[openstack-dev] [Heat] Convergence proof-of-concept showdown
Zane Bitter
zbitter at redhat.com
Thu Nov 27 02:20:49 UTC 2014
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
===============
https://github.com/zaneb/heat-convergence-prototype/tree/distributed-graph
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
resources
- 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?
Michał's Proposal
=================
https://github.com/inc0/heat-convergence-prototype/tree/iterative
Note that a version modified by me to use the same test scenario format
(but not the same scenarios) is here:
https://github.com/zaneb/heat-convergence-prototype/tree/iterative-adapted
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.
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
- 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.
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.
cheers,
Zane.
More information about the OpenStack-dev
mailing list