4. Architecture

COMPUTE
computation

%3 graphtik-v4.1.0 flowchart cluster_compute compute operations operations compose compose operations->compose network network compose->network compile compile network->compile inputs input names inputs->compile outputs output names outputs->compile predicate node predicate predicate->compile plan execution plan compile->plan execute execute plan->execute solution solution execute->solution values input values values->execute The definition & execution of networked operation is split in 1+2 phases:

… it is constrained by these IO data-structures:

… populates these low-level data-structures:

… and utilizes these main classes:

graphtik.op.FunctionalOperation
graphtik.netop.NetworkOperation
graphtik.network.Network
graphtik.network.ExecutionPlan
graphtik.network.Solution
compose
COMPOSITION

The phase where operations are constructed and grouped into netops and corresponding networks.

Tip

compile
COMPILATION
The phase where the Network creates a new execution plan by pruning all graph nodes into a subgraph dag, and deriving the execution steps.
execute
EXECUTION
sequential

The phase where the ExecutionPlan calls the underlying functions of all operations contained in execution steps, with inputs/outputs taken from the solution.

Currently there are 2 ways to execute:

  • sequential
  • parallel, with a multiprocessing.ProcessPool

Plans may abort their execution by setting the abort run global flag.

parallel
parallel execution
execution pool
task

Execute operations in parallel, with a thread pool or process pool (instead of sequential). Operations and netop are marked as such on construction, or enabled globally from configurations.

Note that sideffects are not expected to function with process pools, certainly not when marshalling is enabled.

process pool

When the multiprocessing.Pool() class is used for parallel execution, the tasks must be communicated to/from the worker process, which requires pickling, and that may fail. With pickling failures you may try marshalling with dill library, and see if that helps.

Note that sideffects are not expected to function at all. certainly not when marshalling is enabled.

thread pool
When the multiprocessing.dummy.Pool() class for parallel execution, the tasks are run in process, so no marshalling is needed.
marshalling

Pickling parallel operations and their inputs/outputs using the dill module. It is configured either globally with set_marshal_tasks() or set with a flag on each operation / netop.

Note that sideffects do not work when this is enabled.

configurations

The functions controlling compile & execution globally are defined in config module; they underlying global data are stored in contextvars.ContextVar instances, to allow for nested control.

All boolean configuration flags are tri-state (None, False, True), allowing to “force” all operations, when they are not set to the None value. All of them default to None (false).

graph
network graph

The Network.graph (currently a DAG) contains all FunctionalOperation and _DataNode nodes of some netop.

They are layed out and connected by repeated calls of Network._append_operation() by Network constructor.

This graph is then pruned to extract the dag, and the execution steps are calculated, all ingredients for a new ExecutionPlan.

dag
execution dag
solution dag

There are 2 directed-acyclic-graphs instances used:

steps
execution steps

The ExecutionPlan.steps contains a list of the operation-nodes only from the dag, topologically sorted, and interspersed with instruction steps needed to compute the asked outputs from the given inputs.

It is built by Network._build_execution_steps() based on the subgraph dag.

The only instruction step is for performing evictions.

evictions
The _EvictInstruction steps erase items from solution as soon as they are not needed further down the dag, to reduce memory footprint while computing.
solution

A Solution instance created internally by NetworkOperation.compute() to hold the values both inputs & outputs, and the status of executed operations. It is based on a collections.ChainMap, to keep one dictionary for each operation executed +1 for inputs.

The results of the last operation executed “wins” in the final outputs produced, BUT while executing, the needs of each operation receive the solution values in reversed order, that is, the 1st operation result (or given input) wins for some needs name.

Rational:

During execution we want stability (the same input value used by all operations), and that is most important when consuming input values - otherwise, we would use (possibly overwritten and thus changing)) intermediate ones.

But at the end we want to affect the calculation results by adding operations into some netop - furthermore, it wouldn’t be very useful to get back the given inputs in case of overwrites.

overwrites
Values in the solution that have been written by more than one operations, accessed by Solution.overwrites:
net
network
the Network contains a graph of operations and can compile an execution plan or prune a cloned network for given inputs/outputs/node predicate.
plan
execution plan

Class ExecutionPlan perform the execution phase which contains the dag and the steps.

Compileed execution plans are cached in Network._cached_plans across runs with (inputs, outputs, predicate) as key.

inputs

The named input values that are fed into an operation (or netop) through Operation.compute() method according to its needs.

These values are either:

outputs

The dictionary of computed values returned by an operation (or a netop) matching its provides, when method Operation.compute() is called.

Those values are either:

  • retained in the solution, internally during execution, keyed by the respective provide, or
  • returned to user after the outer netop has finished computation.

When no specific outputs requested from a netop, NetworkOperation.compute() returns all intermediate inputs along with the outputs, that is, no evictions happens.

An operation may return partial outputs.

returns dictionary
When an operation is marked with this flag, the underlying function is not expected to return a sequence but a dictionary; hence, no “zipping” of outputs/provides takes place.
operation

Either the abstract notion of an action with specified needs and provides, or the concrete wrapper FunctionalOperation for arbitrary functions (any callable), that feeds on inputs and update outputs, from/to solution, or given-by/returned-to the user by a netop.

The distinction between needs/provides and inputs/outputs is akin to function parameters and arguments during define-time and run-time.

netop
network operation
The NetworkOperation class holding a network of operations.
needs

A list of (positionally ordered) names of the data needed by an operation to receive as inputs, roughly corresponding to the arguments of the underlying callable. The corresponding data-values will be extracted from solution (or given by the user) when Operation.compute() is called during execution.

Modifiers may annotate certain names as optionals, sideffects, or map them to differently named function arguments.

The graph is laid out by matching the needs & provides of all operations.

provides

A list of names to be zipped with the data-values produced when the operation’s underlying callable executes. The resulting outputs dictionary will be stored into the solution or returned to the user after Operation.compute() is called during execution.

Modifiers may annotate certain names as sideffects.

The graph is laid out by matching the needs & provides of all operations.

modifiers
Annotations on specific arguments of needs and/or provides such as optionals & sideffects (see graphtik.modifiers module).
optionals

Needs corresponding either:

  • to function arguments-with-defaults (annotated with optional), or
  • to *args (annotated with vararg & varargs),

that do not hinder execution of the operation if absent from inputs.

sideffects
Fictive needs or provides not consumed/produced by the underlying function of an operation, annotated with sideffect. A sideffect participates in the compilation of the graph, and is updated into the solution, but is never given/asked to/from functions.
prune
pruning

A subphase of compilation performed by method Network._prune_graph(), which extracts a subgraph dag that does not contain any unsatisfied operations.

It topologically sorts the graph, and prunes based on given inputs, asked outputs, node predicate and operation needs & provides.

unsatisfied operation

The core of pruning & rescheduling, performed by network._unsatisfied_operations() function, which collects all operations that fall into any of these 2 cases:

  • they have needs that do not correspond to any of the given inputs or the intermediately computed outputs of the solution;
  • all their provides are NOT needed by any other operation, nor are asked as outputs.
reschedule
rescheduling
partial outputs
partial operation
canceled operation

The partial pruning of the solution’s dag during execution. It happens when any of these 2 conditions apply:

  • an operation is marked with the FunctionalOperation.rescheduled attribute, which means that its underlying callable may produce only a subset of its provides (partial outputs);
  • endurance is enabled, either globally (in the configurations), or for a specific operation.

the solution must then reschedule the remaining operations downstream, and possibly cancel some of those ( assigned in Solution.canceled).

endurance

Keep executing as many operations as possible, even if some of them fail. Endurance for an operation is enabled if set_endure_operations() is true globally in the configurations or if FunctionalOperation.endurance is true.

You may interrogate Solution.executed to discover the status of each executed operations or call scream_if_incomplete().

predicate
node predicate
A callable(op, node-data) that should return true for nodes to be included in graph during compilation.
abort run

A global configurations flag that when set with abort_run() function, it halts the execution of all currently or future plans.

It is reset automatically on every call of NetworkOperation.compute() (after a successful intermediate compilation), or manually, by calling reset_abort().