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 dredisNote: The LMDB backend doesn't require external dependencies.
$ dredis --dir /tmp/dredis-dataTo 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 127.0.0.1)
--port PORT server port (defaults to 6377)
--dir DIR directory to save data (defaults to a temporary
directory)
--backend {lmdb,leveldb,memory}
key/value database backend (defaults to leveldb)
--backend-option BACKEND_OPTION
database backend options (e.g., --backend-option
map_size=BYTES)
--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 upThis 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--diroptioncreate_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--diroptionmap_size:1073741824(1GB)map_async:Truewritemap:Truereadahead:Falsemetasync:False
This is experimental and doesn't persist to disk. It was created to have a baseline to compare persistent backends.
None.
| Command signature | Type |
|---|---|
| COMMAND* | Server |
| CONFIG GET parameter | Server |
| CONFIG HELP | Server |
| CONFIG SET parameter value | Server |
| DBSIZE | Server |
| FLUSHALL | Server |
| FLUSHDB | 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 |
| SMEMBERS key | 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 |
Footnotes:
- *
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). - **
EXPIREdoesn't set key expiration yet, it's a no-op command - ***
RESTOREdoesn't work with Redis strings compressed with LZF or encoded asOBJ_ENCODING_INT; also doesn't work with sets encoded asOBJ_ENCODING_INTSET, nor hashes and sorted sets encoded asOBJ_ENCODING_ZIPLIST. CONFIG GET,CONFIG HELP, andCONFIG SETare specific to dredis. The commands' signature and behavior are equivalent to the ones in Redis
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:
- https://github.com/yinqiwen/ardb
ardbseems to be the most similar in scope and a good candidate for contributions or a fork. Their sorted sets implementation has a bug with negative scores. It's a large C++ project with lots of features.
- https://github.com/Qihoo360/pika
- no Lua support. This is a large C++ project and its documentation is in Chinese. The project seems to be stable and is used by large Chinese companies.
- https://github.com/KernelMaker/blackwidow
- a C++ library, not a Redis-like server
- https://github.com/siddontang/ledisdb
- Similar to Redis but different commands
- https://github.com/reborndb/qdb
- No Lua support and no longer maintained
- https://github.com/alash3al/redix
- No Lua and no sorted set support
- https://github.com/meitu/titan
- No Lua support and not enough sorted set support
Lua is supported through the lupa library.
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).
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.
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.
@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".
- Make sure you have all important changes in the top section of
CHANGELOG.md - Make sure your PyPI credentials are correct in
~/.pypirc - Run
make release - Enter the new version (e.g.,
1.0.0)