Notesstorage🌱 seedling
Memtable
Also: Memory Table · Write Buffer
The in-memory, sorted buffer at the top of an [[LSM-Tree]] where every write lands first. When it fills, it is frozen and flushed to disk as one sorted, immutable [[SSTable]].
Why it exists
Writes arrive in random key order, but an LSM-tree only ever writes sorted files sequentially. The memtable absorbs random inserts in memory, where sorting is cheap, and turns them into one large sequential write.
How it works
- Each write is appended to the Write-Ahead Log for durability, then inserted into the memtable: usually a skip list, sometimes a balanced tree.
- Reads check the memtable first, because it holds the newest version of any recent key.
- At a size limit (RocksDB's default is 64 MB), the memtable becomes immutable and a fresh one takes new writes while the old one is flushed in the background.
- Once the flush completes, its log records are no longer needed; a WAL file is deleted once everything in it is flushed.
- Deletes are inserted as tombstones, like any other write.
Trade-offs
| Bigger memtable | Smaller memtable |
|---|---|
| fewer, larger flushes; fewer L0 files | less memory, faster recovery |
| more updates to the same key coalesce in RAM | more flushes and more compaction work |
| longer WAL replay after a crash | less coalescing of hot keys |
In practice
- RocksDB lets several immutable memtables queue for flush; if they pile up faster than flushes finish, it stalls writes.
- Cassandra flushes memtables per table, by size or on a commit-log threshold.
Pitfalls
- Memtable memory is per column family or table. Many tables with generous memtables can quietly exhaust RAM.
- A crash loses the memtable, so durability depends entirely on the WAL settings, not on the memtable.
Connections
- Learn first
- Write-Ahead Log
- Builds on this
- LSM-Tree
- Related
- SSTable
Recall
These cards are also in your review deck, scheduled with spaced repetition.