Skip to content


Repository files navigation

DRedis: Disk-based Redis implementation

Redis is a great key-value database and it's extremely fast because it's in-memory. Some people want Redis's rich data types without having to worry about the memory limitations. For those that can afford slower performance and want unlimited storage, DRedis may be an alternative.

WARNING: This project is still experimental and it doesn't implement all Redis commands!


Make sure to install the LevelDB C++ library (apt-get install libleveldb-dev or brew install leveldb) and then run:

$ pip install dredis

Note: The LMDB backend doesn't require external dependencies.


$ dredis --dir /tmp/dredis-data

To know about all of the options, use --help:

$ dredis --help
usage: dredis [-h] [-v] [--host HOST] [--port PORT] [--dir DIR]
              [--backend {lmdb,leveldb,memory}]
              [--backend-option BACKEND_OPTION] [--rdb RDB] [--debug]
              [--flushall] [--readonly] [--requirepass REQUIREPASS]
              [--gc-interval GC_INTERVAL] [--gc-batch-size GC_BATCH_SIZE]

optional arguments:
  -h, --help            show this help message and exit
  -v, --version         show program's version number and exit
  --host HOST           server host (defaults to
  --port PORT           server port (defaults to 6377)
  --dir DIR             directory to save data (defaults to a temporary
  --backend {lmdb,leveldb,memory}
                        key/value database backend (defaults to leveldb)
  --backend-option BACKEND_OPTION
                        database backend options (e.g., --backend-option
  --rdb RDB             RDB file to seed dredis
  --debug               enable debug logs
  --flushall            run FLUSHALL on startup
  --readonly            accept read-only commands
  --requirepass REQUIREPASS
                        require clients to issue AUTH <password> before
                        processing any other commands
  --gc-interval GC_INTERVAL
                        key gc interval in milliseconds (defaults to 500)
  --gc-batch-size GC_BATCH_SIZE
                        key gc batch size (defaults to 10000)

Running dredis with Docker locally (port 6377 on the host):

$ docker-compose up

Asynchronous Deletions (Key Garbage Collection)

This is a new and experimental feature that was necessary at Yipit to offload dredis on RENAME and DEL operations. When a sorted set, hash, or set is stored using any of the backends, dredis creates a few keys in the backend for each element of those collections. With this new feature, on RENAME and DELETE, only the key holding the key ID/pointer is replaced/deleted. There's a background thread that periodically deletes those related keys from the storage backend (you can tweak the GC options via --gc-interval and --gc-batch-size). Despite the asynchronous deletions, the deletion operation still ensures strong consistency because of the use of key IDs/pointers.

If you don't want this experimental feature, you need to go back to DRedis 2.5.3.


There's support for LevelDB, LMDB, and an experimental memory backend. All backend options should be passed in the command line as --backend-option NAME1=value1 --backend-option NAME2=value2 (the values must be JSON-compatible).


LevelDB is the easiest persistent backend because it doesn't require any option tweaking to get it to work reliably.


We use plyvel as the LevelDB backend. All available options are parameters of plyvel.DB.

The current default options for LevelDB are:

  • name: The same value as the --dir option
  • create_if_missing: True


The performance of LMDB can be better than LevelDB and we're considering making it the default backend in the future.


We use py-lmdb as the LMDB backend. All available options are parameters of lmdb.Environment. We recommend that you think ahead and change the map_size parameter according to your needs — this is the maximum size of the LMDB database file on disk.

The current default options for LMDB are:

  • path: The same value as the --dir option
  • map_size: 1073741824 (1GB)
  • map_async: True
  • writemap: True
  • readahead: False
  • metasync: False


This is experimental and doesn't persist to disk. It was created to have a baseline to compare persistent backends.



Supported Commands

Command signature Type
CONFIG GET parameter Server
CONFIG SET parameter value Server
SAVE Server
DEL key [key ...] Keys
DUMP key Keys
EXISTS key [key ...] Keys
EXPIRE key ttl** Keys
KEYS pattern Keys
RENAME key newkey Keys
RESTORE key ttl serialized-value [REPLACE]*** Keys
TTL key Keys
TYPE key Keys
AUTH Connection
PING [msg] Connection
SELECT db Connection
GET key Strings
GETRANGE key start end Strings
INCR key Strings
INCRBY key increment Strings
SET key value Strings
SADD key value [value ..] Sets
SCARD key Sets
SISMEMBER key value Sets
EVAL script numkeys [key ...] [arg ...] Scripting
ZADD key [NX|XX] score member [score member ...] Sorted Sets
ZCARD key Sorted Sets
ZCOUNT key min_score max_score Sorted Sets
ZRANGE key start top [WITHSCORES] Sorted Sets
ZRANGEBYSCORE key min_score max_score [WITHSCORES] [LIMIT offset count] Sorted Sets
ZRANK key member Sorted Sets
ZREM key member [member ...] Sorted Sets
ZSCAN key cursor [MATCH pattern] [COUNT count] Sorted Sets
ZSCORE key member Sorted Sets
ZUNIONSTORE destination numkeys key [key ...] [WEIGHTS weight [weight ...]] Sorted Sets
HDEL key field [field ...] Hashes
HGET key value Hashes
HGETALL key Hashes
HINCRBY key field increment Hashes
HKEYS key Hashes
HLEN key Hashes
HSET key field value [field value ...] Hashes
HSETNX key field value Hashes
HVALS value Hashes


  • *COMMAND's reply is incompatible at the moment, it returns a flat array with command names (their arity, flags, positions, or step count are not returned).
  • **EXPIRE doesn't set key expiration yet, it's a no-op command
  • ***RESTORE doesn't work with Redis strings compressed with LZF or encoded as OBJ_ENCODING_INT; also doesn't work with sets encoded as OBJ_ENCODING_INTSET, nor hashes and sorted sets encoded as OBJ_ENCODING_ZIPLIST.
  • CONFIG GET, CONFIG HELP, and CONFIG SET are specific to dredis. The commands' signature and behavior are equivalent to the ones in Redis

How is DRedis implemented

Initially DRedis had its own filesystem structure, but then it was converted to use LevelDB, which is a lot more reliable and faster (nowadays there's also the LMDB backend).

Other projects implement similar features to what's available on DRedis, but they aren't what Yipit needed when the project started. Some of them rely on multiple threads and compromise on consistency, don't implement Lua scripts, or don't implement sorted sets correctly. We ran the DRedis tests against a few solutions and they failed (which means they're not fully compatible).

Similar projects:

Lua support

Lua is supported through the lupa library.


Data Consistency

We rely on the backends' consistency properties and we use batches/transactions to stay consistent. Tweaking the backend options may impair consistency (e.g., sync=false for LMDB).

Cluster mode & Replication

Replication, key distribution, and cluster mode are not supported. If you want higher availability you can create multiple servers that share or replicate a disk (consistency may suffer when replicating). Use DNS routing or a network load balancer to route requests properly.


The command SAVE creates a snapshot in the same format as Redis's RDB version 7 (compatible with Redis 3.x). We recommend you to run SAVE on a secondary dredis process, otherwise the server will hang during the snapshot (consistency guarantees are higher with LMDB as the backend). The command BGSAVE may be supported in the future.

Other backups solutions involve backing up the files created by the backend. A straightforward approach is to have periodic backups to an object storage such as Amazon S3 orr use a block storage solution and perform periodic backups (e.g., AWS EBS).

If you use SAVE from a secondary process or backup the data directory, there shouldn't be any significant impact on the main server.

Why Python

Because it's a good language to get things off the ground quickly and everybody at Yipit knows it well. If this becomes a valuable project, other languages will be evaluated — the language of choice won't affect much of the I/O bottleneck, but it may bring good performance benefits. We're experimenting with Cython to get better performance without having to rewrite large chunks in C.

The project will migrate to Python 3 soon.

Didn't you have better names?

@andrewgross suggested REDISK. The name will only matter if this project survives, it's still an experiment. Also, other projects use the name redisk.

@nadlerjessie suggested we pronounce dredis as "Doctor Redis".

Releasing dredis

  1. Make sure you have all important changes in the top section of
  2. Make sure your PyPI credentials are correct in ~/.pypirc
  3. Run make release
  4. Enter the new version (e.g., 1.0.0)


Disk-based Redis implementation







No packages published