We will be shipping WebRender for AMD Desktop with 68. We would like to have a gradual rollout again (like we did in 67)

Aiming for full rollout in 1 week with this plan in mind:
Feature is off by default at mainline release go-live/launch.
Turn on feature at 25% update Monday post go-live.
Push to 50% 3 days later, Thursday
Push to 100% a week later IF there are no blocking issues.

An important part of this is that we want to have Nvidia at 100% at the initial release and only roll out to AMD gradually.

Jeff, in the default state (ie, if Normandy doesn't do anything) will Nvidia users already have the feature turned on? Or will Normandy need to turn this on for 100% of Nvidia users as well as the rollout to AMD?

Do you know how we can identify AMD users based on Telemetry?

My very fuzzy understanding is that if Normandy doesn't do anything Nvidia users will have the feature turned off, so Normandy will need to turn it on for 100% of Nvidia users as well as the rollout ot AMD.

I believe something like normandy.telemetry.main.environment.system.gfx.adapters[.vendorID == '0x1002'][0] should be the appropriate target for AMD users.

Note that there is an experiment involving this feature as well. Because of that, this feature won't increase to 100% for the duration of that experiment. The experiment is slated to end August 13.

Sounds like we need some clarification here. My understanding is that if normandy doesn't do anything then WR will be enabled on qualified hardware (Nvidia and AMD). That was turned off in 67 by bug 1541489. I'm going to need to know today if we need to do the same in 68, and then have normandy enable it.

I believe Julien is correct in that it will be enabled by default for both Nvidia and AMD in 68, because the all.qualified pref is true on 68. If we can target just AMD with Normandy (and it sounds like we can) then perhaps the simplest option is to use Normandy to turn it OFF for AMD users at initial release and gradually turn it back on. That way we don't mess with nvidia users at all.

• Whitelisted AMD cards
• AMD-only bugs: bug 1539902
• List of bugs marked as affecting 68/Win10

if normandy doesn't do anything then WR will be enabled on qualified hardware (Nvidia and AMD)

If that's the case, then the experiment I linked in comment 4 is at serious risk, since it changes from a very easy type of experiment to a very hard type of experiment. For the harder type of experiments (called a split-recipe holdback study), we usually want at least a week of lead time before the target release goes live. As it stands, we only have 2 business days before the release (in the US).

use Normandy to turn it OFF for AMD users at initial release and gradually turn it back on

This is difficult and unreliable to do with Normandy. The only realistic way to do this again requires at least a week of lead up time before the release to be reliable. Given the way that WebRender works, the most common case we'd see here is that a user gets WebRender turned on when they upgraded to 68, has WebRender on for "a while", and then if they are an AMD user it gets shut back off some time later. This would be a pretty bad user experience.

If WebRender ships default-on in 68, I don't think this rollout or the related experiment are viable.

After some discussion we decided the best way forward is to not do a gradual rollout at all on 68. AMD users will just all get WR on with 68. We decided this is acceptable because:

• there were no reported issues from nvidia users in 67
• the AMD population is smaller than the nvidia users we already have
• we had very few bugs reported against AMD in pre-release population (IIRC only one bug actually)

The other alternative would be to uplift a patch to disable gfx.webrender.all.qualified in 68 and then using normandy to turn it back for all nvidia users and do a gradual rollout for AMD users. That seems to have higher risk because we might end up with some nvidia users who had it on 67 missing the new normandy study and getting it turned back off in 68. Also the 68 final build is happening imminently and rushing a patch in last minute isn't ideal.

So we're ditching the gradual rollout, and we will still consider doing an experiment on AMD users. The experiment will not be a "holdback" type experiment; instead some AMD users will get the qualified pref flipped back to false at some point after release, so they'll go back to non-WR for the purposes of being a control group. But we can discuss that in the experiment bugs.