Optimization

The objective of optimization is to remove as many tasks from the graph as possible, as efficiently as possible. This helps reduce costs and deliver results as quickly as possible. For example, if a push only modifies a single test, then ideally we only run the tasks necessary to run that test.

A task is said to be “optimized” when it is either replaced with an equivalent, already existing task, or dropped from the graph entirely.

Optimization Strategies

Each task has a single named optimization strategy, and can provide an argument to that strategy. Each strategy is defined as an OptimizationStrategy.

Each task has a task.optimization property describing the optimization strategy that applies, specified as a dictionary mapping strategy to argument. For example:

task.optimization = {'skip-unless-changed': ['js/**', 'tests/**']}

Strategy implementations are shared across all tasks, so they may cache commonly-used information as instance variables.

See Optimization Strategies for an overview of the strategies that are included with Taskgraph.

Optimization Process

Optimization proceeds in three phases: removing tasks, replacing tasks, and finally generating a subgraph containing only the remaining tasks.

Assume the following task graph as context for these examples:

flowchart LR R(decision) R --> A R --> B R --> C A --> D B --> D C --> H D --> G D --> H

Removing Tasks

This phase begins with tasks on which nothing depends (leaf nodes) and follows the dependency graph backward from there – right to left in the diagram above. If a task is not removed, then nothing it depends on will be removed either. Thus if G and H are both removed, C and D may be removed as well. We then evaluate the new leaf nodes, so if D is removed, then A and B may be removed as well. And so on until there are no tasks left to evaluate.

If G is removed but H is not removed, then the entire graph will be kept as all tasks (other than G) are in H’s dependency chain.

For each task with no remaining dependencies, the decision whether to remove is made by calling the optimization strategy’s should_remove_task method. If this method returns True, the task is removed.

The optimization process takes a do_not_optimize argument containing a list of tasks that cannot be removed under any circumstances. This can be used to “force” running specific tasks in various contexts.

Replacing Tasks

This phase begins with tasks without dependencies (other than the decision task) and follows the reversed dependency graph from there – left to right in the diagram above. If a task is not replaced, then anything depending on that task cannot be replaced. Replacement is generally done on the basis of some hash of the inputs to the task. In the diagram above, if both A and B are replaced with existing tasks, then D is a candidate for replacement. But if B has no replacement, then replacement of D will not be considered.

It is possible to replace a task with nothing. This is similar to optimizing away, but is useful for utility tasks (possibly like G). If such a task is considered for replacement, then all of its dependencies (here, D) have already been replaced and there is no utility in running the task and no need for a replacement task. It is an error for a task on which others depend to be replaced with nothing.

The do_not_optimize set applies to task replacement, as does an additional existing_tasks dictionary which allows the caller to supply a set of known, pre-existing tasks. This is used for action tasks, for example, where it contains the entire task-graph generated by the original decision task.

Subgraph Generation

The first two phases annotate each task in the existing taskgraph with their fate: removed, replaced, or kept. The tasks that are replaced also have a replacement taskId.

The last phase constructs a subgraph containing the retained tasks, and simultaneously rewrites all dependencies to refer to taskIds instead of labels. To do so, it assigns a taskId to each retained task and uses the replacement taskId for all replaced tasks.

The soft-dependencies are then solved for each task, by adding all the remaining tasks in the subgraph from that list to its dependencies.

The result is an optimized taskgraph with tasks named by taskId instead of label. At this phase, the edges in the task graph diverge from the task.dependencies attributes, as the latter may contain dependencies outside of the taskgraph (for replacement tasks).

As a side-effect, this phase also expands all {"task-reference": ".."} and {"artifact-reference": ".."} objects within the task definitions.