taskcluster/docs/optimization-process.rst
author Jared Wein <jwein@mozilla.com>
Thu, 31 Jan 2019 23:40:33 +0000
changeset 512997 e4d259ea825ac517f8d2de85430fbd955be2df92
parent 509158 5a7bea3fb23b1c208725be620aa881645af40f8a
child 520117 5caf48a420eb337a296c8b337332f2933884f467
permissions -rw-r--r--
Bug 1521599 - Delete the failures data after it has been output. r=Gijs a=lizzard The failures were remaining in the data object, which later had any remaining keys printed in the diagnostics section. This bug was introduced because we stopped using Array objects to generate formatted strings. In the previous code, this would have ended up just printing out the first failure listed in the diagnostics section (a partial duplicate of the actual Failure Log). Differential Revision: https://phabricator.services.mozilla.com/D18289

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::

    TC1 <--\     ,- UP1
          , B1 <--- T1a
    I1 <-|       `- T1b
          ` B2 <--- T2a
    TC2 <--/     |- T2b
                 `- UP2

Removing Tasks
--------------

This phase begins with tasks on which nothing depends 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 T1a and T1b are both removed, B1 may be removed as well. But if T2b is
not removed, then B2 may not be removed either.

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 is used to
"force" running specific tasks.

Replacing Tasks
---------------

This phase begins with tasks having no dependencies 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 TC1 and I1 are replaced with existing
tasks, then B1 is a candidate for replacement. But if TC2 has no replacement,
then replacement of B2 will not be considered.

It is possible to replace a task with nothing.  This is similar to optimzing
away, but is useful for utility tasks like UP1. If such a task is considered
for replacement, then all of its dependencies (here, B1) 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 as 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 retained. 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 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.