When studying RNTuple IO behavior I saw that somehow RNTuple ends up using hundreds of MB memory when storing trigger::TriggerEvent and reco::GlobalHaloData (as part of AOD). Using the streamer mode for both data products reduces the peak allocated memory usage by ~533 MB. The purpose of this issue is to document the investigation and discuss options forward.
I'm currently using CMSSW_16_0_0_pre4 as the base for tests, and the AOD file is from /TTtoLNu2Q_TuneCP5_13p6TeV_powheg-pythia8/RunIII2024Summer24DRPremix-140X_mcRun3_2024_realistic_v26-v2/AODSIM.