中文帮助请点这里, click here for Chinese documents

RedRock_old (depreciated) Homepage

Introduction

RedRock is a combination of Redis and RocksDB.

Why

Redis is a wonderful NoSQL based on memory. But memory costs more than disk. We hope something,

Same fast and strong as Redis
Extend to disk (HDD or SSD)

As SSD is becoming cheaper and has good performance, it is why I, enigneer Stone, code RedRock.

I wish one stone to hit two birds.

NOTE: I only code the project for a couple of months and only use my off-busy-work time. It is not as mature as a project for ten years. Wish you use it more and leave feedbacks at Github. Thank you.

Key Design

RedRock stores all Redis keys in memory, and can store ONLY the keys.

If memory is not enough for holding all key/values, RedRock will save some values to disk and later retrieve the values from the disk if clients need them. Usually the size of value is bigger than key. If value size is 100 times than key, the dataset can be 100 times than the limit of memory. If more key/values are coming over the 100 times limit, RedRock will evict some keys totally by LRU/LFU.

For the performance, if most requests visit key/value in memory, RedRock is same as Redis, usually hundreds kilo qps with 99% latnecy of sub 1ms for one machine. If more visits are coming to the keys which values are in disk, the performance can degrade rapidlly. You can check the benchmark for different patterns.

So the best user case for RedRock is hot key/value fitting in memory with a huge warm/cold values in disk. This way, we achieve the performance of latency similar to Redis and much-over-meomroy-size-limit dataset like traditional DB.

RedRock uses the Redis persistent ways like AOF/RDB. Because the dataset would be much bigger than memory, you need some advice for the persistency.

RedRock Features

Pure Redis. When no enable disk, RedRock is running almost the same codes as Redis
Use Redis Protocol, so no need to adjust one line of your client codes or config files
Keep all keys in memory but only hot key's value in memory with most other values in disk
Your dataset volume can be one hundred times of your memory
Support all Redis commands except module's commands
Support expire and eviction, you can use it as Cache
Support all Redis data structures, including String,List,Set,Hash,SortedSet,Stream,HyperLogLog,Geo
Support persistence, including RDB and AOF, in fork() child process or main thread
Can config maximum memory
Policy for dumping value to disk supports LRU and LFU
If use eviction, you can specify LRU or LFU, and the max keys with value in memory
Support Replication: Leader/Follower(i.e. Master/Slave) replica
I think it would support Redis Sentinel
I think it would support Reids Cluster. So no need for Twemproxy and Codis
I think it would support distributed lock, RedLock, but I think Martin Kleppmann is right
Support Redis Pipeline
Support Redis Transaction, including WATCH command
Support Redis Blocking
Support Script(LUA)
Support Redis Subscribe/Publish
Support Redis Stream
Support original stats for Redis including SlowLog, plus RedRock specific stats
Good for mass intensive writing to disk. For regular SSD, you can write new data at speed 20M/s
Data is compressed in disk, you only need disk volume from 10% to half of your dataset
When main load is for hot keys, performance is almost the same as Pure Redis, i.e. Million rps for one node
When coming to avarage random key visit pattern, performance maybe degrade by 1 maganitude order，I guess near Hundred Kilo rps for one node
Using Rocksdb library, but overload of memory in db engine is limited to tens MB
Keep the main logic in main thread as Redis, low level for thread switch and race risks.
Key/value length limit is same as Redis
Support multi database and can change the number