Your 1MB write hits disk as 30MB on a healthy RocksDB instance

Ajouté : 2026-05-17

[00:00:00]Your 1 MB write hits disk as 30 MB, same [music] data written eight times over.

[00:00:05]Writes land in memory first. When that buffer fills, it dumps to disk as a sorted file. So far, one write equals one write. Data lives in levels. Each one is 10 times bigger than the level above. When a level fills, the engine merges it into the next one down, rewriting every byte already there. Four levels deep, that's eight times the write you asked for.

[00:00:26]>> [music] >> Here's the catch. Fewer levels, less rewriting, but slower lookups. More levels, faster queries, but bigger write amp. So, pick the compaction strategy that fits your workload. Write [music] heavy logs go size tiered, like Cassandra. Read heavy lookups go leveled, like RocksDB. Your storage engine choice is your write amp curve.

[00:00:46]Same bytes eight times over. Pick the strategy, not the database.