Bad things happen otherwise.  One case in point is create database
followed by a crash.  Create database requests a checkpoint after
inserting new tuple into pg_database.  Crash happens right after
create database commits and before clog update is not flushed to disk.
Relcache initialization before xlog replay will set HEAP_XMIN_INVALID
hint bit in the newly created database's tuple because clog did not
report the xmin as committed.