moli/moli-leveldb-parser/TESTS.md

Chromium LevelDB test ports

The source is Chromium's actual LevelDB dependency, pinned to 7ee830d02b623e8ffe0b95d59a74db1e58da04c5. Chromium revision d9fa5a15858b7189aa1a70cebe811169154ebcf0 selects that gitlink and identifies it as LevelDB 1.23 in README.chromium. The tests retain the upstream license and authors.

The suite has 133 tests: 68 ports/adaptations of the scenarios listed below, 56 additional parser regressions, and the 9 original snapshot tests. This is coverage of the read-only parser, not the complete Chromium LevelDB engine suite. Each original scenario keeps its snake_case name, with a suffix where its observable behavior was adapted. Supporting fixtures are test-only code.

cargo test -p moli-leveldb-parser -p moli-cookie-import --lib
cargo fmt --all
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo nextest run --no-fail-fast

WAL physical records

Source: db/log_test.cc. Port: src/tests/log.rs, 31 tests, including these 23 upstream cases:

Upstream cases	What is checked
Empty, ReadWrite, ManyBlocks	Empty records and 100,000 records over many physical blocks
Fragmentation, RandomRead	50/100 KB values and 500 skewed-size records, fixed seed 301
MarginalTrailer, MarginalTrailer2, ShortTrailer, AlignedEof	Exact block ends, short padding and empty FIRST fragments
OpenForAppend	Reading a log whose writer resumed at a nonzero block offset
ReadError	An injected I/O error propagates
BadRecordType, BadLength, ChecksumMismatch	Unknown types, records crossing a block, and bad CRC32C
UnexpectedMiddleType, UnexpectedLastType, UnexpectedFullType, UnexpectedFirstType	Invalid fragment ordering
TruncatedTrailingRecordIsIgnored, BadLengthAtEndIsIgnored, MissingLastIsIgnored, PartialLastIsIgnored	Torn final appends never emit partial records
ErrorJoinsRecords	Corruption must not join fragments belonging to different logical records

Upstream's reader reports corruption and can continue salvaging records. The parser returns Err at complete corruption; the public snapshot API exposes no partial result. Tests assert errors instead of upstream dropped-byte counts and salvaged data. Incomplete final appends retain upstream EOF behavior.

The remaining 15 upstream cases (SkipIntoMultiRecord, ReadStart, ReadSecond*, ReadThird*, ReadFourth*, ReadInitialOffsetIntoBlockPadding, ReadEnd, ReadPastEnd) exercise an initial-offset/seek API. This parser always starts at byte zero, so those API-specific cases are outside its scope.

Additional tests cover the empty-FIRST compatibility exception documented in db/log_reader.cc, checksums of every fragment kind, zero-filled preallocation, corrupt padding, short/interrupted reads, callback errors, and cuts around every fragment boundary.

Coding, CRC and WriteBatch

The four additional layout tests use literal wire bytes at every alignment modulo 8 to check borrowed headers, trailers and restart arrays, little-endian values, truncation without cursor advancement, and nonempty blocks with zero restarts. Pointer checks ensure views refer to the original input; the existing corruption tests continue to validate semantics.

Source	Ported cases	Local tests
util/coding_test.cc	All 10: Fixed32, Fixed64, EncodingOutput, Varint32, Varint64, both overflow cases, both truncation cases, Strings	coding.rs
util/crc32c_test.cc	StandardResults, Values, Mask	coding.rs
db/write_batch_test.cc	Empty, Multiple, Corruption, Append	write_batch.rs

EncodingOutput checks little-endian decoding against literal bytes; the parser has no encoder. CRC.Mask checks the production masking wrapper against its unmasked checksum. CRC.Extend and WriteBatch.ApproximateSize test APIs absent from this crate. Append checks the combined on-disk batch; the parser retains the latest version per key rather than all memtable history.

Added regressions cover truncated fixed fields and length prefixes, overlong varint32 encodings, high bits that would overflow, every batch truncation point, counts too small/large, unsupported tags, 56-bit sequence limits, empty keys and values, ordered deletes, and conflicting values at the same sequence.

MANIFEST / VersionEdit

Source: db/version_edit_test.cc. Port: src/tests/manifest.rs, 16 tests.

VersionEditTest.EncodeDecode preserves the large (1 << 50) file, size and sequence values, additions, removals and compaction pointers. Assertions check decoded manifest state instead of encoding it again. Its arbitrary foo comparator is tested as an explicit unsupported-comparator error; supported snapshots use leveldb.BytewiseComparator.

Additional cases cover metadata split across records, missing metadata, previous-log updates, every VersionEdit field's truncation, unknown tags, invalid levels/internal keys, sequence overflow, file deletion and level moves, deletions-before-additions in one edit, CURRENT validation/selection, incomplete final edits, and complete edits with invalid checksums. ComparatorCheck also corresponds to the custom-comparator mismatch scenario in db/db_test.cc.

SST blocks and internal keys

Sources: table/table_test.cc, db/dbformat_test.cc. Port: src/tests/table.rs, 23 tests.

Upstream cases	Adaptation
Harness.Empty, SimpleEmptyKey, SimpleSingle, SimpleMulti, SimpleSpecialKey	Block and table forward scans; restart intervals 1/16/1024; raw and Snappy data/index blocks
Harness.ZeroRestartPointsInBlock	The four-byte empty block emitted by Java LevelDB
Harness.Randomized	Original 0..2000 entry-count schedule, binary-key alphabet and skewed lengths; fixed seed 306
Harness.RandomizedLongDB	Checked-in native fixture from 100,000 writes, fixed seed 301; Rust independently reconstructs all 25,479 final entries
TableTest.ApproximateOffsetOfPlain	Original values up to 300 KB; verifies full contents instead of an offset-estimation API
CompressionTableTest.ApproximateOffsetOfCompressed	Original 25%-compressible workload; verifies Snappy-decoded contents
FormatTest.InternalKey_EncodeDecode, InternalKey_DecodeFromEmpty	Snapshot decoding of empty/long keys and sequence boundaries; rejection of short internal keys

The read-only API exposes a full scan. Reverse/custom comparators, backward iterators, seeks, memtable constructors, key-shortening/debug formatting, and approximate offsets have no corresponding API here. The pinned Chromium port_chromium.h returns false for Zstd support, so the Zstd parameter is not ported; compression types other than raw/Snappy are explicitly rejected.

Added regressions cover footer size/magic, file-size disagreement with MANIFEST, out-of-bounds block handles, data/index CRCs, invalid compression with valid CRCs, restart counts/offsets/order, illegal shared prefixes, entries crossing restart arrays, unknown internal value types, and callback errors. Existing snapshot tests also cover malicious Snappy lengths and missing/legacy tables.

Recovery, versions and tombstones

Sources: db/recovery_test.cc, db/db_test.cc, db/corruption_test.cc. Port: src/tests/recovery.rs, 21 tests.

Upstream cases	Read-only expectation
RecoveryTest.ManifestReused, LargeManifestCompacted	Appended edits and a 3 MB zero-padded MANIFEST read successfully; source files stay unchanged
RecoveryTest.NoLogFiles, LogFileReuse	Missing logs on a valid snapshot and empty/nonempty current WALs
RecoveryTest.MultipleMemTables	Original 1,000 padded keys recovered entirely in memory
RecoveryTest.MultipleLogFiles	Recover newer unregistered WALs while ignoring obsolete logs, even with larger sequences
RecoveryTest.ManifestMissing	Fail without creating/replacing files
DBTest.GetLevel0Ordering, GetOrderedByLevels	Newest visible version wins across overlapping L0 files and different levels
DBTest.Recover, RecoveryWithEmptyLog	Repeated updates across WAL/SST transitions
DBTest.IterMultiWithDelete, IterMultiWithDeleteAndCompaction	Deleted entries absent from output with both WAL-only and SST+WAL data
CorruptionTest.SequenceNumberRecovery	Recovered sequence ordering keeps subsequent updates visible; repair itself is outside scope
CorruptionTest.Recovery	Corruption in the original two WAL locations returns an error instead of salvaged partial data

These are 15 upstream scenarios. Added regressions cover native Chromium SST/WAL compatibility, future SST versions, tombstones newer than WAL values, empty batches, malformed batches without partial output, and duplicate numeric WAL filenames. Recovery checks compare the input files before and after reads.

The nine existing snapshot tests moved from the Chrome importer. They retain engine comparison, mtime/content checks, a held writer lock, read-only filesystem permissions, obsolete-file resurrection checks, .sst fallback/missing live tables (DBTest.StillReadSST/MissingSSTFile scenarios), corruption, previous WAL selection, and UTF-agnostic byte handling. Chrome string/JSON import tests remain in moli-cookie-import.

Native fixtures, their C++ generator, checksums and regeneration instructions are in tests/fixtures. Normal tests require no C++ compiler or network. Compaction scheduling, writable recovery/repair, concurrency, caches/Bloom-filter performance, locking policy, and write-failure injection for a database engine are outside the parser's API and are not claimed as ported.