This adds variants of the SnapshotVerify*(...) methods that allow for imprecise comparisons, i.e. using the perPixelTolerance and overallTolerance parameters.
Why is this necessary?
Adding tolerances has been a highly requested feature (see #63) to work around some simulator changes introduced in iOS 13. Historically the simulator has supported CPU-based rendering, giving us very stable image representations of views that we can compare pixel-by-pixel. Unfortunately, with iOS 13, Apple changed the simulator to use exclusively GPU-based rendering, which means that the resulting snapshots may differ slightly across machines (see uber/ios-snapshot-test-case#109).
The negative effects of this were mitigated in iOSSnapshotTestCase by adding two tolerances to snapshot comparisons: a per-pixel tolerance that controls how close in color two pixels need to be to count as unchanged and an overall tolerance that controls what portion of pixels between two images need to be the same (based on the per-pixel calculation) for the images to be considered unchanged. Setting these tolerances to non-zero values enables engineers to record tests on one machine and run them on another (e.g. record new reference images on their laptop and then run tests on CI) without worrying about the tests failing due to differences in GPU rendering. This is great in theory, but from our testing we've found even the lowest tolerance values to consistently handle GPU differences between machine types let through a significant number of visual regressions. In other words, there is no magic tolerance threshold that avoids false negatives based on GPU rendering and also avoids false positives based on minor visual regressions.
This is especially true for accessibility snapshots. To start, tolerances seem to be more reliable when applied to relatively small snapshot images, but accessibility snapshots tend to be fairly large since they include both the view and the legend. Additionally, the text in the legend can change meaningfully and reflect only a small number of pixel changes. For example, I ran a test of full screen snapshot on an iPhone 12 Pro with two columns of legend. Even an overall tolerance of only 0.0001 (0.01%) was enough to let through a regression where one of the elements lost its .link trait (represented by the text "Link." appended to the element's description in the snapshot). But this low a tolerance wasn't enough to handle the GPU rendering differences between a MacBook Pro and a Mac Mini. This is a simplified example since it only uses overallTolerance, not perPixelTolerance, but we've found many similar situations arise even with the combination.
Some teams have developed infrastructure to allow snapshots to run on the same hardware consistently and have built a developer process around that infrastructure, but many others have accepted tolerances as a necessity today.
Why create separate "imprecise" variants?
The simplest approach to adding tolerances would be adding the perPixelTolerance and overallTolerance parameters to the existing snapshot methods, however I feel adding separate methods with an "imprecise" prefix is better in the long run. The naming is motivated by the idea that it needs to be very obvious when what you're doing might result in unexpected/undesirable behavior. In other words, when using one of the core snapshot methods, you should have extremely high confidence that a test passing means there's no regressions. When you use an "imprecise" variant, it's up to you to set your confidence levels according to your chosen tolerances. This is similar to the "unsafe" terminology around memory in the Swift API. You should generally feel very confident in the memory safety of your code, but any time you see "unsafe" it's a sign to be extra careful and not gather unwarranted confidence from the compiler.
Longer term, I'm hopeful we can find alternative comparison algorithms that allow for GPU rendering differences without opening the door to regressions. We can integrate these into the core snapshot methods as long as they do not introduce opportunities for regressions, or add additional comparison variants to iterate on different approaches.
This adds variants of the
SnapshotVerify*(...)methods that allow for imprecise comparisons, i.e. using theperPixelToleranceandoverallToleranceparameters.Why is this necessary?
Adding tolerances has been a highly requested feature (see #63) to work around some simulator changes introduced in iOS 13. Historically the simulator has supported CPU-based rendering, giving us very stable image representations of views that we can compare pixel-by-pixel. Unfortunately, with iOS 13, Apple changed the simulator to use exclusively GPU-based rendering, which means that the resulting snapshots may differ slightly across machines (see uber/ios-snapshot-test-case#109).
The negative effects of this were mitigated in iOSSnapshotTestCase by adding two tolerances to snapshot comparisons: a per-pixel tolerance that controls how close in color two pixels need to be to count as unchanged and an overall tolerance that controls what portion of pixels between two images need to be the same (based on the per-pixel calculation) for the images to be considered unchanged. Setting these tolerances to non-zero values enables engineers to record tests on one machine and run them on another (e.g. record new reference images on their laptop and then run tests on CI) without worrying about the tests failing due to differences in GPU rendering. This is great in theory, but from our testing we've found even the lowest tolerance values to consistently handle GPU differences between machine types let through a significant number of visual regressions. In other words, there is no magic tolerance threshold that avoids false negatives based on GPU rendering and also avoids false positives based on minor visual regressions.
This is especially true for accessibility snapshots. To start, tolerances seem to be more reliable when applied to relatively small snapshot images, but accessibility snapshots tend to be fairly large since they include both the view and the legend. Additionally, the text in the legend can change meaningfully and reflect only a small number of pixel changes. For example, I ran a test of full screen snapshot on an iPhone 12 Pro with two columns of legend. Even an overall tolerance of only
0.0001(0.01%) was enough to let through a regression where one of the elements lost its.linktrait (represented by the text "Link." appended to the element's description in the snapshot). But this low a tolerance wasn't enough to handle the GPU rendering differences between a MacBook Pro and a Mac Mini. This is a simplified example since it only usesoverallTolerance, notperPixelTolerance, but we've found many similar situations arise even with the combination.Some teams have developed infrastructure to allow snapshots to run on the same hardware consistently and have built a developer process around that infrastructure, but many others have accepted tolerances as a necessity today.
Why create separate "imprecise" variants?
The simplest approach to adding tolerances would be adding the
perPixelToleranceandoverallToleranceparameters to the existing snapshot methods, however I feel adding separate methods with an "imprecise" prefix is better in the long run. The naming is motivated by the idea that it needs to be very obvious when what you're doing might result in unexpected/undesirable behavior. In other words, when using one of the core snapshot methods, you should have extremely high confidence that a test passing means there's no regressions. When you use an "imprecise" variant, it's up to you to set your confidence levels according to your chosen tolerances. This is similar to the "unsafe" terminology around memory in the Swift API. You should generally feel very confident in the memory safety of your code, but any time you see "unsafe" it's a sign to be extra careful and not gather unwarranted confidence from the compiler.Longer term, I'm hopeful we can find alternative comparison algorithms that allow for GPU rendering differences without opening the door to regressions. We can integrate these into the core snapshot methods as long as they do not introduce opportunities for regressions, or add additional comparison variants to iterate on different approaches.