Thanks for tackling this — JVM auto-sizing in containers reading host memory is a real and well-known footgun, and the cgroup v1/v2 dual handling here is the right shape. A few inline suggestions below; nothing blocking. Two file-level notes:

1. No automated tests. Cgroup-gated shell behavior is awkward to unit-test, but a small harness that parameterizes the cgroup mount point (default /sys/fs/cgroup) and runs the function against fixture files would catch regressions cheaply. Acceptable to skip given the PR description confirms manual cluster verification.
2. Duplication. The same ~17-line block is copy-pasted across confignode-env.sh and datanode-env.sh. The existing calculate_memory_sizes() is already duplicated in both files, so this is consistent with the current style — but a follow-up that sources a shared helper from scripts/conf/iotdb-common.sh would reduce drift.
3. AINode unaffected, right? scripts/sbin/start-ainode.sh is Python-based and presumably doesn't share this Bash sizing machinery — worth confirming so nobody assumes container-aware sizing applies there too.