The behaviour you describe happens exactly when two processes e.g. concurrently hold HANDLEs on the same file and one of those deletes the file then. Windows keeps file names until all open HANDLEs are closed and depending on how those HANDLEs have been opened by the first app, concurrent deletion is perferctly fine for Windows. Though, a such deleted file can't be opened easily anymore and looks like it has lost permissions only. But that's not the case, it's deleted already. It might be that this happens for Postgres to itself somehow when some other app has an open HANDLE. I don't think that some other app is deleting that file by purpose instead, reading it for some reason seems more likely to me.
Using Process Monitor, Thorsten's explanation above appears to
correctly diagnose what is happening. ProcMon data shows
postgres.exe performing "CreateFile" operations on the affected
WAL files, with the result status "DELETE PENDING". Which
according to https://stackoverflow.com/a/29892104 means:
"Windows allows a process to delete a file, even though it is still opened by another process (e.g. Windows indexing service or Antivirus). It gets internally marked as "delete pending". The file does not actually get removed from the file system, it is still there after the File.Delete call. Anybody that tries to open the file after that gets an access denied error. The file doesn't actually get removed until the last handle to the file object gets closed"
which is the same behaviour Thorsten describes above (great info,
thanks Thorsten).
The mystery now is that the only process logged as touching the
affected WAL files is postgres.exe (of which there are many
separate processes). Could it be that one of the postgres.exe
instances is holding the affected WAL files in use after another
postgres.exe instance has flagged the file as deleted? (or to put
it the other way, a postgres.exe instance is flagging the file as
deleted while another instance still has an open handle to the
file)? If it is some other process such as the indexer (disabled)
or AV (excluded from pgdata) is obtaining a handle on the WAL
files, it isn't being logged by ProcMon.
Kind regards,
Guy