Compaction Foundry

Controls

Compaction strategy

Leveled merges overlapping files into clean, non-overlapping levels — more rewriting, less wasted space. Size-tiered merges same-sized files together less eagerly — less rewriting, more stale data survives longer.

Write burst (N keys)

Writes a batch of keys — some repeats, some deletes — into the in-memory memtable.

Flush memtable to L0

Forces the current memtable to disk as a new Level-0 file, even if it hasn't hit its size threshold yet.

Run compaction pass

Triggers one merge step — the lowest over-budget level gets folded into the next one down.

What to watch for

• Run the same sequence of write bursts and compaction passes under Leveled, note Write amplification, then switch to Size-tiered and repeat — leveled's number is consistently higher.
• Compare Space amplification the same way — size-tiered's is higher.
• If you flush repeatedly without compacting, the L0 SSTables count grows until the lane's label turns red — a real read/write stall condition.

Why it happens

Leveled compaction keeps each level's key ranges non-overlapping, which means a key gets rewritten every time it cascades one level deeper — more total bytes written per logical byte, but old or duplicate versions get cleaned up quickly. Size-tiered only merges same-sized files together and tolerates overlapping ranges for longer, so a key is rewritten less often (lower write amplification) but stale versions of it can survive on disk much longer (higher space amplification). Neither is 'better' — it's the same tradeoff RocksDB, Cassandra, and every other LSM-based store expose as a tuning knob.