Experiments v2 is out
Issue 02 · Aug 31, 2026
Experiments v2 lands this week
We teased this last issue, and it goes out this week. Experiments v2 is a rebuild from the assignment engine up, and the headline is that experiments are now part of your configuration instead of a copy of it. In v1 a group froze a whole copy, so later edits never reached a running test and only one experiment could touch a player at a time. Now a group stores only the paths it changes, so several tests can run at once as long as they don’t overlap.
- Target on your own data. Any property your game sends can be a condition, and player tenure is built in, so “first seen in the last 7 days” is a rule you can write.
- New ways to assign. Two to eight groups, per player or per server. Pick a deterministic weighted hash or a weighted round robin, and assignments are sticky by default.
- Break the glass. Put a specific player or server into a group by hand, pull someone out, or hand them back to automatic assignment. Every change is audited.
- Reading the result. Break any metric down by experiment or group, and filter to players in a test or out of it. A new results view built for v2 will come later.
You can keep several tests running at once without them stepping on each other, aim each one with your own data, and trust that a player who enrolls stays put for the length of the test.
Read the experiments docs, or reply and we’ll walk your team through it.
Ten days to the RDC mixer
- Thursday, September 10, 3:00 to 6:00 PM PT
- Courage Anyone, 399 S 1st St, Floor 2, San Jose
Twin Atlas is hosting a mixer with us on day one of RDC. Drinks, food, and a room of Roblox creators, community members, and studio teams. Come talk shop with our team or just unwind after the day’s sessions.
Registration is subject to host approval and the room only holds so many. If you’re going to be in San Jose, put your name in now.
Major heatmap upgrade
This one is live in your dashboard now. The heat can follow your zoom: turn on Scale to view and the colors are recomputed for the area you’re looking at, so a busy room no longer flattens every other room into the same shade.
The legend has two handles, and they do different jobs. Pull the low one up and everything quieter than that drops off the map, leaving only the areas that clear your bar. Pull the high one down and nothing disappears; the color range caps there, so a single hot spot stops eating the whole gradient and the moderate traffic behind it separates out. Size, opacity, and rotation are yours to set, and hovering reads its value from the same data the colors come from, so the number always agrees with what you’re looking at.
Queries that finish what they started
Most queries come back instantly, and that hasn’t changed. This is about the far end, where a question spans many billions of rows: those used to die at thirty seconds and retry themselves. Query execution has now been pulled off the request that asked for it, so a slow query keeps running instead of being cut off mid-flight, and the ceiling doubled to sixty seconds.
- You can see it working. Tiles tell you whether they’re queued or running, and show progress as the query goes, so a slow load looks like progress instead of a stuck spinner.
- One query, not one each. When your whole team opens the same board, they join the same execution instead of each kicking off separate work over the same data.
- No more pointless reruns. Boards stop re-running everything every ten seconds. Cached results appear immediately and refresh behind you, and a query nobody is watching any more gets canceled instead of finishing into the void.
The questions that used to time out are worth asking again, boards load when you open them, and it stops mattering how many of you are looking at once.
Straight from your feedback
Small things you told us were annoying, now fixed. Keep them coming, we read every one.
- Speed — Queries that filter or break down on the properties the SDK sends by default run noticeably faster
- Layout — The query builder behaves on small screens, and its pickers stop collapsing in short windows
What we’re working on next
community pulse for everyone
Pulse has been in front of a small group of studios, and their notes are what we’ve been working through: what belongs on the page, which signals actually get acted on, where it needs to be clearer. It shares a road with everything else in this issue, so it has moved in steps rather than in one jump. Worth the wait, and getting closer.
what players say, meet what they do
- Sentiment and behavior currently live in different rooms. We want to knock the wall down: chart sentiment for a topic over time next to your other metrics, and build audiences out of it.
- Take the players who keep saying leveling is too slow, and put exactly them into an experiment that speeds it up. Then check whether they stopped complaining.
- It runs the other way too. For any experiment, read how each group actually felt about the change, not just what the numbers moved.
What players say stops being a separate report you read after the fact, and becomes something you can segment on and measure against. We’re getting underway on this soon, and we’d love your feedback to keep shaping it and what comes after.
swap the map without touching the query
Right now, changing a heatmap’s background image means editing the query behind it. That’s backwards when you just want to look at the same data on a different map, so we’re moving it up to the dashboard.