Hi Michael, yes, I understood your main concern about “half-measures”, and I agree with the sentiment. Any TOSCA orchestrator that wants to claim “service template portability”
must be able to establish a “contract” with any external entity to which it delegates to make sure that orchestrator state can be sync’d up after the delegation finishes. This is true whether you just delegate to a designer-provided script/artifact (e.g. to
implement a lifecycle operation), or to an entire workflow.
We’re currently in the process of incorporating this “contract” approach, but only for artifacts that implement lifecycle operations. Running a template designer-provided
script is in fact an example of an orchestrator “delegating” to an external entity, and then taking back control after the script finishes. We’re looking at the types of additional metadata that need to be provided for operations in order to establish the
“contract” between the script and the orchestrator. For example: what is the node state in which the script can be run? What is the state of the node after the script completes. What node attributes may get modified by the script, and how do we retrieve the
updated values? How long should the orchestrator wait before the script “times out”?
We’re currently doing this work for lifecycle operations only, but not yet for workflows that operate on entire topologies. It’s proving challenging (and time consuming) enough
with the small group to do this for operations, and I imagine that doing this correctly for entire topology workflows may be significantly more complex. In fact, we’re speculating that doing this for workflows might require a standardized instance model (and
associated APIs) into which the resulting state can be reflected. It could take several months (if not several quarters) to get this right. That’s why I recommended we should be OK in the meantime with the “delegate and forget” approach that’s currently being
used in systems like Open-O. At least it gets people using TOSCA models, which will hopefully be a step towards using TOSCA’s declarative approach (augmented with TOSCA imperative workflows where necessary). If we don’t embrace this approach, we run the risk
of disenfranchising people who were early adopters of TOSCA, and who may give up on the standard as a result. That wouldn’t be good for the TOSCA ecosystem.
By the way, by way of comparison, there is a very rich YANG ecosystem out there that has generated thousands of YANG models (https://yangcatalog.org/).
Given how much richer TOSCA is than YANG in terms of functionality, our ecosystem should strive to be orders of magnitude more active.
Chris
From: Michael Brenner [mailto:michael@gigaspaces.com]
Sent: Wednesday, April 26, 2017 1:22 PM
To: Chris Lauwers <lauwers@ubicity.com>
Cc: Matt Rutkowski <mrutkows@us.ibm.com>; Huabing Zhao <zhao.huabing@zte.com.cn>; tosca@lists.oasis-open.org
Subject: Re: [tosca] Groups - Issue_TOSCA318_Lack of BPMN BPEL support-v-2017-04-25.pptx uploaded
Hi Chris,
I appreciate the challenges. Let me clarify that I do not discourage the native imperative workflows defined in TOSCA, and I don't think that what I suggest is in conflict with Matt's view. What I am not a big fan of is half-measures that
lead to lack of portability/inter-operability ... and that was the main point of my comments.
Let me try to parse what I am saying:-)
When TOSCA promotes its native (declarative OR imperative workflows), all one describes is the intent at the end of the workflow, and nothing more. I interpret this (and any TOSCA-consuming orchestrator would do the same) that using the
service template and the provided artifacts, the TOSCA-consuming orchestrator is able to accomplish the task - and it absolutely does not matter through what mechanism it does so (and that could include the use of BPEL/BPMN by some particular TOSCA-based orchestrator,
if that is how it implements workflow). But TOSCA service template in this case is "workflow implementation neutral" (i.e. unaware of the engine, BPEL/BPMN/embedded in the orchestrator, or whatever else you may have) which is in-line with the TOSCA philosophy
of under-specifying things (a good principle when it comes to standards that are build to last). It allows ANY orchestrator to fulfill the intent the best way possible, and does not favor or disadvantage one implementation or another. Which makes the spec
support portability/inter-operability. And it gets you away from specifying all the complexities I mentioned to make the spec right. Over-standardizing is one the fireproof mechanism to "date" a standard, and it won't last.
But if we choose to over-specify (I don't have a Quijotian syndrome:-) then we MUST specify in such a way that the spec continues to be portable/inter-operable - which means that once we go down the slippery slope of explicitly supporting/endorsing
external workflow engines, we better specify the mechanism in full, because otherwise we deliberately introduce lack of portability and inter-operability (let aside of delaying implementations or making them more complex than necessary).
Do we want to take the red pill or the blue pill?
Btw Chris, here is the attached write-up on workflows. By now we have advanced to further analyzing the topic in-depth on this thread and in the TOSCA TC meetings, but I remembered you showed interest a few weeks ago - so I am sharing it
here.
On Wed, Apr 26, 2017 at 2:52 PM, Chris Lauwers <lauwers@ubicity.com> wrote:
I agree with Michael that our goal should be to strengthen the appeal of TOSCA. But I also agree with
Matt that we should be pragmatic about this. The reality we’re faced with is that (as far as I know) there aren’t any interoperable implementations of TOSCA out there today that adopt a truly declarative paradigm. The implementations I’m aware of use TOSCA
mostly for modeling but not necessarily for orchestration. Instead, they rely largely on imperative logic (such as the BPNM workflows in Open-O) provided by the service template designer (or worse, hardcoded in the Orchestrator for very specific use cases)
to get a TOSCA service model deployed. What we need to get to instead is a “true” TOSCA orchestrator that provides a “TOSCA runtime” that is able to process any standards-based declarative TOSCA service template. Said a different way, we need to get to the
point where people use TOSCA for its orchestration features rather than (or in addition to) its modeling features. We’re not quite there yet (although some of us are working on it
J )
It appears we have differing opinions about the best way to get there. We could try to discourage the
“imperative” use of TOSCA (as Michael suggests) with the goal of promoting the declarative paradigm, or we could support it with the goal of getting broader adoption of TOSCA models (as Matt suggests) with the goal of replacing service-specific imperative
logic (such as BPMN workflows) with generic declarative support in the Orchestrator over time.
Chris
From:
tosca@lists.oasis-open.org [mailto:tosca@lists.oasis-open.org]
On Behalf Of Michael Brenner
Sent: Wednesday, April 26, 2017 7:07 AM
To: Matt Rutkowski <mrutkows@us.ibm.com>
Cc: Huabing Zhao <zhao.huabing@zte.com.cn>;
tosca@lists.oasis-open.org
Subject: Re: [tosca] Groups - Issue_TOSCA318_Lack of BPMN BPEL support-v-2017-04-25.pptx uploaded
Hi Matt,
Thanks for the feedback, and certainly, I wish too to have been part of the discussion ... it just my wishes/interests and my committed tasks have the tendency to be able to not
always coordinate:-). Like for many of us...
I'll do my best to be in other calls dedicated to this topic - but discussing over email, and reaching some tentative agreement on the issues/requirements before you propose solutions,
works best for the kind of schedule I have, at least. I think if the requirements are agreed, the solution may be more straightforward to be agreed.
On Wed, Apr 26, 2017 at 9:42 AM, Matt Rutkowski <mrutkows@us.ibm.com> wrote:
Hi Michael,
Wish you had been on the call yesterday (perhaps you can read my meeting notes posted to TOSCA?)... We touched on most of these issues with Luc echoing many of your thoughts.
In the end, TOSCA has taken the philosophy of recognizing that exiting providers use existing tooling for installs/configs (including BPEL/BPMN workflows) and that we wanted to embrace these providers
and recognize/point out the pitfalls and encourage them (when able) to move to declarative. In addition, if we can treat these tools (including ansible, or other) as "black boxes" that can return us to a known state (consuming inputs in the model and providing
access to outputs) that is the general view of processing. Of course, we have other proposals for operational "hints" on expected processing time (will not call it timeout) for the orchestrator to gauge how long it should continue.
All agreed, that an instance model/API would be needed to allow full introspection after the "black boxes" complete and that these workflows should not cause changes to the other parts of the topology.
These are all things we will discuss in text. Please also account for the Artifact Processor entity which will also have information for the orchestrator (for invocation) and can also have properties/configurations, etc.
In addition, we agreed not to call this "delegate" as our existing delegate does not match this behavior proposed.
Can you attend the WG call in subsequent weeks?
Kind regards,
Matt
From: Michael Brenner <michael@gigaspaces.com>
To: Huabing Zhao <zhao.huabing@zte.com.cn>
Cc: tosca@lists.oasis-open.org
Date: 04/26/2017 07:57 AM
Subject: Re: [tosca] Groups - Issue_TOSCA318_Lack of BPMN BPEL support-v-2017-04-25.pptx uploaded
Sent by: <tosca@lists.oasis-open.org>
Hi Huabing, all:
I reviewed the proposed changes to support external workflow engines via "delegate" mechanism, and I think this is something that can complement TOSCA in special cases - e.g. re-using workflows previously implemented outside of TOSCA, in particular those requiring
interactions between external entities.
The problem that I have is that the current proposal leaves us with many unsolved issue, and renders a TOSCA-consuming orchestrator with more questions than the answers that an external workflow engine can provide. I would prefer us to work on resolving all
the issues to provide a workable and portable solution, rather than inserting a half-baked mechanism that we do not know yet how it will converge to the full solution.
Here are some of the issues I observe are unsolved:
1. TOSCA philosophy is to specify "what" (intent) and not "how" (implementation). How a TOSCA consuming orchestrator resolves the TOSCA-driven requirements is typically irrelevant, but the expectation is that the result fulfills the requirements. This philosophy
is now infringed by the proposal, because of several issues - see below.
1.a: If we "standardize" the artifact, i.e. have some standard artifact name or suffix that a TOSCA parser needs to recognize, we are going from "what" to "how". And we also create a precedent that would require constant changes - because other such "keynames"
will emerge. Perhaps this could be solved with a separate registry, but I don't think it should be standardized as part of the TOSCA spec.
1.b: on the other hand, if the parser does not have a way to identify whom to delegate, it is at a loss about what to do with the artifact. So it is sort of a catch 22 situation.
2. Assuming we find the right solution for 1. above, we still have a major portability issue. Again, typically a TOSCA-consuming orchestrator should not understand/parse the content of the artifacts, but just execute them or pass them along to other artifacts
for consumption. The proposal does not specify what the TOSCA-consuming orchestrator is supposed to do with the artifact that is passed to the delegate workflow. And that gets us into a different catch22.
2.a: If the delegate artifact is opaque to TOSCA-consuming orchestrators, 2 different orchestrators may process the artifact quite differently in the best case, or would not know what to do it (beyond parsing) in the worst case. So this requires a clear specification
of how the TOSCA-consuming orchestrator is supposed to process the artifact, in order to consistently get the same result. Obviously, that moves us from "what" to "how".
2.b: Assuming there is a good solution to 2.a above, we run into a different issue, and that becomes an implementation and compliance issue. Each TOSCA-consuming orchestrator would have to be able to support the mechanism described for processing the delegate
artifact, but for every different artifact (e.g. BPEL vs. BPMN ... and this is just the beginning precedent), the mechanism would have to have specific parts, in addition to the generic parts, most likely. That basically forces a TOSCA-based orchestrator implementation
to support in a standard way ANY potential mechanism to delegate to any potential workflow engine. This is not something one should ask a vendor that will be forced into compliance, and it is completely atypical for a standard to ask for this. The solution
would be to define a one-time generic mechanism of passing the artifact, regardless of which what external workflow engine is required to process the artifact.
3. TOSCA service template defines a topology, with nodes that have certain states. When control is being passed to the delegated external workflow engine, it is not clear in the proposal what the expectations (and constraints) are with respect to changes to
the topology and/or nodes and/or states of the nodes by the external entity. A stable solution would be advantaged by a single source of truth for topology, nodes and states - in this case that source of truth is the TOSCA-consuming orchestrator that implements
and maintains the TOSCA service template. That would argue that the delegated workflow engine cannot change topology, nodes or node states? If this is the intent, it should be stated in the changes to the TOSCA spec. If the intent is different (i.g. external
workflow engine can change topology, nodes, and their states) then how do we ensure stability of the solution? No mechanism is included in the proposal for communicating back to the TOSCA-consuming orchestrator what has changed and how.
4. That brings us to another general issue. While the proposal attempts to define a mechanism that gives an external entity the delegation to execute a workflow (and I described above why that portion needs more work to ensure portability), the proposal does
not address how the control is returned to the TOSCA-consuming orchestrator, neither is it defined when the control is returned. TOSCA-consuming orchestrator is the master/manager of the topology/nodes/states that it deployed or is in course of deploying.
It uses the topology and the relationship between nodes to traverse the topology in a certain order, and execute the lifecycle management of the different nodes in the topology based on the declaration of intent in the model. Therefore, any mechanism delegating
work outside the TOSCA-consuming orchestrator is incomplete unless it describes how the control is returned to the TOSCA-consuming orchestrator, and when, and with what results.
On the how/when, I can see several options that would have to be defined:
4. a: delegate and forget. In this case, the TOSCA-consuming orchestrator does not block to wait the result of the external workflow execution, and instead continues to work on implementing the intent described in the service template. No results are expected
from the external workflow processing that are needed by the TOSCA-consuming orchestrator to know about.
4.b: delegate and sync up later. This is a similar case as 4.a, but in this case there may be results from the external workflow execution available later on, that are of interest. In this case, a mechanism needs to be defined to allow a TOSCA service template
to describe how to wait/block for those results later on.
4.c: delegate and wait/block indefinitely. In this case, as soon as TOSCA-consuming orchestrator delegates a workflow externally, it would stop executing any further, and wait for the workflow to complete and return control - possibly with results.
4.d: delegate and wait/block for a period of time. This is similar to 4.c above, but a timer is associated, and if the external workflow engine does not return control to the TOSCA-based orchestrator within the defined time interval, the latter will continue
execution. A further option would be to add also the mechanism described in 4.b, i.e. to allow for a timed-out case to sync up again later.
There may be other situation we need to consider. In any case, the main point I want to make is that we do need to think about these issues and provide a consistent and portable solution. This is a complex mechanism if we want to do it right. And not doing
it right ... well, it is just not right, and weakens, rather than strengthening the appeal of TOSCA specification. I am sure that we can all agree that our interest is to make the TOSCA spec better.
Best regards,
Michael
On Tue, Apr 25, 2017 at 10:47 PM, Huabing Zhao <zhao.huabing@zte.com.cn> wrote:
--
--
--
|