[openstack-dev] [Heat] Convergence proof-of-concept showdown

Zane Bitter zbitter at redhat.com
Wed Dec 10 17:47:13 UTC 2014

You really need to get a real email client with quoting support ;)

On 10/12/14 06:42, Murugan, Visnusaran wrote:
> Well, we still have to persist the dependencies of each version of a resource _somehow_, because otherwise we can't know how to clean them up in the correct order. But what I think you meant to say is that this approach doesn't require it to be persisted in a separate table where the rows are marked as traversed as we work through the graph.
> [Murugan, Visnusaran]
> In case of rollback where we have to cleanup earlier version of resources, we could get the order from old template. We'd prefer not to have a graph table.

In theory you could get it by keeping old templates around. But that 
means keeping a lot of templates, and it will be hard to keep track of 
when you want to delete them. It also means that when starting an update 
you'll need to load every existing previous version of the template in 
order to calculate the dependencies. It also leaves the dependencies in 
an ambiguous state when a resource fails, and although that can be 
worked around it will be a giant pain to implement.

I agree that I'd prefer not to have a graph table. After trying a couple 
of different things I decided to store the dependencies in the Resource 
table, where we can read or write them virtually for free because it 
turns out that we are always reading or updating the Resource itself at 
exactly the same time anyway.

>> This approach reduces DB queries by waiting for completion notification on a topic. The drawback I see is that delete stack stream will be huge as it will have the entire graph. We can always dump such data in ResourceLock.data Json and pass a simple flag "load_stream_from_db" to converge RPC call as a workaround for delete operation.
> This seems to be essentially equivalent to my 'SyncPoint' proposal[1], with the key difference that the data is stored in-memory in a Heat engine rather than the database.
> I suspect it's probably a mistake to move it in-memory for similar reasons to the argument Clint made against synchronising the marking off of dependencies in-memory. The database can handle that and the problem of making the DB robust against failures of a single machine has already been solved by someone else. If we do it in-memory we are just creating a single point of failure for not much gain. (I guess you could argue it doesn't matter, since if any Heat engine dies during the traversal then we'll have to kick off another one anyway, but it does limit our options if that changes in the future.)
> [Murugan, Visnusaran] Resource completes, removes itself from resource_lock and notifies engine. Engine will acquire parent lock and initiate parent only if all its children are satisfied (no child entry in resource_lock). This will come in place of Aggregator.

Yep, if you s/resource_lock/SyncPoint/ that's more or less exactly what 
I did. The three differences I can see are:

1) I think you are proposing to create all of the sync points at the 
start of the traversal, rather than on an as-needed basis. This is 
probably a good idea. I didn't consider it because of the way my 
prototype evolved, but there's now no reason I can see not to do this. 
If we could move the data to the Resource table itself then we could 
even get it for free from an efficiency point of view.
2) You're using a single list from which items are removed, rather than 
two lists (one static, and one to which items are added) that get 
compared. Assuming (1) then this is probably a good idea too.
3) You're suggesting to notify the engine unconditionally and let the 
engine decide if the list is empty. That's probably not a good idea - 
not only does it require extra reads, it introduces a race condition 
that you then have to solve (it can be solved, it's just more work). 
Since the update to remove a child from the list is atomic, it's best to 
just trigger the engine only if the list is now empty.

> It's not clear to me how the 'streams' differ in practical terms from just passing a serialisation of the Dependencies object, other than being incomprehensible to me ;). The current Dependencies implementation
> (1) is a very generic implementation of a DAG, (2) works and has plenty of unit tests, (3) has, with I think one exception, a pretty straightforward API, (4) has a very simple serialisation, returned by the edges() method, which can be passed back into the constructor to recreate it, and (5) has an API that is to some extent relied upon by resources, and so won't likely be removed outright in any event.
> Whatever code we need to handle dependencies ought to just build on this existing implementation.
> [Murugan, Visnusaran] Our thought was to reduce payload size (template/graph). Just planning for worst case scenario (million resource stack) We could always dump them in ResourceLock.data to be loaded by Worker.

If there's a smaller representation of a graph than a list of edges then 
I don't know what it is. The proposed stream structure certainly isn't 
it, unless you mean as an alternative to storing the entire graph once 
for each resource. A better alternative is to store it once centrally - 
in my current implementation it is passed down through the trigger 
messages, but since only one traversal can be in progress at a time it 
could just as easily be stored in the Stack table of the database at the 
slight cost of an extra write.

I'm not opposed to doing that, BTW. In fact, I'm really interested in 
your input on how that might help make recovery from failure more 
robust. I know Anant mentioned that not storing enough data to recover 
when a node dies was his big concern with my current approach.

I can see that by both creating all the sync points at the start of the 
traversal and storing the dependency graph in the database instead of 
letting it flow through the RPC messages, we would be able to resume a 
traversal where it left off, though I'm not sure what that buys us.

And I guess what you're suggesting is that by having an explicit lock 
with the engine ID specified, we can detect when a resource is stuck in 
IN_PROGRESS due to an engine going down? That's actually pretty interesting.

> Based on our call on Thursday, I think you're taking the idea of the Lock table too literally. The point of referring to locks is that we can use the same concepts as the Lock table relies on to do atomic updates on a particular row of the database, and we can use those atomic updates to prevent race conditions when implementing SyncPoints/Aggregators/whatever you want to call them. It's not that we'd actually use the Lock table itself, which implements a mutex and therefore offers only a much slower and more stateful way of doing what we want (lock mutex, change data, unlock mutex).
> [Murugan, Visnusaran] Are you suggesting something like a select-for-update in resource table itself without having  a lock table?

Yes, that's exactly what I was suggesting.


More information about the OpenStack-dev mailing list