Re: [Netconf] LC on subscribed-notifications-10

["Eric Voit (evoit)" <[e-mail]>]

And <Eric8>.

Comments below reflected within:
https://github.com/netconf-wg/rfc5277bis/blob/master/draft-ietf-netconf-subscribed-notifications-13.txt

From: Kent Watsen, May 23, 2018 8:44 PM

Hi Eric,

Now <Kent7>.    BTW, I wish we switched to ASCII a while back…
Also, I removed highlighting because it was getting confusing.

Kent


<snip/>


<Eric3>  Added the descriptive paragraph requested in the middle of the three paragraphs below...

It is possible to place a start time on a configured subscription.  This enables streaming of logged information immediately after restart.

Replay of events records created since restart can be quite useful.  This allows event records generated before transport connectivity was supportable by a publisher to be passed to a receiver.  In addition, event records logged before restart are not sent.  This avoids the potential for accidental event record duplication.  Such duplication might otherwise be likely as a configured subscription’s identifier before and after the reboot is the same, and there may be not be evidence to a receiver that a restart has occurred.  By establishing restart as the earliest potential time for event records to be included in notification messages, a well-understood timeframe for replay is defined.

Therefore, when configured replay subscription receivers first become ACTIVE, buffered event records (if any) will be sent immediately after the "subscription-started" notification.  And the leading event record sent will be the first event record subsequent to the latest of four different times: the "replay-log-creation-time", "replay-log-aged-time", "replay-start-time", or the most recent publisher boot time.

<Kent3> Hmmm, I'm having a negative reaction to the "event records logged before restart are not sent" bit.  I know what you are trying to do, but I worry that this behavior might drop important logs, perhaps to the advantage of an adversary.  Note that some devices implement an <edit-config> with a restart.  Maybe the solution should require publishers to maintain a per configured-subscription awareness of (roughly) which log was sent last?   - and notify the receiver when a restart has occurred, or when the replaying of events occurs, so that they can be aware that there might be some duplicates?

<Eric4>  The current solution guarantees no duplicates, and also informs the receiver of each new “start-time”.  This allows the receiver to attempt to reconstruct any gaps from the last event previously pushed, should the choose to attempt such reconstruction.   As a dynamic subscription has no such boundary constraints on replay and boot time, all a subsequent dynamic subscription needs to do is to request the events between the last received event previously received from that configured subscription and the new replay-start-time.

<Kent4> So, the receiver is informed of each new "start-time" via the "subscription-started" control message, and then MUST do a short-lived dynamic subscription to scoop-up any possibly-missed logs, for which there may be none?   If we choose to keep this behavior, the draft should say this more clearly, perhaps in the Security Considerations section…

<Eric5> Added the following text as the last paragraph in the Implementation Considerations Section...

For configured replay subscriptions, the receiver is protected from duplicated events being pushed after a publisher is rebooted.  However it is possible that a receiver might want to acquire event records which failed to be delivered just prior to the reboot. Delivering these event records be accomplished by leveraging the “eventTime” from the last event record received prior to the receipt of a “subscription-started” state change notification.  With this “eventTime” and  the “replay-start-time” from the “subscription-started” notification, an independent dynamic subscription can be established which retrieves any event records which may have been generated but not sent to the receiver.

 <KENT5> Is this really limited to reboots, or any restarting of a configured subscription?

<Eric6> The protection from event replication is just for reboots – this is because the configured replay start time might duplicate events already sent from a previous boot.

<Kent6> So, in the case of a suspend/resume, the device picks up where it left off?   Or is it also just starting with "current"?  (strangely, I'm having trouble finding this answer in -13).  If it picks up where it left off, does this mean that the device has index of some sort and, if so, why can't that state be persisted in the case of a graceful shutdown/restart?

<Eric7> It would be *very* difficult to persist on the publisher which event actually was received last by a receiver.  However the receiver can make that determination, as it does maintain that state.  Look for my proposal below on how to handle this with a new leaf in “subscription-started” “replay-previous-event-time”

<Kent7> good point about what was *received* as opposed to what was *sent*, as they may differ, but then would the publisher be able to say the last event it sent?  As I recall your proposal from before, we the receiver *always* needs a dynamic subscription to fill in any *potentially* missing gaps.  My goal is to try to change the "always" to a "sometimes" by giving the receiver enough info to know there is a gap, or perhaps just know when there wasn't a gap but, in either case, it becomes less than "always".

<Eric8> Yes, with the proposal now embodied in the latest draft version linked above, it is now a “sometimes”.

In the “subscription-started” notification, the controller is passing the timestamp of the immediately prior event stored in its replay buffer.  Whether the publisher tried and failed to send this immediately prior event a receiver or not doesn’t actually matter.  All that matters is that that receiver can sees whether this time matches the its last received event record.  If it does not, then a dynamic replay can be initiated.





But the mechanism to retrieve lost events is common between reboots and restarts.  I.e., using this solution, a receiver never needs to know a reboot occurred.  It just knows that a subscription restart occurred, and there might be a gap in events which should be retrieved via a dynamic subscription.

By having the receiver follow the same process for both reboots and restarted subscriptions, it simplifies recovery procedures.

<Kent6> I'm unsure about the "same process" (mentioned in my previous comment).

<Eric7> It is the same process.  Note that below, the “replay-previous-event-time” tweak allows a receiver to determine if there are events missed.  And with the tweak, the fact a reboot has happened is still not explicitly exposed.

<Kent7> Sorry, I didn't mean I didn't know if the process was the same between the two interactions, but what the "same process" is.   Nonetheless, what you write above ("allows a receiver to determine if there are events missed") is music to my ears, and I see now that you've just now added the "replay-previous-event-time” (that wasn't clear to me at first, I had to look at -13 to be sure).  I'll look at it more closely in the section where it is introduced below.  Regarding your last sentence, I take it that it not being explicitly exposed is a good thing, right?

<Eric8>  Yes, a publisher not being absolutely forced to expose a reboot is good.




Next, I view this as a Security Consideration, since nefarious things can happen when a device reboots and note that a DoS could extend the gap to hours or days.  I think that this draft (the text above) is watering down the issue.  In my view, this is a huge consideration, along the lines of a receiver really MUST always try to use a dynamic subscription to fill in any gaps.

<Eric6> I agree this is a best practice, but a MUST would require telemetry receivers to have to support dynamic subscriptions.  I can see IoT receiver implementations where this wouldn’t be likely.

<Kent6> Good point, but I don't see in the YANG module establish-subscription being optional to implement.  Is that an oversight?

<Eric7> This is ok.  It is totally fine for a publisher to support dynamic subscriptions.  But a receiver need not.  Perhaps a lightweight IoT client just will just be a configured subscription receiver.

<Kent7> I think you misunderstood me.  Using your IoT example, even though a device (or the entire IoT space) only uses configured subscriptions, the current module doesn't enable a server to not support dynamic subscriptions.   For constrained devices, having to implement something never used could be a problem…

<Eric8>  The requirement is that a publisher must support a dynamic subscription.  There is no requirement for that on a configured receiver.   To clarify this, I have tweaked early “Configured Subscriptions” section text to say:

On the publisher, supporting configured subscriptions is optional and advertised using the "configured" feature. On a configured receiver, support for dynamic subscriptions is optional except where replaying missed event records is required.






To cover your desire to highlight this in security, I have added the following requirement to the Security Considerations Section:

When a configured receiver gets a new “subscription-started” message for a known subscription where it is already consuming events, the receiver SHOULD retrieve any event records generated since the last event record was received.  This can be accomplish by establishing a separate dynamic replay subscription with the same filtering criteria with the publisher.

<Kent6> that's helpful.  Perhaps tack onto the end of the last sentence something like ", assuming the publisher supports the ['dynamic' and] 'replay' feature[s]"?

<Eric7>  Added
“, assuming the publisher supports the "replay" feature.”
As dynamic isn’t an optional feature.

<Kent7> okay, that's expected, but see my previous Kent7 comment, I think that we might want dynamic to be optional.

<Eric8> Per above, receiver support for dynamic is what is optional.



It almost begs the question for why configured subscriptions are supported at all.

<Eric6> Because there have been lots of industry requests.  And lots of implementations with non-IETF solutions.  Personally I prefer dynamic subscriptions.

<Kent6> I can see the appeal of configured subscriptions, but if each receiver of a configured subscription SHOULD always do a dynamic subscription on every publisher reboot, it is already being coded to support dynamic subscriptions, at which point it seems that maybe it should've just been implemented to support dynamic subscriptions from the start (on top of the publisher-initiated call-home transport, of course).   Can you say some more about these industry requests and, in particular, if the requests are still strong even when made aware of the need to also do dynamic subscriptions?

<Eric7>  Here is an industry request:  stream event record entries being placed into security logs (e.g., Integrity Measurement Architecture audit events).   These events are only relevant starting at boot time.  Losing events means a system cannot be validated.  And days after a boot, streaming all the events since boot would result in too many events.   So, in this environment, doing a dynamic subscription to fill the gap is acceptable.

<Kent7> meh




<snip/>



<Kent4> this I agree with, but I really don't like the fact that receiver MUST do a short-lived dynamic subscription to scoop-up any possibly-missed logs, for which there may be none.  Perhaps we could add more values into the "subscription-started" notification message that would enable to receiver to make a local determination if such a dynamic subscription would be  helpful?



<Eric5> I recommend against providing extra objects/reasons in the “subscription-started” at this time.  Publishers might not want to advertise a reboot, and they might not want to advertise why there was loss in event continuity.   All that should matter to a receiver is that such a discontinuity existed, and they have a way to try to fill event the gap should they care.  If the need for more data and the cause of the discontinuity turns out to be required, we can always augment here with future objects.



<KENT5> first, I'm still not 100% sure if this is just a reboot problem, or any time the subscription is restarted/resumed.



<Eric6> Per above: retrieving missing event records is not a reboot specific problem.  But unintentionally replicating event records is reboot specific.  (Otherwise the configured replay-start-time would drive a repeat of everything on each and every reboot.)



<Kent6> okay, I think I got it this time.  Having a *configurable* replay-start-time is so confusing.  Is it really worth having?



<Eric7>   Yes it is worth having.

(a) In many environments, reboot is very infrequent.  Without configurable start time, an operator setting up a configured subscription would not have the ability to designate what to send.  It could only send the full log (at whatever size).

(b) on-publisher security or troubleshooting diagnostics might identify a breach or some event where streaming recent historical event records is a MUST.  As a result, it might want to stream a subset of event records off a box going back in time to potential events which might have been evidence or contributing factors.



<Kent7> Let me come at this another way.  Assume we drop all support for *configurable* replay-start-time.  As such, configured subscriptions always start with the next-generated event (no replay at all).   This covers most use-cases, right?   For those receivers that really wanted the older logs, can't they just do a dynamic subscription to collect them, same as we've been discussing above?



<Eric8> Some reasons this might not always be practical:

(a) IoT devices just might want to passively listen to event streams of Telemetry.  (I.e., this would force configured receivers to support dynamic subscriptions.)

(b) This forces complexity onto applications which only ever need to track what has happened since boot.  (E.g., per above, continuous Integrity Measurement Architecture (IMA) boot log streaming and evaluation.)

(c) Publisher access permissions for who can use the establish-subscription RPC might have to be expanded to include lots of configured receivers.  This might open up a vector to control plane DDoS.  Right now the access permissions would just have to allow the receiver read access to the event records.

(d) A publisher may choose to firewall classes of receivers (or locations of receivers) into a listen-only mode without the ability to establish subscriptions.







Next, can you explain what info could be provided that might allow a client to make a local decision as to whether a dynamic subscription is needed of not, and does the "replay-start-time" node provide it?



<Eric6> The local decision is easy.   As a configured receiver, I have just received a new subscription-started state change notification for a subscription I am tracking.  If I *need* to know about any events which occurred since the last event record I received for that subscription-id, then create a separate replay to cover that time gap.



<Kent6> it's not "if I need to know" so much as "I *know* that I'm missing some events because X and Y and, therefore I will create a dynamic subscription otherwise, since I know that I didn't miss any events, I will not create a dynamic subscription."   I'm trying to figure out what X and Y might be…



<Eric7>  Right now the evidence a configured receiver has is:

X = A subscription #3333 exists, the last event record came in at 2:01:55UTC.  I never got a “subscription-terminated” or “subscription-concluded”.

Y = a new “subscription-started” arrives.



Thinking about how to accomplish what you are asking, I do think there might be a possibility.   We could place a new object “replay-previous-event-time” in the subscription-started notification:



       +---n subscription-started {configured}?

          +--ro (target)

             +--:(stream)

            +--ro replay-start-time?       yang:date-and-time        {replay}?

            +--ro replay-previous-event-time?       yang:date-and-time        {replay}?



What this object would do is to provide the timestamp of the event record just prior to the “replay-start-time” in the event buffer (should such a previous event exist).   If this timestamp doesn’t match the last event a receiver got (e.g., 2:01:55UTC), then a dynamic subscription could be established to fill the information gap.



<Kent7> Right, something like that.  Do you think this does it, or are there a bunch of disclaimers around its effectiveness?



<eric8>  I know of no disclaimers.



/Eric8





<snip/>



<Eric4>  Tweaked a Section 2.4.2.1 sentence to say:



This document puts no restrictions on the size or form of the log, where it resides within the publisher, or when event record entries in the log are purged.



I suggest adding text that clarifies this, and details the need for a short-lived dynamic-subscription.



<Eric4> The tweak above, with the suggested text in the Implementation Considerations section above hopefully covers this.



<KENT5> the "purged" part helps, but whyis this information buried inside a section titled "Requesting a replay of event records"?



<Eric6>   This is the second paragraph of the first section which discussed replay.  It is the sentence after the optional feature of replay is introduced.



<Kent6> I understand how it fits into this section, but it seems like it should be in a section called something like "replay log", since it's equally applicable for configured subscriptions (and s2.4 is about dynamic subscriptions).  Of course, if we remove replay from configured subscriptions, then leaving it here makes sense…



<Eric7> Per above, there is lots of value in configured replay.



<Kent7> Yes, but see again my challenge to that assertion.  I'm leaving this here in case the decision is flipped…





<snip/>