db-09 — References

Primary sources

Sanjay Ghemawat & Jeff Dean, LevelDB design document, Google, 2011. https://github.com/google/leveldb/blob/main/doc/index.md
LevelDB implementation notes, doc/impl.md — describes the on-disk layout (MANIFEST, log files, table files) and the recovery procedure that db-09 mirrors almost verbatim. https://github.com/google/leveldb/blob/main/doc/impl.md
Patrick O'Neil et al., The Log-Structured Merge-Tree (LSM-Tree), Acta Informatica 33, 1996. The original paper; introduces the C0/C1 levels, the merge step, and the amortized write-cost analysis.

LevelDB (Google) — direct ancestor of our design. https://github.com/google/leveldb
RocksDB (Facebook/Meta) — adds column families, leveled compaction, bloom filters per file, and many tuning knobs but keeps the same WAL → memtable → flush → L0 → … shape. https://github.com/facebook/rocksdb/wiki
Pebble (CockroachDB) — RocksDB-compatible engine in Go; very readable. https://github.com/cockroachdb/pebble
HBase HFile / Cassandra SSTables — same on-disk philosophy.

Mark Callaghan, Read, write & space amplification, 2018 — explains why the "newest source wins" rule is required and how compaction trades read amplification for write amplification. https://smalldatum.blogspot.com/2018/09/read-write-and-space-amplification.html
Pebble's docs/rocksdb.md for an excellent diff-style walkthrough of how a modern engine differs from LevelDB while preserving the same correctness invariants.

Pillai, Chidambaram, et al., All File Systems Are Not Created Equal, OSDI '14. The "fsync the file, then fsync the directory" rule we follow for SST publish and MANIFEST rewrite comes from this work.

db-03 (WAL) — record framing, torn-tail tolerance, WalIter.
db-05 (MemTable) — sorted map with explicit tombstones.
db-06 (SSTable) — on-disk sorted-string file format with a footer and trailing checksum.
db-08 (BlockCache + MergingIterator) — k-way merge with newer-source-wins and optional tombstone dropping; canonical SerializeStream used by Db::serialize_view.