The Rredis Package: Bryan W. Lewis and Contributions From Many Others July 5, 2015
The Rredis Package: Bryan W. Lewis and Contributions From Many Others July 5, 2015
Bryan W. Lewis
[email protected]
and contributions from many others
July 5, 2015
1 Introduction
The rredis package provides a native R interface to Redis. Redis is an in memory key/value
database with many innovative features written by Salvatore Sanfilippo1 . It supports data per-
sistence, networked client/server operation, command pipelining, structured value types, server
replication, data expiration, clustering, multicast-like publish/subscribe, and it’s very fast.
The following simple example illustrates a typical use of the rredis package:
> library("rredis")
> redisConnect()
> redisSet("x",rnorm(5))
[1] "OK"
> redisGet("x")
[1] 0.808448325 0.341482747 -0.728739322 -0.105507214 -0.002349064
The key name “x” is associated with the R vector produced by rnorm(5) and stored in Redis. Note
that the R object associated with “x” is available to other rredis clients, and indeed to any Redis
client that can deserialize R objects. Neither the Redis server nor the rredis clients need reside on
the machine on which the result was generated. Depending on the Redis server settings, “x” can
be persistent–that is the value and its association with “x” will persist even if the Redis server is
terminated and re-started.
Values in Redis are classified by type. Value types are perhaps the most distinguishing feature of
Redis.
The canonical string type holds general-purpose objects, for example any serializable R object,
text, or arbitrary binary data.
1
https://github.com/antirez
The rredis Package
The list type represents lists of Redis string objects, ordered by insertion order. Data can be
accessed from lists with stack-like PUSH and POP operations, or by directly indexing ranges
of elements. Importantly, redis lists support atomic blocking and asynchronous operation.
Redis sets are unordered collections of unique Redis strings equipped with typical set opera-
tions like unions and intersections. Uniqueness is enforced by Redis at insertion-time. Redis
also supports operations on ordered sets, with set commands prefixed by “Z.”
Redis hashes are collections of Redis strings indexed by a hashed set of secondary keys.
Expiration intervals or absolute expiration times may be set on any Redis value. The Redis server
can handle lots of small transactions with aplomb, easily exceeding 50,000 transactions/second even
on very limited hardware2 .
2 Supported Platforms
The Redis server is written in ANSI C and supported on most POSIX systems including GNU/Linux,
Solaris, *BSD, and Mac OS X. Microsoft supports a fork of Redis for Windows available here:
https://github.com/MSOpenTech/redis.
The rredis package for R is supported on all supported R platforms, including Microsoft Windows,
and can connect to a Redis server running on a supported platform.
Redis clients are available for lots of languages other than R, including Java, C, C#, Ruby, Python,
PHP, Tcl, Perl, Erlang, Clojure, JavaScript, Scala, and more...
Redis is an open-source project available from http://redis.io, with source code available from
Github at http://github.com/antirez/redis. Redis is also available as an installable package for
most modern GNU/Linux operating systems.
The Redis server is completely configured by the file redis.conf. In order to run the Redis server
as a background process, edit this file and change the line:
daemonize no
to:
daemonize yes
Most default Redis configuration files are set to disconnect connected clients after an inactivity time
out interval. It’s possible to disable that behavior in the redis.conf file with:
2
Redis easily exceeds 100,000 transactions/second on typical high-end workstations
2
The rredis Package
timeout 0
You may wish to peruse the rest of the configuration file and experiment with the other server
settings as well. Finally, start up the Redis server with
redis-server ./redis.conf
Note that some packaged versions of Redis on GNU/Linux set the Redis server to start on boot as
a service.
Redis strings represent the canonical value type. They are used to store any R object that can be
serialized to a bit-stream. Most R objects are serializable. Notable exceptions include objects with
open connections and external reference pointers.
We assume from now on that the rredis package is loaded in the running R session using either
require("rredis")
or
library("rredis")
3
The rredis Package
> print(x)
[1] -0.3297596 1.0417431 -1.3216719 -0.8186305 -0.2705817
> redisSet("x",x)
[1] "OK"
> y <- redisGet("x")
> print(y)
[1] -0.3297596 1.0417431 -1.3216719 -0.8186305 -0.2705817
> all.equal(x,y)
[1] TRUE
> redisGet("z")
NULL
Note that one must explicitly specify a key name (“x” in the above example) and that Redis key
names need not correspond to R variable names.
The SET/GET operations are atomic–that is, multiple SET and or GET operations are guaranteed
not to simultaneously occur. And redisGet always returns immediately, even if a value is not
available in which case it returns NULL (see the example).
The true power of Redis becomes apparent when we share values across multiple clients. For
example, start up a new R session and try:
> library("rredis")
> redisConnect()
> y <- redisGet("x")
> print(y)
[1] -0.3297596 1.0417431 -1.3216719 -0.8186305 -0.2705817
The default behavior of Redis is to make the database persistent, so the value associated with
“x” in the above examples will last until it is overwritten or explicitly removed, even if the Redis
server is re-started. One may immediately purge Redis of all key/value pairs with the (dangerous)
redisFlushAll command.
Redis supports multiple distinct key workspaces, indexed by number. Access may be switched
between workspaces with the redisSelect function as illustrated below. We also use redisKeys
to list all key names in the current workspace.
> redisKeys()
[[1]]
[1] "x"
> redisSelect(1)
[1] "OK"
> redisKeys()
NULL
4
The rredis Package
redisSelect(0)
> redisKeys()
[[1]]
[1] "x"
The number of available workspaces is user-configurable in the redis.conf file (the default is 16).
Note also that index values in Redis begin with 0.
One may easily store and retrieve multiple objects in one operation with redisMSet and redis-
MGet. The example also illustrates how values may be expired (in this case, after one second) with
redisExpire.
> redisMSet(list(x=pi,y=runif(5),z=sqrt(2)))
[1] "OK"
> redisMGet(c("x","y","z"))
$x
[1] 3.141593
$y
[1] 0.85396951 0.80191589 0.21750311 0.02535608 0.11929247
$z
[1] 1.414214
> redisExpire("z",1)
[1] 1
> Sys.sleep(1)
> redisGet("z")
NULL
Redis is a particularly convenient system for sharing data between many different applications.
We illustrate cross-application communication with simple examples using R and the redis-cli
command-line program that is included with the Redis server.
Store a sample value in the Redis database with the redis-cli program from the command line
as follows:
redis-cli set shell "Greetings, R client!"
OK
5
The rredis Package
And, voilà, R and shell communicate text through Redis. The returned character string value is
decorated with an attribute named ”redis string value” to indicate that the original value in Redis
was not a serialized R object.
When R values with the ”redis string value” attribute are uploaded to Redis, R serialization is
bypassed and the value is uploaded as raw binary data. This mechanism was introduced to preserve
the original nature of Redis values through download/upload sequences. For example:
> x <- redisGet("shell")
> redisSet("hello", x)
We can now use the redis-cli client to recover the value inserted by R from the new key:
redis-cli get hello
"Greetings, R client!"
In general, however, R objects are serialized by default before uploading to Redis. Consider the
following example:
> redisSet("R", "Greetings, shell client!")
This example produces undesirable results because the default behavior of the R redisSet command
is to store data as R objects, which the shell client cannot decipher. Instead, we must encode the
R object (in this case, a character string) in a format that shell can understand:
> redisSet("R", charToRaw("Greetings, shell client!"))
[1] "OK"
It can be tricky to share arbitrary R objects with other languages, but raw character strings usually
provide a reasonable, if sometimes inefficient, common tongue.
The raw=TRUE option may be set on most package functions that receive data, for example redisGet.
Use the RAW option to leave the message data as is (otherwise the functions try to deserialize it to
a standard R object). The RAW format is useful for binary exchange of data with programs other
than R.
6
The rredis Package
Redis list value types provide us with a remarkably powerful and rich set of operations. Redis lists
may be used to set up data queues and they may be accessed either synchronously or asynchronously.
We walk through basic Redis list operation in the first example below. The example shows how
redisLPush pushes values onto a list from the left, and redisRPush pushes values from the right.
> redisLPush("a",1)
[1] 1
> redisLPush("a",2)
[1] 2
> redisLPush("a",3)
[1] 3
> redisLRange("a",0,2)
[[1]]
[1] 3
[[2]]
[1] 2
[[3]]
[1] 1
> redisLPop("a")
[1] 3
> redisLRange("a",0,-1)
[[1]]
[1] 2
[[2]]
[1] 1
> redisRPush("a","A")
[1] 3
> redisRPush("a","B")
[1] 4
> redisLRange("a",0,-1)
[[1]]
[1] 2
[[2]]
[1] 1
[[3]]
[1] "A"
[[4]]
[1] "B"
7
The rredis Package
> redisRPop("a")
[1] "B"
Like the redisGet function, redisLPop and redisRPop always return immediately, even when no
value is available in which case they return NULL. Redis includes a blocking variant of the list “Pop”
commands that is illustrated in the next example.
> redisBLPop("b",timeout=1)
NULL
> redisLPush("b",runif(5))
[1] 1
> redisBLPop("b",timeout=1)
$b
[1] 0.3423658 0.4188430 0.2494071 0.9960606 0.5643137
In the first case above, the NULL value is returned after a one-second timeout because no value
was immediately available in the list. Once populated with data, the second attempt consumes the
list value and returns immediately.
We can also block on multiple lists, returning when data is available on at least one of the lists:
> redisFlushAll()
[1] "OK"
> redisLPush("b",5)
[1] 1
> redisBLPop(c("a","b","c"))
$b
[1] 5
Although blocking list operations seem simple, they provide an extraordinarily powerful environment
for coordinating events between multiple R (and other client) processes. The following example
illustrates a simple event stream in which data is emitted periodically by a shell script, and consumed
and processed as events arrive by an R process.
First, open an R window and block on the “a” and “b” lists:
> redisFlushAll()
> for (j in 1:5) {
+ x <- redisBLPop(c("a","b"))
+ print (x)
+ }
8
The rredis Package
Now, open a terminal window and navigate to the directory that contains the redis-cli program.
Run (the following may all be typed on one line):
for x in 1 2 3 4 5;do sleep $x;
if test $x == "2";
then ./redis-cli lpush a $x;
else ./redis-cli lpush b $x;
fi;
done
And now you will see your R session processing the events as they are generated by the shell script:
$b
[1] "1"
$a
[1] "2"
$b
[1] "3"
$b
[1] "4"
$b
[1] "5"
Now, imagine that events may be processed independently, and that they occur at an extraordinary
rate–a rate too fast for R to keep up. The solution in this case is simple, start up another R process
and it will handle events as they come in, relieving the first R process of about half the event load.
Still not enough, start up another, etc.
Keeping in mind that the R clients can run on different computers, we realize that this simple
example can easily lead to a very scalable parallel event processing system that requires very little
programming effort!
The Redis set value type operates somewhat like Redis lists, but only allowing unique values within
a set. Sets also come equipped with the expected set operations, as illustrated in the following
example.
> redisSAdd("A",runif(2))
[1] 1
9
The rredis Package
> redisSAdd("A",55)
[1] 1
> redisSAdd("B",55)
[1] 1
> redisSAdd("B",rnorm(3))
[1] 1
> redisSCard("A")
[1] 2
> redisSDiff(c("A","B"))
[[1]]
[1] 0.5449955 0.7848509
> redisSInter(c("A","B"))
[[1]]
[1] 55
> redisSUnion(c("A","B"))
[[1]]
[1] 55
[[2]]
[1] 0.5449955 0.7848509
[[3]]
[1] -1.3153612 0.9943198 -0.3725513
Redis changes and sometimes new Redis commands are introduced. Thanks to the elegant design
of the underlying Redis protocol, the rredis package can support any Redis command, even those
not yet explicitly wrapped by R functions in the package.
Use the low-level redisCmd(CMD, ...) function to perform any Redis operation. The CMD argument
must be a character string that represents a valid Redis command, see for example http://redis.io/commands.
Replace the ... with any additional arguments specific to the Redis command. Arguments that
are not already in raw format will be converted to raw byte format, and non-character values will
be serialized as R objects (character strings will be left as unserialized ASCII by default).
Here is a simple example that emulates the redisSet and redisGet functions:
> redisCmd("set","x",runif(5))
10
The rredis Package
[1] "OK"
> redisCmd("get","x")
[1] 0.0006776408 0.5652672122 0.0985793471 0.5748420910 0.4098101135
That performance seems terrible. After all, isn’t Redis capable of tens of thousands of transactions
per second?
There are at least three approaches to improving the performance of the last example: Redis pipelin-
ing (the best way), enable TCP NODELAY, or use Wush Wu’s alternate hiredis-based support
package for rredis or an alternative hiredis-based R client package.
The rredis package supports Redis pipelining. Quoting from the Wiki page http://redis.io/commands:
11
The rredis Package
Redis pipelining is enabled in the rredis package with the redisSetPipeline(TRUE) function. Once
enabled, results are not immediately returned to the R client but instead cached at the Redis server
until explicitly requested with the redisGetResponse function.
Here is our example with pipelining enabled:
> library("rredis")
> redisConnect()
> redisSetPipeline(TRUE)
> t1 <- proc.time()
> for(j in 1:100) redisSet("x", j)
> resp <- redisGetResponse()
> print(proc.time() - t1)
Enabling the TCP NODELAY option is one performance approach that hiredis (the official Redis
C library client) takes: see https://github.com/redis/hiredis. It disables the TCP Nagle congestion
control algorithm and can flood your network with lots of Redis traffic resulting in heartburn for
your system administrators. With that caveat, we recently added the ability to support this in the
rredis package using standard R connections. Here is an example:
> library("rredis")
> redisConnect(nodelay=TRUE)
> t1 <- proc.time()
> for(j in 1:100) redisSet("x", j)
> print(proc.time() - t1)
We see that enabling TCP NODELAY gives similar performance to the pipelined approach (at least
on a wimpy laptop) without the added inconvenience of requiring getResponse.
12
The rredis Package
Wush Wu and Dirk Eddelbuettel have an R client for Redis based on the hiredis C library that gener-
ally gives better performance than the rredis package a the expense of portability and disabling Nagle
all the time. Wush is also working on a pluggable hiredis support package for the rredis package that
should be available soon. Here are some links of interest: https://github.com/eddelbuettel/rhiredis
https://github.com/wush978/rredis
5 Transactions
Redis supports batch submission of multiple Redis operations. Aggregating operations with trans-
actions can in many cases significantly increase performance. The following description is adapted
from the Redis documentation at http://redis.io:
The redisMulti, redisExec, redisDiscard and redisWatch form the foundation of transactions
in Redis. They allow the execution of a group of commands in a single step, with two important
guarantees:
1. All the commands in a transaction are serialized and executed sequentially. It can never
happen that a request issued by another client is served in the middle of the execution of
a Redis transaction. This guarantees that the commands are executed as a single atomic
operation.
2. Either all of the commands or none are processed. The redisExec command triggers the
execution of all the commands in the transaction, so if a client loses the connection to the
server in the context of a transaction before calling the redisMulti command none of the
operations are performed, instead if the redisExec command is called, all the operations are
performed. When using the append-only file Redis makes sure to use a single write(2) syscall
to write the transaction on disk. However if the Redis server crashes or is killed by the system
administrator in some hard way it is possible that only a partial number of operations are
registered. Redis will detect this condition at restart, and will exit with an error. Using
the redis-check-aof tool it is possible to fix the append only file that will remove the partial
transaction so that the server can start again.
Queued Redis commands may be discarded with the redisDiscard function. Upon successful
execution of a transaction, the results from all of the queued commands are returned as a list.
The redisWatch function provides a check and set style conditional transaction. Use redisWatch
to monitor any number of Redis keys. If any watched key values change prior to calling redisExec
the entire queued sequence will be discarded. Conditioning transactions with redisWatch is quite
useful in multi-client asynchronous settings.
13
The rredis Package
The following extremely basic example illustrates transactions conditional on no change in value
corresponding to the “z” key:
> redisWatch("z")
[1] "OK"
> redisMulti()
[1] "OK"
> redisSet("x",runif(3))
[1] "QUEUED"
> redisSet("y",pi)
[1] "QUEUED"
> redisGet("x")
[1] "QUEUED"
> redisExec()
[[1]]
[1] "OK"
[[2]]
[1] "OK"
[[3]]
[1] 0.7620601 0.5982853 0.8274721
6 Publish/Subscribe
The publish/subscribe functions let Redis clients reliably multicast (publish) messages over “chan-
nels” that any client may subscribe to. Channels are identified by name (character string). Use the
redisSubscribe function to subscribe to one or more channels. Use the redisPublish function
to transmit messages over a channel. Once subscribed, channels must be monitored for incom-
ing messages using the redisGetResponse function, usually in an event loop. Beware that the
redisGetResponse function indefinitely blocks for an incoming message on subscribed channels.
Here is a simple example:
> redisSubscribe("channel1")
14
The rredis Package
[[2]]
[1] "channel1"
[[3]]
[1] "message3"
Note that the only Redis functions that may be used in between the redisSubscribe and re-
disUnsubscribe functions are redisGetResponse, redisSubscribe, and redisMonitorChannels
functions.
7 Miscellaneous
We cover a few miscellaneous features of the package here.
Redis may be configured with renamed commands. This is one of the more obscure features of
Redis. For example, suppose that we desire to rename the Redis commands ”SET” and ”GET” as
”NEWSET” and ”NETGET,” respectively. That can be achieved by adding the following two lines
to your redis.conf file:
rename-command SET NEWSET
rename-command GET NEWGET
Note that after restarting Redis with this configuration, the redisSet and redisGet won’t work
anymore since those commands have been renamed.
In order to use the new names in the rredis package, create an R list with the replacement names
as follows and place that list in a special environment used by the rredis package used to store state
associated with open Redis connections:
> assign("rename", list(SET="NEWSET", GET="NEWGET"), envir=rredis:::.redisEnv)
And then the normal redisSet and redisGet functions work again! Similarly rename any other
renamed Redis commands. Note that this renaming strategy applies to the current running R
session (it applies to all open Redis connections in that session).
15
The rredis Package
7.2 Authentication
Redis supports a really basic password authentication scheme. If you use this, it’s probably a good
idea to use stunnel or an SSH tunnel or something similar to encrypt all network traffic between
Redis clients and Redis. Otherwise passwords are transmitted in clear text over the network.
Use the redisAuth(password) function to authenticate with Redis, or optionally supply the password
argument to the redisConnect function.
16