A tiny redis client for script boys with high performance. Optimized especially for massive insertion.
pip install redisio
TLDR: The following is the document but don't read it. Instead read the code. It's much shorter.
import redisio
rd = redisio.Redis(host='127.0.0.1', port=6379, db=0, password='')
The arguments above are set as default so can be omitted.
Connecting via unix sockets is also available.
import redisio
rd = redisio.Redis(socket='/tmp/redis.sock')
See the list of Redis commands at redis.io.
Since redisio
is designed to be implemented strictly in protocol with little syntax sugar on calling but not any modification on data itself, any future commands of redis
can be properly supported without update.
Instance redisio.Redis
is callable.
Any direct calling on it sends the commands (either single command or multiple in list) to the server and then return the client instance itself immediately without reading replies in order to be called in chain conveniently:
assert rd == rd('SET', 'x', 3)('GET', 'x')(['SET', 'x', 3], ['GET', 'x'])
Method redisio.Redis.next
returns the first reply in the message queue from the server:
assert 'OK' == rd('SET', 'x', 3)('GET', 'x').next()
assert '3' == next(rd)
Note: It will be blocked to call next
when the queue is empty.
Reply in the queue can be reached by the index:
assert '3' == rd('SET', 'x', 3)('GET', 'x')('SET', 'x', 4)[-2]
Note: It will first iterate the whole queue and then return the specific reply, with of course a side-effect to empty the queue.
Each redis command is mapped to a method with the same name.
Calling it in this method-like way will send the command, then read all the replies, and return the last one.
assert '3' == rd('SET', 'x', 3).get('x')
Note: It will be blocked to call rd.shutdown()
because this command will never be answered (Dead Men Tell No Tales). rd('SHUTDOWN')
should be used.
Instance redisio.Redis
is iterable.
So iterating it to get all the replies.
r, = rd("HGET", key, field)
r1, r2= rd("HGET", key, field)("HGET", key, field2)
r1, r2= list(rd(["GET", key], ("GET", key2))(["SET", "X", "Y"]))[:2]
If you want to insert a large amount of data into redis without the care of the results, you can close the connection after sending it by the use of del
to avoid parsing the replies.
rd(*large_scale_of_cmds).__del__()
Benefit from this the massive insertion is blazingly fast: sending a million of HSET cost only 5.355 seconds via redisio
while it costs 23.918 seconds via redis-py
.
Note: Replies are buffered on the server if the client have not read them while the connection keeps alive. This will eventually make the server crash because of the increasing occupied memory. So be aware.
redisio
will automatically reset the connection before sending a command in the method-like way while there are more than 1024 replies to read.
rd(*large_scale_of_cmds).dbsize()
rd(*large_scale_of_cmds)('DBSIZE')[-1]
The former is usually faster than the latter because no massive replies need to be read and parsed.
rd.monitor()
# rd.subscribe('channel')
while 1:
print next(rd)
Q: Why not using redis-py but redisio?
A: To accomplish the majority of tasks redis-py is recommended and mostly used even by myself. Frankly speaking You may never need redisio. redisio is written and optimized especially for massive insertion to be memory-saving and much faster.
Q: How to use redisio in bash?
A: redis-cli is available in bash.
Q: How to use redisio with thread safety?
A: redis-py is thread safe.
Q: Why the result of hgetall is not a dict but a list?
A: This is the original format of replies from redis-server. Once you get used to this original style, you will be able to process results fluently from redis-cli or Lua Script without the mess of confusing types or structures brought by other brilliant libraries. Anyway let's get down to brass tacks. You can get your dict like this:
hash_values = rd.hgetall('a_hash_key')
hash_dict = dict(zip(hash_values[::2], hash_values[1::2]))