In the first post of this series I introduced the Komposition screencast editor, and briefly explained the fundamentals of property-based testing (PBT). Furthermore, I covered how to write testable code, regardless of how you check your code with automated tests. Lastly, I highlighted some difficulties in using properties to perform component and integration testing.

If you haven’t read the introductory post, I suggest doing so before continuing with this one. You’ll need an understanding of what PBT is for this case study to make sense.

This post is the first case study in the series, covering the timeline flattening process in Komposition and how it’s tested using PBT. The property tests aren’t integration-level tests, but rather unit tests. This case study serves as a warm-up to the coming, more advanced, ones.

Before we look at the tests, we need to learn more about Komposition’s hierarchical timeline and how the flattening process works.

The Hierarchical Timeline

Komposition’s timeline is hierarchical. While many non-linear editing systems have support for some form of nesting1 they are primarily focused on flat timeline workflows. The timeline structure and the keyboard-driven editing in Komposition is optimized for the screencast editing workflow I use.

It’s worth emphasizing that Komposition is not a general video editor. In addition to its specific editing workflow, you may need to adjust your recording workflow to use it effectively2.

Video and Audio in Parallels

At the lowest level of the timeline are clips and gaps. Those are put within the video and audio tracks of parallels. The following diagram shows a parallel consisting of two video clips and one audio clip.

Clips and gaps are placed in video and audio tracks
Clips and gaps are placed in video and audio tracks

The tracks of a parallel are played simultaneously (in parallel), as indicated by the arrows in the above diagram. The tracks start playing at the same time. This makes parallels useful to synchronize the playback of specific parts of a screencast, and to group closely related clips.

Gaps

When editing screencasts made up of separate video and audio recordings you often end up with differing clip duration. The voice-over audio clip might be longer than the corresponding video clip, or vice versa. A useful default behaviour is to extend the short clips. For audio, this is easy. Just pad with silence. For video, it’s not so clear what to do. In Komposition, shorter video tracks are padded with repeated still frame sections called gaps.

The following diagram shows a parallel with a short video clip and a longer audio clip. The dashed area represents the implicit gap.

Still frames are automatically inserted at implicit gaps to match track duration
Still frames are automatically inserted at implicit gaps to match track duration

You can also add gaps manually, specifying a duration of the gap and inserting it into a video or audio track. The following diagram shows a parallel with manually added gaps in both video and audio tracks.

Adding explicit gaps manually
Adding explicit gaps manually

Manually added gaps (called explicit gaps) are padded with still frames or silence, just as implicit gaps that are added automatically to match track duration.

Sequences

Parallels are put in sequences. The parallels within a sequence are played sequentially; the first one is played in its entirety, then the next one, and so on. This behaviour is different from how parallels play their tracks. Parallels and sequences, with their different playback behaviors, make up the fundamental building blocks of the compositional editing in Komposition.

The following diagram shows a sequence of two parallels, playing sequentially:

A sequence containing two parallels
A sequence containing two parallels

The Timeline

Finally, at the top level, we have the timeline. Effectively, the timeline is a sequence of sequences; it plays every child sequence in sequence. The reason for this level to exist is for the ability to group larger chunks of a screencast within separate sequences.

A timeline containing two sequences, with two parallels
each
A timeline containing two sequences, with two parallels each

I use separate sequences within the timeline to delimit distinct parts of a screencast, such as the introduction, the different chapters, and the summary.

Timeline Flattening

Komposition currently uses FFmpeg to render the final media. This is done by constructing an ffmpeg command invocation with a filter graph describing how to fit together all clips, still frames, and silent audio parts.

FFmpeg doesn’t know about hierarchical timelines; it only cares about video and audio streams. To convert the hierarchical timeline into a suitable representation to build the FFmpeg filter graph from, Komposition performs timeline flattening.

The flat representation of a timeline contains only two tracks; audio and video. All gaps are explicitly represented in those tracks. The following graph shows how a hierarchical timeline is flattened into two tracks.

Timeline flattening transforming a hierarchical timeline
Timeline flattening transforming a hierarchical timeline

Notice in the graphic above how the implicit gaps at the ends of video and audio tracks get represented with explicit gaps in the flat timeline. This is because FFmpeg does not know how to render implicit gaps. All gaps are represented explicitly, and are converted to clips of still frames or silent audio when rendered with FFmpeg.

Property Tests

To test the timeline flattening, there’s a number of properties that are checked. I’ll go through each one and their property test code.

These properties were primarily written after I already had an implementation. They capture some general properties of flattening that I’ve come up with. In other cases, I’ve written properties before beginning on an implementation, or to uncover an existing bug that I’ve observed.

Thinking about your system’s general behaviour and expressing that as executable property tests is hard. I believe, like with any other skill, that it requires a lot of practice. Finding general patterns for properties, like the ones Scott Wlaschin describe in Choosing properties for property-based testing, is a great place to start. When you struggle with finding properties of your system under test, try applying these patterns and see which work for you.

Property: Duration Equality

Given a timeline \(t\), where all parallels have at least one video clip, the total duration of the flattened \(t\) must be equal to the total duration of \(t\). Or, in a more dense notation,

\[\forall t \in T \to duration(flatten(t)) = duration(t)\]

where \(T\) is the set of timelines with at least one video clip in each parallel.

The reason that all parallels must have at least one video clip is because currently the flattening algorithm can only locate still frames for video gaps from within the same parallel. If it encounters a parallel with no video clips, the timeline flattening fails. This limitation is discussed in greater detail at the end of this article.

The test for the duration equality property is written using Hedgehog, and looks like this:

It generates a timeline using forAll and custom generators (1). Instead of generating timelines of any shape and filtering out only the ones with video clips in each parallel, which would be very inefficient, this test uses a custom generator to only obtain inputs that satisfy the invariants of the system under test.

The range passed as the first argument to Gen.timeline is used as the bounds of the generator, such that each level in the generated hierarchical timeline will have at most 5 children.

Gen.timeline takes as its second argument another generator, the one used to generate parallels, which in this case is Gen.parallelWithClips. With Hedgehog generators being regular values, it’s practical to compose them like this. A “higher-order generator” can be a regular function taking other generators as arguments.

As you might have noticed in the assertion (3), durationOf takes as its first argument a value AdjustedDuration. What’s that about? Komposition supports adjusting the playback speed of video media for individual clips. To calculate the final duration of a clip, the playback speed needs to taken into account. By passing AdjustedDuration we take playback speed into account for all video clips.

Sidetrack: Finding a Bug

Let’s say I had introduced a bug in timeline flattening, in which all video gaps weren’t added correctly to the flat video tracks. The flattening is implemented as a fold, and it would not be unthinkable that the accumulator was incorrectly constructed in a case. The test would catch this quickly and present us with a minimal counter-example:

Hedgehog presenting a minimal counter-example
Hedgehog presenting a minimal counter-example

Hedgehog prints the source code for the failing property. Below the forAll line the generated value is printed. The difference between the expected and actual value is printed below the failing assertion. In this case it’s a simple expression of type Duration. In case you’re comparing large tree-like structures, this diff will highlight only the differing expressions. Finally, it prints the following:

This failure can be reproduced by running:
> recheck (Size 23) (Seed 16495576598183007788 5619008431246301857) <property>

When working on finding and fixing the fold bug, we can use the printed size and seed values to deterministically rerun the test with the exact same inputs.

Property: Clip Occurence

Slightly more complicated than the duration equality property, the clip occurrence property checks that all clips from the hierarchical timeline, and no other clips, occur within the flat timeline. As discussed in the introduction on timeline flattening, implicit gaps get converted to explicit gaps and thereby add more gaps, but no video or audio clips should be added or removed.

The hierarchical timeline is generated and flattened like before (1, 2). The two assertions check that the respective video clips (3) and audio clips (4) are equal. It’s using lenses to extract clips from the flat timeline, and the helper functions timelineVideoClips and timelineAudioClips to extract clips from the original hierarchical timeline.

Still Frames Used

In the process of flattening, the still frame source for each gap is selected. It doesn’t assign the actual pixel data to the gap, but a value describing which asset the still frame should be extracted from, and whether to pick the first or the last frame (known as still frame mode.) This representation lets the flattening algorithm remain a pure function, and thus easier to test. Another processing step runs the effectful action that extracts still frames from video files on disk.

The decision of still frame mode and source is made by the flattening algorithm based on the parallel in which each gap occur, and what video clips are present before or after. It favors using clips occurring after the gap. It only uses frames from clips before the gap in case there are no clips following it. To test this behaviour, I’ve defined three properties.

Property: Single Initial Video Clip

The following property checks that an initial single video clip, followed by one or more gaps, is used as the still frame source for those gaps.

The custom video track generator (1) always produces tracks with an initial video clip followed by one or more video gaps. The generated timeline (2) can contain parallels with any audio track shape, which may result in a longer audio track and thus an implicit gap at the end of the video track. In either case, all video gaps should padded with the last frame of the initial video clip, which is checked in the assertion (4).

Still frames being sourced from the single initial video clip
Still frames being sourced from the single initial video clip

Property: Ending with a Video Clip

In case the video track ends with a video clip, and is longer than the audio track, all video gaps within the track should use the first frame of a following clip.

The custom generator (1) produces parallels where the video track is guaranteed to end with a clip, and where the audio track is 100 ms shorter than the video track. This ensures that there’s no implicit video gap at the end of the video track. Generating (2) and flattening (3) is otherwise the same as before. The assertion (4) checks that all video gaps uses the first frame of a following clip.

Still frames being sourced from following video clips when possible
Still frames being sourced from following video clips when possible

Property: Ending with an Implicit Video Gap

The last property on still frame usage covers the case where the video track is shorter than the audio track. This leaves an implicit gap which, just like explicit gaps inserted by the user, are padded with still frames.

The custom generator (1) generates a video track consisting of video clips only, and an audio track that is 100ms longer. Generating the timeline (2) and flattening (3) are again similar to the previous property tests. The assertion (4) checks that all video gaps use the last frame of preceding clips, even if we know that there should only be one at the end.

Still frames being sourced from preceding video clip for last implicit gap
Still frames being sourced from preceding video clip for last implicit gap

Properties: Flattening Equivalences

The last property I want to show in this case study checks flattening at the sequence and parallel levels. While rendering a full project always flattens at the timeline, the preview feature in Komposition can be used to render and preview a single sequence or parallel.

There should be no difference between flattening an entire timeline and flattening all of its sequences or parallels and folding those results into a single flat timeline. This is what the flattening equivalences properties are about.

The first property generates a timeline (1) where all parallels have at least one video clip. It flattens all sequences within the timeline and folds the results together (2). Folding flat timelines together means concatenating their video and audio tracks, resulting in a single flat timeline.

Before the final assertion, it checks that we got a result (3) and not Nothing. As it’s using the Gen.parallelWithClips generator there should always be video clips in each parallel, and we should always successfully flatten and get a result. The final assertion (4) checks that rendering the original timeline gives the same result as the folded-together results of rendering each sequence.

The other property is very similar, but operates on parallels rather than sequences:

The only difference is in the traversal (2), where we apply Render.flattenParallel to each parallel instead of applying Render.flattenSequence to each sequence.

Missing Properties

Whew! That was quite a lot of properties and code, especially for a warm-up. But timeline flattening could be tested more thoroughly! I haven’t yet written the following properties, but I’m hoping to find some time to add them:

  • Clip playback timestamps are the same. The “clip occurrence” property only checks that the hierarchical timeline’s clips occur in the flat timeline. It doesn’t check when in the flat timeline they occur. One way to test this would be to first annotate each clip in original timeline with its playback timestamp, and transfer this information through to the flat timeline. Then the timestamps could be included in the assertion.

  • Source assets used as still frame sources. The “still frames used” properties only check the still frame mode of gaps, not the still frame sources. The algorithm could have a bug where it always uses the first video clip’s asset as a frame source, and the current property tests would not catch it.

  • Same flat result is produced regardless of sequence grouping. Sequences can be split or joined in any way without affecting the final rendered media. They are merely ways of organizing parallels in logical groups. A property could check that however you split or join sequences within a timeline, the flattened result is the same.

A Missing Feature

As pointed out earlier, parallels must have at least one video clip. The flattening algorithm can only locate still frame sources for video gaps from within the same parallel. This is an annoying limitation when working with Komposition, and the algorithm should be improved.

As the existing set of properties describe timeline flattening fairly well, changing the algorithm could be done with a TDD-like workflow:

  1. Modify the property tests to capture the intended behaviour
  2. Tests will fail, with the errors showing how the existing implementation fails to find still frame sources as expected
  3. Change the implementation to make the tests pass

PBT is not only an after-the-fact testing technique. It can be used much like conventional example-based testing to drive development.

Obligatory Cliff-Hanger

In this post we’ve looked at timeline flattening, the simplest case study in the “Property-Based Testing in a Screencast Editor” series. The system under test was a module of pure functions, with complex enough behaviour to showcase PBT as a valuable tool. The tests are more closely related to the design choices and concrete representations of the implementation.

Coming case studies will dive deeper into the more complex subsystems of Komposition, and finally we’ll see how PBT can be used for integration testing. At that level, the property tests are less tied to the implementation, and focus on describing the higher-level outcomes of the interaction between subsystems.

Next up is property tests for the video classifier. It’s also implemented a pure function, but with slightly more complicated logic that is trickier to test. We’re going to look at an interesting technique where we generate the expected output instead of the input.

  1. Introduction
  2. Timeline Flattening
  3. Video Scene Classification
  4. Integration Testing

Thanks for reading! See you next time.

Credits

Thank you Chris Ford and Ulrik Sandberg for proof-reading and giving valuable feedback on drafts of this post.

Footnotes


  1. Final Cut Pro has compound clips, and Adobe Premiere Pro has nested sequences.

  2. The section on workflow in Komposition’s documentation describes how to plan, record, and edit your screencast in way compatible with Komposition.