load_stats_for_tablet_based_tables() called read_barrier() on the
group0 raft server without holding the group0 gate. This allowed
abort_and_drain() to proceed and destroy the server while
read_barrier() was still in flight — a use-after-free detected by
UBSan at raft/server.cc:1573.
Fix by holding hold_group0_gate() in load_stats_for_tablet_based_tables()
before calling read_barrier(). This prevents abort_and_drain() from
completing until the read_barrier operation finishes, matching the
contract that all group0 clients must hold the gate across async
server operations.
Fixes: SCYLLADB-2071
Backport: The bug is present in 2026.1 and 2026.2 therefore we should also backport there.
Closesscylladb/scylladb#30037
939ea402a4