taskflow/doc/source/persistence.rst
Joshua Harlow c90e36020a Add the database schema to the sqlalchemy docs
In order to show people that are reading the docs what the current
schema for the database is create a helper tool that upgrades the
schema to the newest version and then uses tabulate to create a
restructured text version of that schema which is then included in
the documentation.

Change-Id: Id215f88430971a4a083f9739fb2ec59d971dc8fa
2014-10-10 06:24:39 +00:00

10 KiB

Persistence

Overview

In order to be able to receive inputs and create outputs from atoms (or other engine processes) in a fault-tolerant way, there is a need to be able to place what atoms output in some kind of location where it can be re-used by other atoms (or used for other purposes). To accommodate this type of usage TaskFlow provides an abstraction (provided by pluggable stevedore backends) that is similar in concept to a running programs memory.

This abstraction serves the following major purposes:

  • Tracking of what was done (introspection).
  • Saving memory which allows for restarting from the last saved state which is a critical feature to restart and resume workflows (checkpointing).
  • Associating additional metadata with atoms while running (without having those atoms need to save this data themselves). This makes it possible to add-on new metadata in the future without having to change the atoms themselves. For example the following can be saved:
    • Timing information (how long a task took to run).
    • User information (who the task ran as).
    • When a atom/workflow was ran (and why).
  • Saving historical data (failures, successes, intermediary results...) to allow for retry atoms to be able to decide if they should should continue vs. stop.
  • Something you create...

How it is used

On engine <engines> construction typically a backend (it can be optional) will be provided which satisfies the :py~taskflow.persistence.backends.base.Backend abstraction. Along with providing a backend object a :py~taskflow.persistence.logbook.FlowDetail object will also be created and provided (this object will contain the details about the flow to be ran) to the engine constructor (or associated :pyload() <taskflow.engines.helpers.load> helper functions). Typically a :py~taskflow.persistence.logbook.FlowDetail object is created from a :py~taskflow.persistence.logbook.LogBook object (the book object acts as a type of container for :py~taskflow.persistence.logbook.FlowDetail and :py~taskflow.persistence.logbook.AtomDetail objects).

Preparation: Once an engine starts to run it will create a :py~taskflow.storage.Storage object which will act as the engines interface to the underlying backend storage objects (it provides helper functions that are commonly used by the engine, avoiding repeating code when interacting with the provided :py~taskflow.persistence.logbook.FlowDetail and :py~taskflow.persistence.backends.base.Backend objects). As an engine initializes it will extract (or create) :py~taskflow.persistence.logbook.AtomDetail objects for each atom in the workflow the engine will be executing.

Execution: When an engine beings to execute (see engine <engines> for more of the details about how an engine goes about this process) it will examine any previously existing :py~taskflow.persistence.logbook.AtomDetail objects to see if they can be used for resuming; see resumption <resumption> for more details on this subject. For atoms which have not finished (or did not finish correctly from a previous run) they will begin executing only after any dependent inputs are ready. This is done by analyzing the execution graph and looking at predecessor :py~taskflow.persistence.logbook.AtomDetail outputs and states (which may have been persisted in a past run). This will result in either using there previous information or by running those predecessors and saving their output to the :py~taskflow.persistence.logbook.FlowDetail and :py~taskflow.persistence.backends.base.Backend objects. This execution, analysis and interaction with the storage objects continues (what is described here is a simplification of what really happens; which is quite a bit more complex) until the engine has finished running (at which point the engine will have succeeded or failed in its attempt to run the workflow).

Post-execution: Typically when an engine is done running the logbook would be discarded (to avoid creating a stockpile of useless data) and the backend storage would be told to delete any contents for a given execution. For certain use-cases though it may be advantageous to retain logbooks and there contents.

A few scenarios come to mind:

  • Post runtime failure analysis and triage (saving what failed and why).
  • Metrics (saving timing information associated with each atom and using it to perform offline performance analysis, which enables tuning tasks and/or isolating and fixing slow tasks).
  • Data mining logbooks to find trends (in failures for example).
  • Saving logbooks for further forensics analysis.
  • Exporting logbooks to hdfs (or other no-sql storage) and running some type of map-reduce jobs on them.

Note

It should be emphasized that logbook is the authoritative, and, preferably, the only (see inputs and outputs <inputs_and_outputs>) source of run-time state information (breaking this principle makes it hard/impossible to restart or resume in any type of automated fashion). When an atom returns a result, it should be written directly to a logbook. When atom or flow state changes in any way, logbook is first to know (see notifications <notifications> for how a user may also get notified of those same state changes). The logbook and a backend and associated storage helper class are responsible to store the actual data. These components used together specify the persistence mechanism (how data is saved and where -- memory, database, whatever...) and the persistence policy (when data is saved -- every time it changes or at some particular moments or simply never).

Usage

To select which persistence backend to use you should use the :pyfetch() <taskflow.persistence.backends.fetch> function which uses entrypoints (internally using stevedore) to fetch and configure your backend. This makes it simpler than accessing the backend data types directly and provides a common function from which a backend can be fetched.

Using this function to fetch a backend might look like:

from taskflow.persistence import backends

...
persistence = backends.fetch(conf={
    "connection': "mysql",
    "user": ...,
    "password": ...,
})
book = make_and_save_logbook(persistence)
...

As can be seen from above the conf parameter acts as a dictionary that is used to fetch and configure your backend. The restrictions on it are the following:

  • a dictionary (or dictionary like type), holding backend type with key 'connection' and possibly type-specific backend parameters as other keys.

Types

Memory

Connection: 'memory'

Retains all data in local memory (not persisted to reliable storage). Useful for scenarios where persistence is not required (and also in unit tests).

Files

Connection: 'dir' or 'file'

Retains all data in a directory & file based structure on local disk. Will be persisted locally in the case of system failure (allowing for resumption from the same local machine only). Useful for cases where a more reliable persistence is desired along with the simplicity of files and directories (a concept everyone is familiar with).

Sqlalchemy

Connection: 'mysql' or 'postgres' or 'sqlite'

Retains all data in a ACID compliant database using the sqlalchemy library for schemas, connections, and database interaction functionality. Useful when you need a higher level of durability than offered by the previous solutions. When using these connection types it is possible to resume a engine from a peer machine (this does not apply when using sqlite).

Schema

Logbooks

Name Type Primary Key
created_at DATETIME False
updated_at DATETIME False
uuid VARCHAR True
name VARCHAR False
meta TEXT False

Flow details

Name Type Primary Key
created_at DATETIME False
updated_at DATETIME False
uuid VARCHAR True
name VARCHAR False
meta TEXT False
state VARCHAR False
parent_uuid VARCHAR False

Atom details

Name Type Primary Key
created_at DATETIME False
updated_at DATETIME False
uuid VARCHAR True
name VARCHAR False
meta TEXT False
atom_type VARCHAR False
state VARCHAR False
intention VARCHAR False
results TEXT False
failure TEXT False
version TEXT False
parent_uuid VARCHAR False

Zookeeper

Connection: 'zookeeper'

Retains all data in a zookeeper backend (zookeeper exposes operations on files and directories, similar to the above 'dir' or 'file' connection types). Internally the kazoo library is used to interact with zookeeper to perform reliable, distributed and atomic operations on the contents of a logbook represented as znodes. Since zookeeper is also distributed it is also able to resume a engine from a peer machine (having similar functionality as the database connection types listed previously).

Interfaces

taskflow.persistence.backends

taskflow.persistence.backends.base

taskflow.persistence.logbook

Hierarchy

taskflow.persistence.backends.impl_memory taskflow.persistence.backends.impl_zookeeper taskflow.persistence.backends.impl_dir taskflow.persistence.backends.impl_sqlalchemy