Scenario-1 submission is ~234 MB; the largest field is unread by the leaderboard

[thc1006]

What I measured

Running the example agent through a real evaluation and packing the submission, the payload sizes come out as:

scenario	submission (JSON)
0 (10 static balloons)	~12 MB
1 (100 balloons, Monte Carlo)	~234 MB

For scenario 1, the breakdown by field is:

field	size	read by the leaderboard?
balloon_world_data.balloon_flights	165.7 MB (71%)	no
balloon_world_data.trajectories	64.4 MB	yes (the only field it reads)
balloon_world_data.rocket_flight	4.2 MB	no
everything else	~0 MB	mostly no

Why it is worth a look

The upload endpoint reads the whole file into memory and json.loads it, so a 234 MB submission turns into a few GB of Python objects per request. Any upload size limit has to sit above ~250 MB or it will reject real scenario-1 submissions, and with ~100 teams the storage adds up quickly.

The largest field, balloon_flights, is never read by the leaderboard. It is also the same data the viewer already gets from trajectories (balloon_flights is per-balloon, trajectories[...].balloon_states is per-step, but it is the same balloon positions and velocities), and it is fully determined by the run's seed plus scenario_parameters, so it can be recomputed rather than shipped.

Options (all measured)

• Drop balloon_flights from the submission. Producer-side only, the leaderboard needs no change. 234 MB → ~68 MB. The one thing to confirm first: that no offline re-verification / anti-cheat step relies on the shipped array rather than recomputing it.
• gzip the submission (.json.gz). 234 MB → ~83 MB, which is smaller than the current pickle. This one does need the leaderboard to gunzip before parsing, so it touches both sides.
• Both together land around ~24 MB.

Context

This came up while moving the submission format off pickle to close the pre-auth deserialization issue on the leaderboard (ARRC-Rocket/BalloonPoppingLeaderboard#7). That format change is a separate, size-neutral PR; I did not fold any trimming into it. I wanted to surface the size question on its own so we can decide the direction together. Happy to send the balloon_flights drop as a small producer-side PR if that is the way we want to go.

[thc1006]

Fixed by #74, which is on develop. These merged into develop rather than the default branch, so nothing closed them automatically. Closing now that the fix is in; it reaches main with the next release.