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. Changing strategy resets the experiment to an empty run, clearing staged keys and files; note the first result before switching.

Write burst (8 keys)

Adds 8 writes, small enough to remain in memory. Use Flush memtable to L0 to create a file; repeated writes can also reach the byte limit and flush automatically.

Flush memtable to L0

Writes all staged keys and deletion markers to one new Level-0 file. In Size-tiered, flushing a tombstone with a write burst can make its file closer in size to the burst files it needs to merge with.

Delete latest key

Available after the latest key has a persisted value. It stages a tombstone over that value; flush the marker before compaction can consider it. Only one tombstone is followed at a time, and its status stays visible until a safe merge clears it.

Run compaction pass

Becomes available when the selected strategy has an eligible merge: leveled moves files from an over-budget level, while size-tiered combines four similarly sized files. A pass can merge other files first; a tombstone stays until the selected merge includes its older copies.

Reset run

Clears the current files and restarts the selected strategy at the same fixed starting state. Repeating the same actions produces the same result.

What to watch for

• When the schematic L0 threshold pauses writes, compare how a compaction pass clears the stall and changes both amplification readouts.
• Several unique-key bursts can look similar under both strategies. Continue until the key sequence wraps and overwrites earlier keys; stale versions may then make the measured trade-off visible.
• For Size-tiered, write and flush a burst, delete its latest key, add one more burst before flushing the marker, then flush two more bursts. The four-file merge clears the marker only if it includes the older copy; a pass over other files leaves it visible.
• Flush repeatedly without compacting and watch the schematic L0 threshold turn red. Replay the same operation sequence under both strategies; results depend on this model’s merge rules.

Why it happens

This schematic model uses simplified file sizes and merge passes, not RocksDB, Cassandra or another engine. Space amplification counts on-disk SSTable bytes divided by all live logical bytes; staged memtable keys count as live data but are not on disk yet, so this interim ratio can be below 1 until flush. In this workload, leveled passes rewrite more bytes while reducing overlapping versions; size-tiered passes rewrite less often while retaining stale versions longer. Actual amplification depends on workload and implementation details.