Files
Libation/Source/_Tests/LibationSearchEngine.Tests/CorruptIndexRecoveryTests.cs
T
Cursor Agentandrmcrackan 77bbc1b0a0 Treat a Windows sharing violation on write.lock as a lock conflict
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>
2026-08-16 17:25:01 +00:00

254 lines
8.4 KiB
C#

using System;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using AssertionHelper;
using DataLayer;
using LibationSearchEngine;
using Lucene.Net.Index;
using Lucene.Net.Store;
using Microsoft.VisualStudio.TestTools.UnitTesting;
using Directory = System.IO.Directory;
namespace SearchEngineTests;
/// <summary>
/// A damaged search index used to block library imports forever: Lucene 3 reports an unreadable segments file as a
/// plain <see cref="IOException"/>, which Libation mistook for a write.lock conflict and retried instead of repairing,
/// and passing create/overwrite to <see cref="IndexWriter"/> does not repair it either.
/// </summary>
[TestClass]
public class CorruptIndexRecoveryTests
{
private string indexDirectory = null!;
[TestInitialize]
public void Initialize()
{
indexDirectory = Path.Combine(Path.GetTempPath(), "LibationSearchEngineTests", Guid.NewGuid().ToString("N"));
Directory.CreateDirectory(indexDirectory);
}
[TestCleanup]
public void Cleanup()
{
try
{
if (Directory.Exists(indexDirectory))
Directory.Delete(indexDirectory, recursive: true);
}
catch (IOException)
{
// Windows refuses to delete a file Lucene still holds open, and a leftover temp directory is not
// worth failing a test over
}
}
private static LibraryBook book(string asin, string title)
{
var contributor = Contributor.GetEmpty();
var b = new Book(new AudibleProductId(asin), title, null, null, 1, ContentType.Product, [contributor], [contributor], "us");
return new LibraryBook(b, new DateTime(2026, 8, 15), "account");
}
private static readonly List<LibraryBook> library
= [book("B0TEST0001", "Hound of the Baskervilles"), book("B0TEST0002", "Sign of the Four")];
private SearchEngine reIndex()
{
var engine = new SearchEngine(indexDirectory);
engine.CreateNewIndex(library);
return engine;
}
private void assertIndexIsUsable()
{
var engine = new SearchEngine(indexDirectory);
engine.Search("Baskervilles").Docs.Select(d => d.ProductId).Should().BeEquivalentTo(["B0TEST0001"]);
engine.Search(SearchEngine.ALL_QUERY).Docs.Should().HaveCount(2);
}
private string currentSegmentsFile()
=> Directory.GetFiles(indexDirectory, "segments_*").OrderBy(f => f.Length).ThenBy(f => f).Last();
/// <summary>Reproduces the state Lucene leaves behind when a commit is interrupted, eg. by a crash or power loss.</summary>
private void truncateCurrentSegmentsFile(int keepBytes)
{
var file = currentSegmentsFile();
File.WriteAllBytes(file, File.ReadAllBytes(file).Take(keepBytes).ToArray());
}
private static Exception exceptionFromOpeningWriter(string directory)
{
using var index = FSDirectory.Open(directory);
using var analyzer = new Lucene.Net.Analysis.Standard.StandardAnalyzer(SearchEngine.Version);
return Assert.ThrowsExactly<IOException>(() => new IndexWriter(index, analyzer, true, IndexWriter.MaxFieldLength.UNLIMITED).Dispose());
}
[TestMethod]
public void full_reindex_builds_a_searchable_index()
{
reIndex();
assertIndexIsUsable();
}
[TestMethod]
public void full_reindex_repairs_a_zero_length_segments_file()
{
reIndex();
truncateCurrentSegmentsFile(0);
reIndex();
assertIndexIsUsable();
}
[TestMethod]
public void full_reindex_repairs_a_partially_written_segments_file()
{
reIndex();
truncateCurrentSegmentsFile(File.ReadAllBytes(currentSegmentsFile()).Length / 2);
reIndex();
assertIndexIsUsable();
}
/// <summary>
/// The nastiest variant: the current commit is intact, so the index looks fine, but IndexWriter's IndexFileDeleter
/// reads every segments_* file it finds and one unreadable leftover poisons the whole directory.
/// </summary>
[TestMethod]
public void full_reindex_repairs_a_stale_unreadable_segments_file_beside_a_valid_commit()
{
reIndex();
File.WriteAllBytes(Path.Combine(indexDirectory, "segments_9"), []);
reIndex();
assertIndexIsUsable();
}
[TestMethod]
public void full_reindex_repairs_a_checksum_mismatch()
{
reIndex();
var file = currentSegmentsFile();
var bytes = File.ReadAllBytes(file);
bytes[^1] ^= 0xFF;
File.WriteAllBytes(file, bytes);
reIndex();
assertIndexIsUsable();
}
/// <summary>Cloud sync (eg. OneDrive) leaves conflict copies whose names Lucene 3's segments parser rejects.</summary>
[TestMethod]
public void full_reindex_repairs_cloud_sync_debris()
{
reIndex();
File.WriteAllText(Path.Combine(indexDirectory, "segments_2 (1)"), "conflict copy");
reIndex();
assertIndexIsUsable();
}
/// <summary>
/// A garbled segments.gen sends Lucene looking for an absurd generation, and Lucene 3's base-36 filename
/// formatter overruns its buffer on the way. Damage does not always announce itself as an IOException.
/// </summary>
[TestMethod]
public void full_reindex_repairs_a_segments_gen_pointing_at_a_bogus_generation()
{
reIndex();
var gen = Path.Combine(indexDirectory, "segments.gen");
// format(int) then the generation as a long, written twice; corrupt the high bytes of both copies
var bytes = File.ReadAllBytes(gen);
bytes[7] = 0x63;
bytes[15] = 0x63;
File.WriteAllBytes(gen, bytes);
reIndex();
assertIndexIsUsable();
}
[TestMethod]
public void search_repairs_a_zero_length_segments_file()
{
reIndex();
truncateCurrentSegmentsFile(0);
// the query path recovers via SearchEngineCommands, which keys off IsRecoverableCorruptIndexException
var ex = Assert.ThrowsExactly<IOException>(() => new SearchEngine(indexDirectory).Search("Baskervilles"));
SearchEngine.IsRecoverableCorruptIndexException(ex).Should().BeTrue();
}
[TestMethod]
public void an_unreadable_segments_file_is_classified_as_recoverable_corruption()
{
reIndex();
truncateCurrentSegmentsFile(0);
var ex = exceptionFromOpeningWriter(indexDirectory);
(ex is CorruptIndexException).Should().BeFalse();
ex.Message.Should().Be("read past EOF");
SearchEngine.IsRecoverableCorruptIndexException(ex).Should().BeTrue();
}
/// <summary>
/// A held write.lock is transient and must keep being retried. It derives from <see cref="IOException"/>, so
/// widening the corruption check to cover IOException must not swallow it and delete a healthy index.
/// </summary>
[TestMethod]
public void a_write_lock_conflict_is_not_classified_as_corruption()
{
SearchEngine.IsRecoverableCorruptIndexException(new LockObtainFailedException("Lock obtain timed out")).Should().BeFalse();
SearchEngine.IsRecoverableCorruptIndexException(new UnauthorizedAccessException()).Should().BeFalse();
// Windows raises the sharing violation on the lock file before Lucene can turn it into a
// LockObtainFailedException, so this arrives as a plain IOException. Mistaking it for corruption would
// delete the index the other holder is using.
SearchEngine.IsRecoverableCorruptIndexException(
new IOException(@"The process cannot access the file 'C:\Users\me\Libation\SearchEngine\write.lock' because it is being used by another process."))
.Should().BeFalse();
}
/// <summary>
/// The risk in repairing anything that is not a lock conflict is over-reach, so a held write.lock has to keep
/// being retried and must leave the index alone. The lock is held for the whole retry budget rather than
/// released part way through, which would race with Lucene 3's own lock bookkeeping.
/// </summary>
[TestMethod]
public void a_write_lock_conflict_is_retried_and_leaves_the_index_intact()
{
reIndex();
var indexedBefore = indexFiles();
using var index = FSDirectory.Open(indexDirectory);
var writeLock = index.MakeLock(IndexWriter.WRITE_LOCK_NAME);
writeLock.Obtain().Should().BeTrue();
try
{
// Linux surfaces this as Lucene's LockObtainFailedException, Windows as a sharing violation on the
// lock file. Either way it must not be read as corruption, and the index must survive.
var ex = Assert.Throws<IOException>(() => reIndex());
SearchEngine.IsRecoverableCorruptIndexException(ex).Should().BeFalse();
indexFiles().Should().BeEquivalentTo(indexedBefore);
}
finally
{
// a competing NativeFSLock.Obtain can delete the lock file it failed to take, leaving Release nothing
// to delete. Lucene 3's lock bookkeeping is not what this test is about.
try { writeLock.Release(); }
catch (Exception ex) when (ex is LockReleaseFailedException or IOException) { }
}
}
private List<string> indexFiles()
=> [.. Directory.GetFiles(indexDirectory)
.Select(Path.GetFileName)
.OfType<string>()
.Where(f => f != IndexWriter.WRITE_LOCK_NAME)
.OrderBy(f => f)];
}