leads
to the key index cannot be deleted, so you have several disparate range, so you
need to delete several disparate range first and then you create then a new
range with larger range, that's a difference with NETCONF, we want to address
Rob W: okay, so that's a different problem, potentially, to solve and I think,
again, we need to look at the data model that you're talking about, so the one
I've been through, IETF, doesn't have that issue, as the VLANs are just
information on a subinterface, so it's not actually something where you have
this concern, it's just manipulating that string, and the ability of a client
to mangle VLAN IDs into string is it's probably bordering on trivial to do, I
mean, it's that an easy of a thing to solve. so it might be that your data
model is different and then hence there's a different requirement coming from
that, so I'm happy to look at what your specific data model is to see what the
changes are, what's different
Qin W: yeah, we do have such a model. we can show you offline about this
Rob W: okay.
Mahesh J: Mahesh, as a contributor, adding to Robert's concern for why we
might need this, is the problem specifically for case where we're talking about
a range, like we're trying to specify whether we're trying to expand it or
break it up? is that the specific use case that we looking a solution for?
Qin W: yeah the case we give is maybe kinda limited, but we really want to
generalize this idea. The general idea, actually, we can provide the operation
for NETCONF for protocols so you can actually improve the NETCONF efficiency,
and here we gave the example the VLAN tag range, the value is interval type,
maybe there's some other case where the value actually is a string type, and
you also do this merge operation
Mahesh J: what other use case would you have?
Qin W: in some case where you haven't bring up actually, so you may transpose
some learned configuration into static configuration or dynamic configuration
data. By the way, we only talk about these cases, but we have some other cases
with haven't bring up
Mahesh J: yeah, okay, I think if you bring those cases, it might help the
workgroup appreciate and understand the problem a little better.
Qin W: we can do that
Kent W: Kent, as a contributor I agree with Rob. I don't understand the
motivation for wanting to solve this. I guess scalability and efficiency, but
does it really get to the level of concern that we need to solve the problem,
and that the solution seems like a point solution, and the fact that it's
NETCONF-only is concerning, I request that we have a solution that works for
both NETCONF and RESTCONF. If it is truly a transaction-like mechanism, I
think that's what Rob was saying, then maybe enabling YANG Patch (note: Kent
accidentally said "push") to be used by NETCONF would be another way of
enabling something like this. Then going to Mahesh's comment right here, if
it's truly just for ranges, then it seems like maybe we'd want to have a
datatype, a "typedef range", and then this operation would be available
whenever that typedef was in play. it's just unclear at the moment, I guess
going to Mahesh's last point, more examples and data analysis is needed, it's
currently unclear why we would want to pursue this.
Qin W: yeah I think the intension is that we provide such kind of solution,
hopefully we can generalize this so we can not only apply to the YANG Push, but
also can apply to the existing NETCONF operation, so not limit to the existing
NETCONF protocol operation. we haven't investigated how these can be applied to
YANG Push either, if that's a case, we think maybe first we need to clarify the
problem space first. I think we should look further to apply to that YANG Push.
Kent W: you're saying "YANG Push", but you mean to say "YANG Patch", right?
Qin W: YANG Push, not a YANG Patch
Kent W: how is it related to YANG Push?
Qin W: oh, you meant, you mention...
Kent W: YANG Patch
Qin W: patch, right patch, oh sorry I miss
Kent W: no worries