x
/
solar
Code Issues Proposed changes

Docs: commit log

This commit is contained in:
Przemyslaw Kaminski 2015-05-22 09:45:44 +02:00
parent 51674c00ad
commit 558b6684ae
1 changed files with 47 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
docs/commit-log.md Normal file
View File

@ -0,0 +1,47 @@
# Commit log analysis
See here https://files.slack.com/files-pri/T03ACD12T-F04V4QC6E/2015-05-21_16.14.50.jpg for details.
We have some data storage (to decide -- one global storage or separate storage for each resource?
One global commit log or separate for each resource?) -- call it DB for simplicity.
User modifies some data of some resources K1, K2, H. This data is not stored immediately in the DB,
instead it is stored in some separate place and queued for execution (we call this 'Staged Log').
The modified data for a resource is represented as a diff in its inputs. So if user adds new resource
and assigns an IP to it, it is represented something like:
```
ip:
from: None
to: 10.20.0.2
```
User commands 'apply'. Orchestrator takes the modified resources and applies appropriate actions
in appropriate order that it computes.
We think that the 'appropriate action' can be inferred from the diff for each resource. So for example
if resource is new and has added IP the action `run` can be inferred because previous state was
`None` and current is something new. If on the other hand previous state was some value `A` and
new state is some value `B` -- the orchestrator decides that the action to be run is `update`. And
if previous state is some `A` and new state is `None` the action will be `remove`.
The 'appropriate order' taken by orchestrator can be just like the data flow graph initially. We
see possibility of optimizing the number of actions taken by orchestrator so that moving Keystone
service to another node can be simplified from 4 actions taken to 3 actions.
After resource action is finished the new state is saved to the commit log and data is updated in
the DB.
We want to support rollbacks via commit log. Rollback is done by replaying the commit log backwards.
In case of separate commit logs per resource we think rollback could be done like this: some resource
`K` is rolled back by one commit log, the diff action is the same as reversed diff action of the
commit we are rolling back. We can update other resources with this new data by analyzing the connections.
So in other words -- we change the data in one resource according to what is in the commit to be rolled
back and then we trigger changes in other connected resources. Then we run orchestrator actions like
described above.
From analysis of resource removal we think that we need to save connection data in each commit --
otherwise when we rollback that resource removal we wouldn't know how to restore its connections
to other resources.