Cassandra SSTable Format Parsing

This skill helps with parsing and understanding Cassandra 5.0+ SSTable file formats.

When to Use This Skill

• Parsing Data.db, Index.db, Statistics.db files
• Debugging binary format mismatches
• Analyzing hex dumps of SSTable data
• Working with compression (LZ4, Snappy, Deflate, Zstd; Noop = stored raw)
• Investigating offset calculation errors
• Understanding BTI (Big Table Index) format
• Validating partition boundaries

Key SSTable Components

Data.db

Contains the actual row data with:

• Partition headers
• Row data (clustering + cells)
• Compression blocks
• Checksums

Index.db

Contains partition index entries with:

• BTI (Big Table Index) format in Cassandra 5.0+
• Partition key → file offset mapping
• Promoted index entries

Statistics.db

Contains serialization metadata:

• Encoding stats (min/max timestamps, TTLs)
• Column definitions
• Schema information
• Compression parameters

Summary.db

Contains sampling of index entries for faster lookups

Format References

Primary Source of Truth: docs/sstables-definitive-guide/

Key chapters:

• Ch.5: Data.db Format - Row layout, flags, V5 row/partition encoding
• Ch.6: Index.db and Summary.db - Partition lookups
• Ch.9: CompressionInfo.db - Compression metadata, chunking
• Ch.17: BTI Formats - Trie-based indexes
• Appendix B: Encoding Cheat Sheet - VInt, cell flags
• Appendix F: Known Limitations - What doesn't work yet

Common Debugging Techniques

Hex Dump Analysis

When debugging parsing errors:

1. Extract hex at specific offset:

   hexdump -C Data.db -s <offset> -n 64

2. Compare with expected format:

   • Check magic bytes (if applicable)
   • Verify VInt encoding
   • Validate flag bytes

3. Look for patterns:

   • Repeated byte sequences may indicate arrays/collections
   • All zeros may indicate padding
   • Non-zero high bytes suggest multi-byte integers

Offset Validation

Track byte consumption at each parsing stage:

• Clustering prefix (may be 0 bytes)
• Row sizes (2 VInts)
• Liveness info (conditional)
• Deletion info (conditional)
• Column bitmap (conditional)
• Cell data

Zero-Copy Considerations

When implementing parsers:

• Use Bytes crate for buffer sharing
• Avoid copying large cell values
• Keep references to original buffer
• Use byte slices not owned Vecs

Integration with Rust Code

Current implementation in cqlite-core/src/storage/sstable/reader/parsing/:

• row_decoder/ — the main V5 row/partition decoder, a directory of ~30 files
  (split out of the former single-file parser by epic #1116). Start at
  row_decoder/row_flags.rs for the flag constants and row_decoder/mod.rs for the
  parser entry point; then
  row_framing.rs (row/partition framing), row_data.rs, cell_value_scalar.rs /
  cell_value_complex.rs (cell decode), complex_column.rs (non-frozen collections),
  frozen.rs (frozen collections), udt.rs, partition_driver.rs.
• Uses zero-copy patterns with Bytes
• Handles compression transparently

The former single-file V5-compressed-legacy parser module no longer exists — it was
deleted by epic #1116 (source splits), commit cb049f7a8, and replaced by the
row_decoder/ directory above. If an older doc or comment points you at a single
.rs file for the V5 parser, that pointer is stale.

PRD Alignment

Supports Milestone M1 (Core Reading Library):

• 100% Cassandra 5 SSTable format support
• All compression formats (LZ4, Snappy, Deflate, Zstd) plus Noop
• Zero-copy deserialization
• Memory target: <128MB for large files

Quick Reference

Flag Bytes (Row) — main flag byte

Value	Name	Meaning
0x01	END_OF_PARTITION	End-of-partition marker — nothing follows this flag byte
0x02	IS_MARKER	Unfiltered is a RangeTombstoneMarker, not a Row
0x04	ROW_HAS_TIMESTAMP	Row-level liveness timestamp present (delta-encoded)
0x08	ROW_HAS_TTL	Row-level TTL present (delta-encoded)
0x10	ROW_HAS_DELETION	Row deletion (tombstone) present
0x20	ROW_HAS_ALL_COLUMNS	All schema columns present — no column bitmap follows
0x40	ROW_HAS_COMPLEX_DELETION	Row carries a non-frozen collection column with deletion info
0x80	ROW_HAS_EXTENDED_FLAGS	A second (extended) flag byte follows

Flag Bytes (Row) — EXTENDED flag byte (only when 0x80 is set)

Value	Name	Meaning
0x01	EXTENDED_IS_STATIC	Static row — has NO clustering prefix

Citations: cqlite-core/src/storage/sstable/reader/parsing/row_decoder/row_flags.rs:12-18
(ROW_HAS_TIMESTAMP/TTL/DELETION/ALL_COLUMNS/COMPLEX_DELETION/EXTENDED_FLAGS),
:24 (END_OF_PARTITION = 0x01), :26 (IS_MARKER = 0x02), :31
(EXTENDED_IS_STATIC = 0x01). Guide: appendix-b-encodings-cheat-sheet.md:206-212.
Cassandra: UnfilteredSerializer.java:102-109 (flags) and :114-122 (extended flags).

⚠️ The two highest-consequence bits. 0x01 on the main byte is
END_OF_PARTITION, NOT IS_STATIC and NOT a marker flag — misreading it means
mis-detecting partition boundaries. IS_STATIC is 0x01 of the EXTENDED byte,
which only exists when ROW_HAS_EXTENDED_FLAGS (0x80) is set. There is exactly ONE
HAS_ALL_COLUMNS and its value is 0x20.

VInt Encoding

Variable-length integer encoding:

• First byte indicates length
• Subsequent bytes contain value
• Used for row sizes, timestamps, offsets

Format authority (#3041)

A CQLite file:line is NEVER format authority. Citing CQLite's own code to justify CQLite's
behavior is circular reasoning. Authority is, in order:

1. The pinned cassandra-5.0.8 Cassandra source — read it as a tag ref, never a working tree:

   git show cassandra-5.0.8:src/java/org/apache/cassandra/db/rows/UnfilteredSerializer.java

   (Browse: https://github.com/apache/cassandra/tree/cassandra-5.0.8.) A local clone is optional and
   branch-sensitive — none is guaranteed to exist, and one that does may sit on trunk/6.0-alpha,
   which is not the 5.0 on-disk format; read it via git -C "$CQLITE_CASSANDRA_REPO" show cassandra-5.0.8:<path>.
2. sstabledump output on a real SSTable.
3. docs/sstables-definitive-guide/.

A CQLite source line is evidence of what CQLite does, never of what is correct.

Next Steps

When parser encounters issues:

1. Log byte offsets at each stage
2. Compare against the pinned Java source (git show cassandra-5.0.8:…/UnfilteredSerializer.java)
3. Validate against sstabledump output
4. Check compression block boundaries
5. Verify delta encoding calculations