Hi, Kent,
Please see my reply inline.
From: Kent Watsen [mailto:kent+i...@watsen.net]
Sent: Friday, July 16, 2021 2:55 AM
To: maqiufang (A) <maqiufa...@huawei.com>
Cc: netmod@ietf.org
Subject: Re: [netmod] system configuration sync mechanism
Hi Qiufang,
[snip]
The question is if the server implementation prunes dangling/unused objects
when <intended> is applied, updating <operational>. My assumption is that the
server will discard any object that doesn’t actually impact the running
configuration of the system (i.e., values are consumed by the underlying
operating system, drivers, etc.). Thusly, it is my opinion that only the
referenced objects are applied. Hence why, to answer your last question, I
wrote that these configurations (manufacturer-defined objects) are not applied
immediately but only after they are referenced. Makes sense?
[Qiufang Ma] Yes, try to sum up our discussion about the categories of the
system configuration:
· Physical-resource-dependent--> whether this sort of system
configuration exists in <system> dependents on if the physical resource is
present(e.g., physical interface).
· Physical-resource-independent-->which is provided by the device system
o Further classification from the perspective of “applied” time(dependents on
whether the system configuration impacts the running of the system)
§ Config that is applied immediately(e.g., the loopback, the predefined
minimum length of password…)
§ Config that is applied only after being referenced by other configs(e.g,
definitions for applications ftp/tftp…)
o Further classification from the perspective of generation time
§ Config that is generated unconditionally at each boot time(e.g, loopback,
predefined minimum length of password, ftp/tftp…)
§ Config that is generated conditionally during the device running(e.g.,
system-generated local-port and remote-port for a new established BGP
connection)
Keep in mind that what is described above is just one aspect of what can be in
<system>. In addition to defining reference-able objects, <system> can also
define/apply configuration immediately (e.g., the loopback interface). That
is, configuration not does not have to be referenced in order to become
activated.
[Qiufang Ma] Noted.
Note that, <running> by itself would not pass validation, due to missing
leafrefs. Thankfully, NMDA never says that validation runs on <running>. But
once <running> and <system> have been merged, to become <intended>, the result
does pass validation.
[Qiufang Ma] The referenced instance must also exist for the data to be valid
since the require-instance defaults to true if not present. Is this what you
had in your mind? Yes, NMDA says that it is <intended> which is subject to
validation. But I also notice that In section 5.1.3 of the NMDA:”<running> MUST
always be a valid configuration data tree, as defined in Section 8.1 of
[RFC7950]. ” So my thought here is that <running> should also conform to the
YANG model constraints and that’s to say, a referenced system-defined data item
should also exists in <running>.
Therefore, if system configurations do not exist in <running>, they still need
to be configured in <running> manually in order for being referenced. In this
case, the original purpose of predefining some system configurations for user
convenience is lost. This is the reason why we would like to define some
mechanism here to synchronize <system> into <running>.
I see in RFC 8342 "<running> MUST always be a valid configuration data tree, as
defined in Section 8.1 of [RFC7950]”.
But the question remains if it is possible for the system is able to validate
<running> without, e.g., expanding templates. There may be a 'leafref' or
‘must’ expression somewhere that will fail because the evaluation occurs
without expanding a template that supplies the missing parts.
[Qiufang Ma] On condition that <running> should be valid, the operators will
need to retrieve from the <intended> or <operational> to get the
template-expanded configurations and then create them in the <running>, right?
My feeling is that it loses the meaning of predefining and seems no differences
between operator-defined configurations if operators have to create system
configures in <running> before they use them. So I am beginning to think, if
it’s possible to expand the system-defined template during the copying between
<system> and <running>?
If this draft “updates” RFC 8342 (NMDA), then it can supply a clarifying
statement about what it means that "<running> MUST always be a valid
configuration data tree”. Either that, or an Errata if it’s determined that
the statement isn’t correct.
You make a good technical point, but I think that we should *want* to avoid
having to copy <system> (or <operational>) configuration into <running> if we
can avoid it. Agreed?
[Qiufang Ma] tend to agree. Maybe we should try to avoid it, unless we have to.
FWIW, also in RFC 8342, Section 5.1.4.:
<intended> is tightly coupled to <running>. Whenever data is written
to <running>, the server MUST also immediately update and validate
<intended>.
<intended> MAY also be updated independently of <running> if the
effect of a configuration transformation changes, but <intended> MUST
always be a valid configuration data tree, as defined in Section
8.1<https://datatracker.ietf.org/doc/html/rfc7950#section-8.1>
of [RFC7950]<https://datatracker.ietf.org/doc/html/rfc7950#section-8.1>.
> I am wondering if these configuration will present in the <operational>
> (which contains all the configuration actually used by the device) before
> they’re referenced.
I think that it would depend in the specific server’s behavior, regarding if
*unused* predefine objects are present in <operational>. Certainly the unused
objects would not have to be present in <operational>. If I were implementing
the server, the unused objects would NOT be present in <operational>.
[Qiufang Ma] Yes, if the predefined system configurations is unused, then I
also tend to agree that they would not be present in <operational> but may also
depend on the vendor implementation.
Yup, this is the same conclusion is in my response above.
[Qiufang Ma] :)
> It would be good if we could determine if there are any other
> "resource-independent” configuration categories here.
> [Qiufang Ma] Do you think there exists conditional system configuration (if
> the preceding configurations you mentioned above is not)? For example, if SSH
> is enabled on a device, SSH-related keys are automatically generated. Such
> configurations are generated at the moment when a special functionality is
> enabled.
I’m unsure what you mean in general by "conditional configuration”, but I can
speak to your specific example. Though I must preface my comments that I
imagine there are a number of ways servers might go about enabling `sshd`.
What follows is my personal view, forged by being around systems for awhile ;)
In general:
- `sshd` is NOT enabled by default.
- `sshd` is enabled via a configuration knob.
- the SSH host key is dynamically generated the first time `sshd` is
enabled.
- the SSH host key itself is in <operational> (not <running>)
This view is consistent with the first paragraph in Section 3 of the “keystore”
draft (reproduced below):
3. Support for Built-in Keys
In some implementations, a server may support built-in keys. Built-
in keys MAY be set during the manufacturing process or be
dynamically
generated the first time the server is booted or a particular
service
(e.g., SSH) is enabled.
As a closing thought, this model (which I stated upfront may not be universal)
would have no presence-in or interaction-with <system>…though, perhaps, there
may be some predefined values for what key-algorithms and/or key-lengths to use
when generating the SSH host key...
[Qiufang Ma] By “conditional system configuration”, I was meaning some of the
system configurations are not generated immediately after the device is powered
on. Instead, they are generated when a specific condition is satisfied during
the device running(e.g., a functionality is enabled due to some client
configurations). I am not sure if it really exists, maybe not, just try to
explore the possibilities of various resource-independent system
configurations.:)
To this point I agree..as does RFC 8342 (NMDA), Section 5.3.3.:
Sometimes, resources are controlled by the device and the
corresponding system-controlled data appears in (and disappears from)
<operational> dynamically. If a system-controlled resource has
matching configuration in <intended> when it appears, the system will
try to apply the configuration; this causes the configuration to
appear in <operational> eventually (if application of the
configuration was successful).
[Qiufang Ma] See above, as I summarized as the system configurations that is
generated conditionally during the device running.
Firstly, I again have to preface my comment that there are likely many ways
that templating mechanisms can be defined. But, in general, once a
templating mechanism has been defined, then it stands to reason that templates
could be defined either in <running> (by operators) or in <system> (by the
manufacturer). In one implementation I’m familiar with, the templates are
objects that are referenced/parameterized by other parts of the configuration.
(Same as with the predefined objects discussion above.)
To answer your questions:
1) Yes, it is my opinion that *activated* templates in <system> will be
expanded and present in <intended>.
2) I would never suggest that the system-defined templates are present in
<running>, though they may be referenced/parameterized by config in <running>.
3) if a config-template is configured in <running> (i.e., it is
operator-defined) then, yes, the expanded configuration in <intended> is
"client configuration” (note, "client configuration” is not a formal term).
That said, it seems fair to say that a template defined in <system> and then
referenced by "client configuration” in <running> is also expanded as "client
configuration” in <intended>.
4) I don’t not understand your last sentence, that the expansion of <system>
templates are only present in <operational>. Maybe you’re saying something
subtle, e.g., that servers currently don’t support GET on <intended>. But, in
theory, the expansion of <system> templates should (IMO) be present in
<intended>, so that they may be subject to validation. Of course, all the
<intended> configuration (whether originating in <running> or <system>) that is
successfully “applied” will also be present in <operational>.
[Qiufang Ma] Assume that there is no <system> and this work, the expansion of
system templates are only present in <operational>. Because this is compatible
with system configuration definition in NMDA.
But if system configurations are only present in <operational>, the predefined
system configurations still need to be retrieved and created into <running>
explicitly when being referenced. I think we’ve reached an agreement on the
need for <system> to exist, and our main point of disagreement is whether
<system> should be copied into <running>. Your point is that being merged into
<intended> is enough to make sure a success validation. But my understanding is
that the referenced system configuration data item must also exist in the
<running> to obey the model constraints.
Yes, I believe that you provided an accurate description of the difference in
our opinions. Per my earlier response, you make a valid technical point, my
goal is to waive that interpretation to the side so that a simpler solution can
emerge. It would be good to get other opinions on list, otherwise we’ll take
it into the meeting.
[Qiufang Ma] OK. Hopefully someone else would share some opinions here.
Otherwise let’s take this into the IETF meeting.
[BTW, in keeping with this thread moving from the NETCONF to the NETMOD mailing
lists, would it make sense to move the IETF 111 presentation slot from NETCONF
to NETMOD too? I think it does and, further, it would help with scheduling
(NETCONF is over, NETMOD is under). Would you be okay with this? AD Rob and
the NETCONF chairs discussed this morning, and think it's okay, but would still
need to confirm with the NETMOD chairs.]
[Qiufang Ma] I am happy with the proposal, if it’s also okay for NETMOD
chairs:-). I have sent an email to the NETMOD chairs to request to move this
presentation slot from NETCONF to NETMOD.
A new version of the draft will also be submitted to NETMOD when the
draft-submitting window reopens.
<big snip>
> I’m beginning to think that:
> · auto-copying into <running> is likely never a good idea, because it
> violates the definition of <running>
> [Qiufang Ma] I am quite aware that different datastores in NMDA represents
> different views of data nodes. And <running> represents a configuration
> datastore holding the current configuration of the device.
> Should we consider system configuration also be part of current configuration
> of the device? From my perspective, the difference between system
> configuration and client-configuration lies only in who provides it.
<running> holds the current *operator-specified* configuration of the device.
System-provided configuration is NOT specified by operators (though
system-defined objects may be referenced by operator-specified config in
<running>). I believe that this arrangement is consistent with the definition
of <running>. Agreed?
[Qiufang Ma] Yes. Actually we are not trying to violate the principles of NMDA
and the definition of <running>. The issue we try to resolve here is that
system configurations cannot be used(referenced or overwritten) by the
operators directly and need to be created into <running> explicitly. This
actually loses the meaning of “predefining and bringing convenience”. If
auto-copying is not a good idea, what do you think about defining an RPC
operation for the operators to do the copy(which is also what Rob suggests at
the meeting)?
If we have to copy into <running>, then I think that I agree an RPC
(<edit-config>?) would be better.
[Qiufang Ma] From my perspective, <edit-config> is feasible but not efficient
because operators still need to retrieve <system>/<operational> firstly. If we
could define a RPC to copy the entire <system> into <running>, it seems more
convenient for operators. However, some system configurations which are not
going to be referenced or modified may also be copied into <running>. I don't
have a strong feeling about which one is preferred. Anyway, we need to figure
out whether it would be fine for <running> to missing referenced system
configurations.
You mention “overwritten” by the operators? Why wouldn’t the operators just
define their own? For instance, if they don’t like the vendor’s
“vendor-foobar” object, they could copy/paste/edit their own “my-foobar” object
with the values needed, yes?
[Qiufang Ma] Yes, defining their own would be okay. By overwriting, I mean
sometimes the operators would like to modify the specific system configuration,
e.g., the MTU value of a specified interface(identified by its name).
If the operators want to modify the system configurations, there is no way but
redefine them in <running>.
Best Regards,
Qiufang Ma
> · having in <operational> doesn’t make sense, since the tweaks
> wouldn’t go thru <running> --> <intended> validation.
>
> I’m wondering if a model like below would work for everyone - thoughts?
> [Qiufang Ma] <intended> represents the configuration after all configuration
> transformations to <running> have been performed, so I think it is only
> coupled to <running>.
> Anyway, the <system> should also interacts with <operational>. Agreed?
I don’t agree that <intended> must only be coupled to <running>. Specifically,
I think that it is okay (compatible with NMDA) to define a <system> that also
impacts <intended>. This is the only (IMO) sane approach, as it enables the
combination <running> + <system> to be validated.
[Qiufang Ma] Please see above. If <running> is OK to miss referenced system
configuration, your proposal makes sense to me.
Ack.
Best Regards,
Qiufang Ma
Kent // contributor
_______________________________________________
netmod mailing list
netmod@ietf.org
https://www.ietf.org/mailman/listinfo/netmod
