Windows CI caught real over-reach. When another holder has write.lock, Windows
raises the sharing violation before Lucene can turn it into a
LockObtainFailedException, so it arrives as a plain IOException. Repairing anything
that is not a recognised lock conflict then meant deleting the index the other
holder was using -- exactly the second-instance case the retry exists for.
An IOException naming Lucene's write lock now counts as a lock conflict. Matching
the file name rather than the message wording keeps it working on non-English
Windows. The end-to-end test asserts the property instead of the exception type,
since the type legitimately differs by platform.
Co-authored-by: rmcrackan <rmcrackan@gmail.com>
Releasing the write.lock part way through the retry budget raced with Lucene 3's
own lock bookkeeping: on Windows a competing Obtain left a handle on the file, so
Release and the temp directory cleanup both failed with a sharing violation. Hold
the lock for the whole budget instead and assert what actually matters, that a lock
conflict is retried and leaves the index files alone.
Co-authored-by: rmcrackan <rmcrackan@gmail.com>
Lucene 3's base-36 filename formatter overruns its buffer when segments.gen names
an absurd generation, so the rebuild path has to survive an IndexOutOfRangeException
as well as the IOException shapes. Damage does not always announce itself as IO.
Co-authored-by: rmcrackan <rmcrackan@gmail.com>
A truncated or zero-length segments file is reported by Lucene 3 as a plain
IOException ("read past EOF") rather than a CorruptIndexException, so it was
misclassified as a write.lock conflict: CreateNewIndex burned its whole backoff
budget and rethrew, and the delete-and-rebuild recovery never ran. Passing
create/overwrite to IndexWriter does not repair it either, because
IndexFileDeleter reads every segments_* file in the directory and tolerates only
missing ones, so a single unreadable segments file -- even a stale one from an
older commit -- leaves the index permanently unopenable. The user's only cure
was deleting the SearchEngine folder by hand.
Retries are now reserved for genuine lock conflicts (LockObtainFailedException,
which derives from IOException, and UnauthorizedAccessException), and any other
open failure gets one delete-and-rebuild pass before giving up with a message
that says which folder to remove. The query path recovers too, since
IsRecoverableCorruptIndexException now recognizes the truncated-segments
signature.
Search index updates are also no longer allowed to fail the library change that
triggered them. Both events fire after the database is committed, so an escaping
exception reported a successful scan as "Error importing library" and, being the
first subscriber, stopped the handlers that refresh the grid and backup counts.
Co-authored-by: rmcrackan <rmcrackan@gmail.com>
