Question: How can we persist the flow so that it can be resumed, restarted or rolled-back on engine failure?
Answer: Since a flow is a set of atoms and relations between atoms we need to create a model and corresponding information that allows us to persist the right amount of information to preserve, resume, and rollback a flow on software or hardware failure.
To allow for resumption TaskFlow must be able to re-create the flow and re-connect the links between atom (and between atoms->atom details and so on) in order to revert those atoms or resume those atoms in the correct ordering. TaskFlow provides a pattern that can help in automating this process (it does not prohibit the user from creating their own strategies for doing this).
The default provided way is to provide a factory function which will create (or recreate your workflow). This function can be provided when loading a flow and corresponding engine via the provided load_from_factory() method. This factory function is expected to be a function (or staticmethod) which is reimportable (aka has a well defined name that can be located by the __import__ function in python, this excludes lambda style functions and instance methods). The factory function name will be saved into the logbook and it will be imported and called to create the workflow objects (or recreate it if resumption happens). This allows for the flow to be recreated if and when that is needed (even on remote machines, as long as the reimportable name can be located).
When a flow is created it is expected that each atom has a unique name, this name serves a special purpose in the resumption process (as well as serving a useful purpose when running, allowing for atom identification in the notification process). The reason for having names is that an atom in a flow needs to be somehow matched with (a potentially) existing AtomDetail during engine resumption & subsequent running.
The match should be:
Names provide this although they do have weaknesses:
Note
Even though these weaknesses names were selected as a good enough solution for the above matching requirements (until something better is invented/created that can satisfy those same requirements).
When new flow is loaded into engine, there is no persisted data for it yet, so a corresponding FlowDetail object will be created, as well as a AtomDetail object for each atom that is contained in it. These will be immediately saved into the persistence backend that is configured. If no persistence backend is configured, then as expected nothing will be saved and the atoms and flow will be ran in a non-persistent manner.
Subsequent run: When we resume the flow from a persistent backend (for example, if the flow was interrupted and engine destroyed to save resources or if the service was restarted), we need to re-create the flow. For that, we will call the function that was saved on first-time loading that builds the flow for us (aka; the flow factory function described above) and the engine will run. The following scenarios explain some expected structural changes and how they can be accommodated (and what the effect will be when resuming & running).
When the factory function mentioned above returns the exact same the flow and atoms (no changes are performed).
Runtime change: Nothing should be done – the engine will re-associate atoms with AtomDetail objects by name and then the engine resumes.
When the factory function mentioned above alters the flow by adding a new atom in (for example for changing the runtime structure of what was previously ran in the first run).
Runtime change: By default when the engine resumes it will notice that a corresponding AtomDetail does not exist and one will be created and associated.
When the factory function mentioned above alters the flow by removing a new atom in (for example for changing the runtime structure of what was previously ran in the first run).
Runtime change: Nothing should be done – flow structure is reloaded from factory function, and removed atom is not in it – so, flow will be ran as if it was not there, and any results it returned if it was completed before will be ignored.
When the factory function mentioned above alters the flow by deciding that a newer version of a previously existing atom should be ran (possibly to perform some kind of upgrade or to fix a bug in a prior atoms code).
Factory change: The atom name & version will have to be altered. The factory should replace this name where it was being used previously.
Runtime change: This will fall under the same runtime adjustments that exist when a new atom is added. In the future TaskFlow could make this easier by providing a upgrade() function that can be used to give users the ability to upgrade atoms before running (manual introspection & modification of a LogBook can be done before engine loading and running to accomplish this in the meantime).
When the factory function mentioned above alters the flow by deciding that a previously existing atom should be split into N atoms or the factory function decides that N atoms should be merged in <N atoms (typically occurring during refactoring).
Runtime change: This will fall under the same runtime adjustments that exist when a new atom is added or removed. In the future TaskFlow could make this easier by providing a migrate() function that can be used to give users the ability to migrate atoms previous data before running (manual introspection & modification of a LogBook can be done before engine loading and running to accomplish this in the meantime).
If manual links were added or removed from graph, or task requirements were changed, or flow was refactored (atom moved into or out of subflows, linear flow was replaced with graph flow, tasks were reordered in linear flow, etc).
Runtime change: Nothing should be done.