Everything but attributes in the light dom is manually compared during
DOM traversal, for example child nodes or scroll offset.
This way we get a bullet-proof solution that works with input values,
scroll offsets, shadow dom and anything else web comes up with.
We also restore scroll only on the document scrolling element, for
performance reasons. We should figure out the story around scrolling.
Changes stationary snapshots from ~0.5ms to ~2.5ms due to DOM traversal.
- Intercept CSSOM modifications and recalculate overridden css text.
- When css text does not change, use "backwards reference" similar
to node references.
- Set 'Cache-Control: no-cache' for resources that could be overridden.
- Switch from html to json ml format.
- Allow node reuse between snapshots with `[nSnapshotsBefore, nodeWithIndexM]`.
- Service worker now lazily serializes snapshot chunks into a single html.
This decreases total snapshot size on random scripts ~10x.
This also decreases snapshot collecting time on mostly static pages to ~0.3ms.
Unfortunate downside for now is that we have to intercept
`Element.prototype.attachShadow` to invalidate nodes. This
also temporary breaks scroll restoration. Needs more research.
This introduces an http server that serves our frontend and our snapshots. There is more work to untangle the big server into a few modules.
This change allows us:
- Maybe eventually serve the trace viewer as a web page.
- Rely on browser caches for fast snapshot rendering. This PR also adds "snapshot on hover" feature, subject to change.
- Instead of capturing snapshots on demand, we now stream them
from each frame every 100ms.
- Certain actions can also force snapshots at particular moment using
"checkpoints".
- Trace viewer is able to show the page snapshot at a particular
timestamp, or using a "checkpoint" snapshot.
- Small optimization to not process stylesheets if CSSOM was not used.
There still is a lot of room for improvement.
