Our sorter writes out some integers by using reinterpret_cast, which has different behavior on platforms of different endianness.

Spilled sorter data is then read back into memory here in a way that that differs based on machine endinaness.

Sorter files are only used in this way for resumable index builds.

In practice, this would be a problem in this specific scenario:

• This server is shut down cleanly during the sort phase of a resumable index build
• The user takes a filesystem-level snapshot of all data files
• The user migrates to a platform with different endianness and starts the server the server on those data files