[openstack-dev] [Heat] Convergence proof-of-concept showdown
anant.patil at hp.com
Mon Dec 1 07:02:29 UTC 2014
On 27-Nov-14 18:03, Murugan, Visnusaran wrote:
> Hi Zane,
> At this stage our implementation (as mentioned in wiki
> <https://wiki.openstack.org/wiki/Heat/ConvergenceDesign>) achieves your
> design goals.
> 1. In case of a parallel update, our implementation adjusts graph
> according to new template and waits for dispatched resource tasks to
> 2. Reason for basing our PoC on Heat code:
> a. To solve contention processing parent resource by all dependent
> resources in parallel.
> b. To avoid porting issue from PoC to HeatBase. (just to be aware
> of potential issues asap)
> 3. Resource timeout would be helpful, but I guess its resource
> specific and has to come from template and default values from plugins.
> 4. We see resource notification aggregation and processing next
> level of resources without contention and with minimal DB usage as the
> problem area. We are working on the following approaches in *parallel.*
> a. Use a Queue per stack to serialize notification.
> b. Get parent ProcessLog (ResourceID, EngineID) and initiate
> convergence upon first child notification. Subsequent children who fail
> to get parent resource lock will directly send message to waiting parent
> task (topic=stack_id.parent_resource_id)
> Based on performance/feedback we can select either or a mashed version.
> 1. Failed Resource tasks can be re-initiated after ProcessLog
> table lookup.
> 2. One worker == one resource.
> 3. Supports concurrent updates
> 4. Delete == update with empty stack
> 5. Rollback == update to previous know good/completed stack.
> 1. Still holds stackLock (WIP to remove with ProcessLog)
> Completely understand your concern on reviewing our code, since commits
> are numerous and there is change of course at places. Our start commit
> is [c1b3eb22f7ab6ea60b095f88982247dd249139bf] though this might not help J
> Your Thoughts.
> Happy Thanksgiving.
> *From:*Angus Salkeld [mailto:asalkeld at mirantis.com]
> *Sent:* Thursday, November 27, 2014 9:46 AM
> *To:* OpenStack Development Mailing List (not for usage questions)
> *Subject:* Re: [openstack-dev] [Heat] Convergence proof-of-concept showdown
> On Thu, Nov 27, 2014 at 12:20 PM, Zane Bitter <zbitter at redhat.com
> <mailto:zbitter at redhat.com>> wrote:
> A bunch of us have spent the last few weeks working independently on
> proof of concept designs for the convergence architecture. I think
> those efforts have now reached a sufficient level of maturity that
> we should start working together on synthesising them into a plan
> that everyone can forge ahead with. As a starting point I'm going to
> summarise my take on the three efforts; hopefully the authors of the
> other two will weigh in to give us their perspective.
> Zane's Proposal
> I implemented this as a simulator of the algorithm rather than using
> the Heat codebase itself in order to be able to iterate rapidly on
> the design, and indeed I have changed my mind many, many times in
> the process of implementing it. Its notable departure from a
> realistic simulation is that it runs only one operation at a time -
> essentially giving up the ability to detect race conditions in
> exchange for a completely deterministic test framework. You just
> have to imagine where the locks need to be. Incidentally, the test
> framework is designed so that it can easily be ported to the actual
> Heat code base as functional tests so that the same scenarios could
> be used without modification, allowing us to have confidence that
> the eventual implementation is a faithful replication of the
> simulation (which can be rapidly experimented on, adjusted and
> tested when we inevitably run into implementation issues).
> This is a complete implementation of Phase 1 (i.e. using existing
> resource plugins), including update-during-update, resource
> clean-up, replace on update and rollback; with tests.
> Some of the design goals which were successfully incorporated:
> - Minimise changes to Heat (it's essentially a distributed version
> of the existing algorithm), and in particular to the database
> - Work with the existing plugin API
> - Limit total DB access for Resource/Stack to O(n) in the number of
> - Limit overall DB access to O(m) in the number of edges
> - Limit lock contention to only those operations actually contending
> (i.e. no global locks)
> - Each worker task deals with only one resource
> - Only read resource attributes once
> Open questions:
> - What do we do when we encounter a resource that is in progress
> from a previous update while doing a subsequent update? Obviously we
> don't want to interrupt it, as it will likely be left in an unknown
> state. Making a replacement is one obvious answer, but in many cases
> there could be serious down-sides to that. How long should we wait
> before trying it? What if it's still in progress because the engine
> processing the resource already died?
> Also, how do we implement resource level timeouts in general?
> Michał's Proposal
> Note that a version modified by me to use the same test scenario
> format (but not the same scenarios) is here:
> This is based on my simulation framework after a fashion, but with
> everything implemented synchronously and a lot of handwaving about
> how the actual implementation could be distributed. The central
> premise is that at each step of the algorithm, the entire graph is
> examined for tasks that can be performed next, and those are then
> started. Once all are complete (it's synchronous, remember), the
> next step is run. Keen observers will be asking how we know when it
> is time to run the next step in a distributed version of this
> algorithm, where it will be run and what to do about resources that
> are in an intermediate state at that time. All of these questions
> remain unanswered.
> Yes, I was struggling to figure out how it could manage an IN_PROGRESS
> state as it's stateless. So you end up treading on the other action's toes.
> Assuming we use the resource's state (IN_PROGRESS) you could get around
> that. Then you kick off a converge when ever an action completes (if
> there is nothing new to be
> done then do nothing).
> A non-exhaustive list of concerns I have:
> - Replace on update is not implemented yet
> - AFAIK rollback is not implemented yet
> - The simulation doesn't actually implement the proposed architecture
> - This approach is punishingly heavy on the database - O(n^2) or worse
> Yes, re-reading the state of all resources when ever run a new converge
> is worrying, but I think Michal had some ideas to minimize this.
> - A lot of phase 2 is mixed in with phase 1 here, making it
> difficult to evaluate which changes need to be made first and
> whether this approach works with existing plugins
> - The code is not really based on how Heat works at the moment, so
> there would be either a major redesign required or lots of radical
> changes in Heat or both
> I think there's a fair chance that given another 3-4 weeks to work
> on this, all of these issues and others could probably be resolved.
> The question for me at this point is not so much "if" but "why".
> Michał believes that this approach will make Phase 2 easier to
> implement, which is a valid reason to consider it. However, I'm not
> aware of any particular issues that my approach would cause in
> implementing phase 2 (note that I have barely looked into it at all
> though). In fact, I very much want Phase 2 to be entirely
> encapsulated by the Resource class, so that the plugin type (legacy
> vs. convergence-enabled) is transparent to the rest of the system.
> Only in this way can we be sure that we'll be able to maintain
> support for legacy plugins. So a phase 1 that mixes in aspects of
> phase 2 is actually a bad thing in my view.
> I really appreciate the effort that has gone into this already, but
> in the absence of specific problems with building phase 2 on top of
> another approach that are solved by this one, I'm ready to call this
> a distraction.
> In it's defence, I like the simplicity of it. The concepts and code are
> easy to understand - tho' part of this is doesn't implement all the
> stuff on your list yet.
> Anant & Friends' Proposal
> First off, I have found this very difficult to review properly since
> the code is not separate from the huge mass of Heat code and nor is
> the commit history in the form that patch submissions would take
> (but rather includes backtracking and iteration on the design). As a
> result, most of the information here has been gleaned from
> discussions about the code rather than direct review. I have
> repeatedly suggested that this proof of concept work should be done
> using the simulator framework instead, unfortunately so far to no avail.
> The last we heard on the mailing list about this, resource clean-up
> had not yet been implemented. That was a major concern because that
> is the more difficult half of the algorithm. Since then there have
> been a lot more commits, but it's not yet clear whether resource
> clean-up, update-during-update, replace-on-update and rollback have
> been implemented, though it is clear that at least some progress has
> been made on most or all of them. Perhaps someone can give us an update.
> AIUI this code also mixes phase 2 with phase 1, which is a concern.
> For me the highest priority for phase 1 is to be sure that it works
> with existing plugins. Not only because we need to continue to
> support them, but because converting all of our existing
> 'integration-y' unit tests to functional tests that operate in a
> distributed system is virtually impossible in the time frame we have
> available. So the existing test code needs to stick around, and the
> existing stack create/update/delete mechanisms need to remain in
> place until such time as we have equivalent functional test coverage
> to begin eliminating existing unit tests. (We'll also, of course,
> need to have unit tests for the individual elements of the new
> distributed workflow, functional tests to confirm that the
> distributed workflow works in principle as a whole - the scenarios
> from the simulator can help with _part_ of this - and, not least, an
> algorithm that is as similar as possible to the current one so that
> our existing tests remain at least somewhat representative and don't
> require too many major changes themselves.)
> Speaking of tests, I gathered that this branch included tests, but I
> don't know to what extent there are automated end-to-end functional
> tests of the algorithm?
> From what I can gather, the approach seems broadly similar to the
> one I eventually settled on also. The major difference appears to be
> in how we merge two or more streams of execution (i.e. when one
> resource depends on two or more others). In my approach, the
> dependencies are stored in the resources and each joining of streams
> creates a database row to track it, which is easily locked with
> contention on the lock extending only to those resources which are
> direct dependencies of the one waiting. In this approach, both the
> dependencies and the progress through the graph are stored in a
> database table, necessitating (a) reading of the entire table (as it
> relates to the current stack) on every resource operation, and (b)
> locking of the entire table (which is hard) when marking a resource
> operation complete.
> I chatted to Anant about this today and he mentioned that they had
> solved the locking problem by dispatching updates to a queue that is
> read by a single engine per stack.
> My approach also has the neat side-effects of pushing the data
> required to resolve get_resource and get_att (without having to
> reload the resources again and query them) as well as to update
> dependencies (e.g. because of a replacement or deletion) along with
> the flow of triggers. I don't know if anything similar is at work here.
> It's entirely possible that the best design might combine elements
> of both approaches.
> The same open questions I detailed under my proposal also apply to
> this one, if I understand correctly.
> I'm certain that I won't have represented everyone's work fairly
> here, so I encourage folks to dive in and correct any errors about
> theirs and ask any questions you might have about mine. (In case you
> have been living under a rock, note that I'll be out of the office
> for the rest of the week due to Thanksgiving so don't expect
> immediate replies.)
> I also think this would be a great time for the wider Heat community
> to dive in and start asking questions and suggesting ideas. We need
> to, ahem, converge on a shared understanding of the design so we can
> all get to work delivering it for Kilo.
> Agree, we need to get moving on this.
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> <mailto:OpenStack-dev at lists.openstack.org>
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
Thanks Zane for your e-mail Zane and summarizing everyone's work.
The design goals mentioned above looks more of performance goals and
constraints to me. I understand that it is unacceptable to have a poorly
performing engine and Resource plug-ins broken. Convergence spec clearly
mentions that the existing Resource plugins should not be changed.
IMHO, and my teams' HO, the design goals of convergence would be:
1. Stability: No transient failures, either in Openstack/external
services or resources themselves, should fail the stack. Therefore, we
need to have Observers to check for divergence and converge a resource
if needed, to bring back to stable state.
2. Resiliency: Heat engines should be able to take up tasks in case of
3. Backward compatibility: "We don't break the user space." No existing
stacks should break.
We started the PoC with these goals in mind, any performance
optimization would be a plus point for us. Note than I am neglecting the
performance goal, just that it should be next in the pipeline. The
questions we should ask ourselves is whether we are storing enough data
(state of stack) in DB to enable resiliency? Are we distributing the
load evenly to all Heat engines? Does our notification mechanism
provides us some form of guarantee or acknowledgement?
In retrospective, we had to struggle a lot to understand the existing
Heat engine. We couldn't have done justice by just creating another
project in GitHub and without any concrete understanding of existing
state-of-affairs. We are not at the same page with Heat core members, we
are novice and cores are experts.
I am glad that we experimented with the Heat engine directly. The
current Heat engine is not resilient and the messaging also lacks
reliability. We (my team and I guess cores also) understand that async
message passing would be the way to go as synchronous RPC calls simply
wouldn't scale. But with async message passing there has to be some
mechanism of ACKing back, which I think lacks in current infrastructure.
How could we provide stable user defined stack if the underlying Heat
core lacks it? Convergence is all about stable stacks. To make the
current Heat core stable we need to have, at the least:
1. Some mechanism to ACK back messages over AMQP. Or some other solid
mechanism of message passing.
2. Some mechanism for fault tolerance in Heat engine using external
tools/infrastructure like Celerey/Zookeeper. Without external
infrastructure/tool we will end-up bloating Heat engine with lot of
boiler-plate code to achieve this. We had recommended Celery in our
previous e-mail (from Vishnu.)
It was due to our experiments with Heat engines for this PoC, we could
come up with above recommendations.
Sate of our PoC
On GitHub: https://github.com/anantpatil/heat-convergence-poc
Our current implementation of PoC locks the stack after each
notification to mark the graph as traversed and produce next level of
resources for convergence. We are facing challenges in
removing/minimizing these locks. We also have two different school of
thoughts for solving this lock issue as mentioned above in Vishnu's
e-mail. I will these descibe in detail the Wiki. There would different
branches in our GitHub for these two approaches.
There are few concerns like huge mass of code and bulky changes. I am
currently addressing this issue. The README at GitHub for this PoC will
be helpful for you to set things up quickly and play around with a dummy
resource I have created, which also provides control on simulating
update-in-place. I am in the process of cleaning up all the not needed
code. Only the code needed to create/update/delete a stack will be there
along with the dummy resource. Most of the changes are in service.py,
stack.py and resource.py file. As Heat cores, I expect that you would be
in better position to review these. I will send out a separate e-mail
once I am done.
We learned in hard way that the graph table with True/False was not
enough for handling concurrent updates. We are currently testing it with
graph table having state rather than a flag. Resource versions are very
handy to rollback to and to handle concurrent updates. We have solved
the delete-after-update, update-on-update and rollback issues (not in
master branch though, I will publish the branch soon) and we are testing
it. Resource timeout will be needed and I agree with Vishnu that the
default value has to come from plug-in. Single resource provisioning
should not cause the entire stack to timeout. It was hard learning for
me, but we are on the verge of finishing this up.
I agree that there are no functional tests which you could run and see.
But we do have sample templates to test the PoC from outside. Our
changes are also mostly limited to the three files I mentioned above.
Nevertheless, I have asked my team to spend efforts in writing the tests
for our PoC and it would be ready soon. I do understand your concern
As I said, performance was our next goal, so we are addressing this
issue. The current stack loads all the resources from DB which can be
simply avoided and only those which need to be converged should be
loaded along with their children to realize the resource definition if
it has a get_attr dependency on children. Apart from this, only stack
lock seems to be a problem.
It was difficult, for me personally, to completely understand Zane's PoC
and how it would lay the foundation for aforementioned design goals. It
would be very helpful to have Zane's understanding here. I could
understand that there are ideas like async message passing and notifying
the parent which we also subscribe to. At lot of places we have common
approach, as far as I could understand.
I agree with Zane that there is lot of scope for us to collaborate. We
do need core team's help in making Heat stable and scalable.
Also, I have appealed my team members to be more active on IRC to
More information about the OpenStack-dev