Size-Tiered Compaction (Cassandra)

Size-tiered compaction (STCS) is the simplest strategy: when N SSTables of similar size accumulate, merge them all into one larger SSTable. Repeat at the next tier when N of those accumulate. Cassandra has used this since version 0.

The Algorithm

A "tier" is informal — it's just a group of files that happen to be the same size. With min_count = 4:

• Flush 4 memtables → 4 small SSTables → STCS sees 4 of similar size → merge into 1 medium SSTable.
• After 4 more flushes → 4 small + 1 medium = STCS sees 4 small, merges to 1 medium. Now have 2 medium.
• After 16 total flushes → 4 medium → merge into 1 large.
• And so on.

Properties

• Low write amplification. A byte is rewritten roughly once per tier it advances through. With log4(N) tiers for N files, WA ≈ log4(N).
• High read amplification. Worst case: a query checks one SSTable per tier — log4(N) for negative lookups, and the bloom filter doesn't help against false positives that ARE present somewhere.
• High space amplification. Until a tier fills up and merges, obsolete versions sit around. STCS can have 2x the live data size on disk during periods of heavy writes.
• Bursty compaction. When a top-tier compaction runs, it merges potentially hundreds of GB at once. Big I/O storm; can stall reads.

Where It Wins

• Write-heavy workloads where read latency tolerance is loose (Cassandra's primary target: time-series, event logs).
• Append-mostly workloads with few updates/deletes (less benefit from aggressive merging).
• Operationally simple — only one tunable (min_count).

Where It Loses

• Read-heavy workloads suffer the log(N) tier penalty.
• Update-heavy workloads pay the space amp cost.
• Tail latency is bad — compaction storms cause P99 spikes.

Cassandra's Tunables

min_threshold: 4         # min SSTables per size class before compacting
max_threshold: 32        # max SSTables to compact in one job
bucket_low: 0.5          # tolerance below average size
bucket_high: 1.5         # tolerance above average size

bucket_low and bucket_high define when files count as "same size."

Why Cassandra Defaulted to STCS

Cassandra was designed for write-amplification-sensitive workloads on rotational disks where every write hurt. STCS's lower WA was critical. Modern Cassandra clusters often switch to Leveled for read-heavy column families and TimeWindow (a STCS variant) for time-series.