Valve needs to rethink Fossilize

I think the amount of data that is being replayed is really excessive and unnecessary. I've been testing out Deadlock lately and the amount of shader pipelines to be compiled is really staggering. I tried to get through it, but it was just to much. I thought of checking the background compile and it felt like it was churning throught these for hours without end in sight. That behaviour is not exclusive to Deadlock, but many popular games have really huge amounts of replayable fragments.

Even after going through it, receiving more fragments with updates, it still takes way to long to validate everything, when you ultimately have almost 700k shader fragments. Most of which are probably duplicates, that only get catched when merging shader binaries.
[2024-09-07 21:48:58] Still replaying 1422450 (99%, 677994/691375).

I believe if it takes you more than 5 minutes to compile most of shader binaries on a decent CPU, then there's something not right with the implementation. Thats usually how PSOs work in many new titles and I think it's a good idea to emulate the behaviour of it. Yes, I know you can skip at any time, but some people will sit down for 30 minutes or even hours on lower end hardware (like this issue https://github.com/ValveSoftware/Fossilize/issues/251) thinking that it will make a difference, when in reality there probably wouldn't be any at this point.

I think Valve needs to rethink how the system works. In my opinion, I believe there needs to be some sort of relevancy system.

• It needs to serve users the most relevant or commonly recorded shaders.
• It needs to perform cleanup of old shader fragments that haven't been used in a while, but only when in the mean time commonly received other fragments. That would make sure we're not discarding fragments just because a game is not popular enough.

I think those options would greatly reduce compile time. Another good idea is to set a timer for 3-5 minutes and automatically kill the replaying job. With an option for the user to allow compilation of all fragments. Then you don't have to care how long it takes to complete. If the relevancy system works, then I believe after 5 minutes, they should be most of the way there anyway. Maybe it would be beneficial to collect stats for how long it takes users to replay shader fragments and perform cleanup based on that.

I personally believe Fossilize is a great system, but most users either:

• skip compilation early on because of how long it took for them before
• disable the feature altogether as a result
• will wait for 30 minutes or an hour+ thinking it has to be done to provide a stutter-less and meaningful difference to the experience.

I'm open to suggestions, but I believe Valve needs to do something for non-Deck users in this regard. I know Valve ships binaries to Deck users and that will not be possible for all configurations. Replay system is great, but it needs improvement to show users it's worth doing.

No files, tests, or entry points are named. Start by identifying the shader-fragment replay and cleanup paths, then measure the reported replay workload, including the 691375-fragment example. Done would require an agreed design and implementation scope for relevance, cleanup, or replay limits, followed by validation that compilation time and behavior improve.