RocksDB is an embedded key-value store written in C++ and optimized for fast storage environments like flash or RAM. It uses a log-structured merge tree to store data by writing new data sequentially to an in-memory log and memtable, periodically flushing the memtable to disk in sorted SSTables. It reads from the memtable and SSTables, and performs background compaction to merge SSTables and remove overwritten data. RocksDB supports two compaction styles - level style, which stores SSTables in multiple levels sorted by age, and universal style, which stores all SSTables in level 0 sorted by time.

1. 안미진
RocksDB
Embedded Key-Value Store for Flash and RAM

2. Contents
1. RocksDB Introduction
2. RocksDB Architecture
3. LSM DB
4. RocksDB Compaction
Overview

3. RocksDB Introduction
• Open source based on LevelDB 1.5, written in C++
• Key-Value persistent store
• Embedded Library
• Pluggable database
• Optimized for fast storage (flash or RAM)
• Optimized for server workloads
• Get(), Put(), Delete()
A persistent key-value store for fast storage environments

4. Three Basic Constructs
of RocksDB
• Memtable
– in-memory data structure
– A buffer, temporarily host the incoming writes
• Logfile
– Sequentially-written file
– On storage

5. Three Basic Constructs
of RocksDB
• SSTable(=SSTfile)
– Sorted Static Table on storage
– A file which contains a set of arbitrary, sorted key-value
pairs inside
– Organized in levels
– Immutable in its life time
– sorted data → to facilitate easy lookup of keys
– Storage of the entire database

6. SSTable-BlockBasedTable
of RocksDB
Data Data Data …
Meta
(filter)
Meta
(stats)
…
Meta
index
Data
index
Footer
The default SSTable format in RocksDB

7. SSTable & Memtable
of RocksDB
• On-disk SSTable indexes are always loaded into memory
• All writes go directly to the Memtable index
• Reads check the Memtable first → the SSTable indexes
• Periodically, the Memtable is flushed to disk as an SSTable
• Periodically, on-disk SSTables are merged
→ update/delete records will overwrite/remove the older
data

8. Simplified RocksDB
Memory
Storage
Memtable
SSTable 1
SSTable 2
Key Offset
Key Offset
… …
Index
Key Value Key Value … …
Simplified SSTable file

9. RocksDB Architecture
Active
Memtable
Read-Only
Memtable
Memory
Log
Log
SSTSSTSST
SSTSSTSST
Persistent Storage
Write Request
Read Request LSM Files
CompactionFlush
Switch Switch

10. RocksDB Architecture
Active
Memtable
Read-Only
Memtable
Memory
Log
Log
SSTSSTSST
SSTSSTSST
Persistent Storage
Write Request
Read Request LSM Files
CompactionFlush
Switch Switch

11. RocksDB Architecture
Active
Memtable
Read-Only
Memtable
Memory
Log
Log
SSTSSTSST
SSTSSTSST
Persistent Storage
Write Request
Read Request LSM Files
CompactionFlush
Switch Switch

12. RocksDB Architecture
Active
Memtable
Read-Only
Memtable
Memory
Log
Log
SSTSSTSST
SSTSSTSST
Persistent Storage
Write Request
LSM Files
CompactionFlush
Switch Switch
Read Request

13. RocksDB Architecture
Active
Memtable
Read-Only
Memtable
Memory
Log
Log
SSTSSTSST
SSTSSTSST
Persistent Storage
Write Request
LSM Files
CompactionFlush
Switch Switch
Read Request

14. RocksDB Architecture
Active
Memtable
(4MB)
Immutable
Memtable
Memory
Disk
Write
Level 0
(4 SSTfile)
Level 1
(10MB)
Level 2
(100MB)
. . .
. . . . . .
Info Log
MANIFEST
CURRENT
Compaction
Log
SSTfile
(2MB)

15. Log-Structured Merge Tree
• LSM-tree
– N-level merge trees
– Splitting a logical tree into several physical pieces
– So that the most-recently-updated portion of data is in a tree in
memory
– Transform random writes into sequential writes using logfile &
in-memory store(Memtable)

16. Log-Structured Merge DB
to minimize “random writes”
Write RequestRead Request
Read Write data
in RAM
Read Only data in RAM
on disk
Periodic
Compaction
Transaction Log

17. Log-Structured Merge DB
to minimize “random writes”
① Data Write(Insert, Update)
• New puts are written to memory(Memtable) & logfile
sequentially
• Memtable is filled up → flushed to a SSTable on disk
• Operated in memory, no disk access → faster than B+ tree
② Data Read
• Memtable → SSTable
• Maintain all the SSTable indexes in memory

18. RocksDB Compaction
Multi-threaded compactions
• Background Multi-thread
→ periodically do the “compaction”
→ parallel compactions on different parts of the database
can occur simultaneously
• Merge SSTfiles to a bigger SSTfile
• Remove multiple copies of the same key
– Duplicate or overwritten keys
• Process deletions of keys
• Supports two different styles of compaction
– Tunable compaction to trade-off

19. RocksDB Compaction
Storage
SSTable 1
SSTable 2
SSTable 3
SSTable 4
SSTable 5

20. 1. Level Style Compaction
• RocksDB default compaction style
• Stores data in multiple levels in the database
• More recent data → L0
The oldest data → Lmax
• Files in L0
- overlapping keys, sorted by flush time
Files in L1 and higher
- non-overlapping keys, sorted by key
• Each level is 10 times larger than the previous one
Inherited from LevelDB

21. Level Style Compaction
Compaction process
cache
log
level1
level2
level3
level0
① Pick one file from level N
② Compact it with all its overlapping
files from level N+1
③ Replace them with new files in
level N+1

22. Level Style Compaction
Compaction example
5 bytes
6 bytes
10 bytes 10 bytes
11 bytes
10 bytes
Level-0
Level-1
Level-2
Stage 1 Stage 2 Stage 3
Two compactions by Level Style Compaction

23. Level 0 → Level 1 Compaction
• Level 0 → overlapping keys
• Compaction includes all files from L1
• All files from L1 are compacted with L0
• L0 → L1 compaction completion
L1 → L2 compaction start
• Single thread compaction → not good throughput
• Solution : Making the size of L0 similar to size of L1
Tricky Compaction

24. 2. Universal Style Compaction
• For write-heavy workloads
→ Level Style Compaction may be bottlenecked on
disk throughput
• Stores all files in L0
• All files are arranged in time order
• Temporarily increase size amplification by a factor of
two
• Intended to decrease write amplification
• But, increase space amplification

25. Universal Style Compaction
① Pick up a few files that are chronologically adjacent to one
another
② Merge them
③ Replace them with a new file in level 0
Compaction process

26. Universal Style Compaction
• size_ratio
- Percentage flexibility while comparing file size
- Default : 1
• min_merge_width
- The minimum number of files in a single compaction
- Default : 2
• max_merge_width
- The maximum number of files in a single compaction
- Default : UINT_MAX
Compaction options

27. Universal Style Compaction
Compaction example
5 bytes
6 bytes
10 bytes 10 bytes
Stage 1 Stage 2
Single compaction by Universal Style Compaction
Level-0