Join 10350+ others. No spamming.
I promise!

Follow us at github.



ptaoussanis / carmine


Redis client and message queue for Clojure


API docs | CHANGELOG | other Clojure libs | Twitter | contact/contrib | current Break Version:

[com.taoensso/carmine "2.12.0"] ; See CHANGELOG for details

Carmine, a Clojure Redis client & message queue

Redis is awesome and it's getting more awesome every day. It deserves a great Clojure client.

Aren't there already a bunch of clients?

Plenty: there's redis-clojure, clj-redis based on Jedis, Accession, and (the newest) labs-redis-clojure. Each has its strengths but these strengths often fail to overlap, leaving one with no easy answer to an obvious question: which one should you use?

Carmine is an attempt to cohesively bring together the best bits from each client. And by bringing together the work of others I'm hoping to encourage more folks to pool their efforts and get behind one banner. (Rah-rah and all that).

What's in the box™?

  • Small, uncomplicated all-Clojure library.
  • Fully documented, up-to-date API, with full support for the latest Redis versions.
  • Great performance.
  • Industrial strength connection pooling.
  • Composable, first-class command functions.
  • Flexible, high-performance binary-safe serialization using Nippy.
  • Full support for Lua scripting, Pub/Sub, etc.
  • Full support for custom reply parsing.
  • Command helpers (atomic, lua, sort*, etc.).
  • Ring session-store.
  • Simple, high-performance message queue (Redis 2.6+, stable v2+).
  • Simple, high-performance distributed lock (Redis 2.6+, stable v2+).
  • Pluggable compression and encryption support. (v2+)
  • Includes Tundra, an API for replicating data to an additional datastore. (Redis 2.6+, v2+)

3rd-party tools, etc.

Getting started


Add the necessary dependency to your Leiningen project.clj and require the library in your ns:

[com.taoensso/carmine "2.12.0"] ; project.clj
(ns my-app (:require [taoensso.carmine :as car :refer (wcar)])) ; ns


You'll usually want to define a single connection pool, and one connection spec for each of your Redis servers.

(def server1-conn {:pool {<opts>} :spec {<opts>}}) ; See `wcar` docstring for opts
(defmacro wcar* [& body] `(car/wcar server1-conn ~@body))

Basic commands

Sending commands is easy:

(wcar* (car/ping)
       (car/set "foo" "bar")
       (car/get "foo"))
=> ["PONG" "OK" "bar"]

Note that sending multiple commands at once like this will employ pipelining. The replies will be queued server-side and returned all at once as a vector. By default, if a single command is in the body of wcar the value of the command is returned. To always return a vector, add the :as-pipeline argument to wcar.

If the server responds with an error, an exception is thrown:

(wcar* (car/spop "foo"))
=> Exception ERR Operation against a key holding the wrong kind of value

But what if we're pipelining?

(wcar* (car/set  "foo" "bar")
       (car/spop "foo")
       (car/get  "foo"))
=> ["OK" #<Exception ERR Operation against ...> "bar"]


The only value type known to Redis internally is the byte string. But Carmine uses Nippy under the hood and understands all of Clojure's rich datatypes, letting you use them with Redis painlessly:

(wcar* (car/set "clj-key" {:bigint (bigint 31415926535897932384626433832795)
                           :vec    (vec (range 5))
                           :set    #{true false :a :b :c :d}
                           :bytes  (byte-array 5)
                           ;; ...
       (car/get "clj-key"))
=> ["OK" {:bigint 31415926535897932384626433832795N
          :vec    [0 1 2 3 4]
          :set    #{true false :a :c :b :d}
          :bytes  #<byte [] [B@4d66ea88>}]

Types are handled as follows:

  • Clojure strings become Redis strings.
  • Keywords become Redis strings. (v2+)
  • Simple Clojure numbers (integers, longs, floats, doubles) become Redis strings.
  • Everything else gets automatically de/serialized.

You can force automatic de/serialization for an argument of any type by wrapping it with car/serialize.

Documentation and command coverage

Like labs-redis-clojure, Carmine uses the official Redis command reference to generate its own command API. Which means that not only is Carmine's command coverage always complete, but it's also fully documented:

(use 'clojure.repl)
(doc car/sort)
=> "SORT key [BY pattern] [LIMIT offset count] [GET pattern [GET pattern ...]] [ASC|DESC] [ALPHA] [STORE destination]

Sort the elements in a list, set or sorted set.

Available since: 1.0.0.

Time complexity: O(N+M*log(M)) where N is the number of elements in the list or set to sort, and M the number of returned elements. When the elements are not sorted, complexity is currently O(N) as there is a copy step that will be avoided in next releases."

Yeah. Andreas Bielk, you rock.


Redis 2.6 introduced a remarkably powerful feature: server-side Lua scripting! As an example, let's write our own version of the set command:

(defn my-set
  [key value]
  (car/lua "return'set', _:my-key, 'lua '.. _:my-val)"
                  {:my-key key}   ; Named key variables and their values
                  {:my-val value} ; Named non-key variables and their values

(wcar* (my-set "foo" "bar")
       (car/get "foo"))
=> ["OK" "lua bar"]

Script primitives are also provided: eval, eval-sha, eval*, eval-sha*. See the Lua scripting docs for more info.


The lua command above is a good example of a Carmine helper.

Carmine will never surprise you by interfering with the standard Redis command API. But there are times when it might want to offer you a helping hand (if you want it). Compare:

(wcar* (car/zunionstore "dest-key" 3 "zset1" "zset2" "zset3" "WEIGHTS" 2 3 5))
(wcar* (car/zunionstore* "dest-key" ["zset1" "zset2" "zset3"] "WEIGHTS" 2 3 5))

Both of these calls are equivalent but the latter counted the keys for us. zunionstore* is another helper: a slightly more convenient version of a standard command, suffixed with a * to indicate that it's non-standard.

Helpers currently include: atomic, eval*, evalsha*, info*, lua, sort*, zinterstore*, and zunionstore*. See their docstrings for more info.

Commands are (just) functions

In Carmine, Redis commands are real functions. Which means you can use them like real functions:

(wcar* (doall (repeatedly 5 car/ping)))

(let [first-names ["Salvatore"  "Rich"]
      surnames    ["Sanfilippo" "Hickey"]]
  (wcar* (mapv #(car/set %1 %2) first-names surnames)
         (mapv car/get first-names)))
=> ["OK" "OK" "Sanfilippo" "Hickey"]

(wcar* (mapv #(car/set (str "key-" %) (rand-int 10)) (range 3))
       (mapv #(car/get (str "key-" %)) (range 3)))
=> ["OK" "OK" "OK" "OK" "0" "6" "6" "2"]

And since real functions can compose, so can Carmine's. By nesting wcar calls, you can fully control how composition and pipelining interact:

(let [hash-key "awesome-people"]
  (wcar* (car/hmset hash-key "Rich" "Hickey" "Salvatore" "Sanfilippo")
         (mapv (partial car/hget hash-key)
               ;; Execute with own connection & pipeline then return result
               ;; for composition:
               (wcar* (car/hkeys hash-key)))))
=> ["OK" "Sanfilippo" "Hickey"]

Listeners & Pub/Sub

Carmine has a flexible Listener API to support persistent-connection features like monitoring and Redis's fantastic Publish/Subscribe feature:

(def listener
  (car/with-new-pubsub-listener (:spec server1-conn)
    {"foobar" (fn f1 [msg] (println "Channel match: " msg))
     "foo*"   (fn f2 [msg] (println "Pattern match: " msg))}
   (car/subscribe  "foobar" "foobaz")
   (car/psubscribe "foo*")))

Note the map of message handlers. f1 will trigger when a message is published to channel foobar. f2 will trigger when a message is published to foobar, foobaz, foo Abraham Lincoln, etc.

Publish messages:

(wcar* (car/publish "foobar" "Hello to foobar!"))

Which will trigger:

(f1 '("message" "foobar" "Hello to foobar!"))
(f2 '("pmessage" "foo*" "foobar" "Hello to foobar!"))

You can adjust subscriptions and/or handlers:

(with-open-listener listener
  (car/unsubscribe) ; Unsubscribe from every channel (leave patterns alone)
  (car/psubscribe "an-extra-channel"))

(swap! (:state listener) assoc "*extra*" (fn [x] (println "EXTRA: " x)))

Remember to close the listener when you're done with it:

(car/close-listener listener)

Note that subscriptions are connection-local: you can have three different listeners each listening for different messages, using different handlers. This is great stuff.

Reply parsing

Want a little more control over how server replies are parsed? You have all the control you need:

(wcar* (car/ping)
       (car/with-parser clojure.string/lower-case (car/ping) (car/ping))
=> ["PONG" "pong" "pong" "PONG"]

Binary data

Carmine's serializer has no problem handling arbitrary byte[] data. But the serializer involves overhead that may not always be desireable. So for maximum flexibility Carmine gives you automatic, zero-overhead read and write facilities for raw binary data:

(wcar* (car/set "bin-key" (byte-array 50))
       (car/get "bin-key"))
=> ["OK" #<byte[] [B@7c3ab3b4>]

Message queue

Redis makes a great message queue server:

(:require [taoensso.carmine.message-queue :as car-mq]) ; Add to `ns` macro

(def my-worker
  (car-mq/worker {:pool {<opts>} :spec {<opts>}} "my-queue"
   {:handler (fn [{:keys [message attempt]}]
               (println "Received" message)
               {:status :success})}))

(wcar* (car-mq/enqueue "my-queue" "my message!"))
%> Received my message!

(car-mq/stop my-worker)

Look simple? It is. But it's also distributed, fault-tolerant, and fast.


  • Messages are persistent (durable) as per Redis config.
  • Each message will be handled once and only once.
  • Handling is fault-tolerant: a message cannot be lost due to handler crash.
  • Message de-duplication can be requested on an ad hoc (per message) basis. In these cases, the same message cannot ever be entered into the queue more than once simultaneously or within a (per message) specifiable post-handling backoff period.

See the API docs for details.

Distributed locks

(:require [taoensso.carmine.locks :as locks]) ; Add to `ns` macro

(locks/with-lock "my-lock"
  {:pool {<opts>} :spec {<opts>}} ; Connection details
  1000 ; Time to hold lock
  500  ; Time to wait (block) for lock acquisition
  (println "This was printed under lock!"))

Again: simple, distributed, fault-tolerant, and fast. See the taoensso.carmine.locks namespace for details.

Tundra (beta)

Redis is a beautifully designed datastore that makes some explicit engineering tradeoffs. Probably the most important: your data must fit in memory. Tundra helps relax this limitation: only your hot data need fit in memory. How does it work?

  1. Use Tundra's dirty command any time you modify/create evictable keys.
  2. Use worker to create a threaded worker that'll automatically replicate dirty keys to your secondary datastore.
  3. When a dirty key hasn't been used in a specified TTL, it will be automatically evicted from Redis (eviction is optional if you just want to use Tundra as a backup/just-in-case mechanism).
  4. Use ensure-ks any time you want to use evictable keys. This will extend their TTL or fetch them from your datastore as necessary.

That's it: two Redis commands, and a worker! Tundra uses Redis' own dump/restore mechanism for replication, and Carmine's own message queue to coordinate the replication worker.

Tundra can be very easily extended to any K/V-capable datastore. Implementations are provided out-the-box for: Disk, Amazon S3 and DynamoDB.

(:require [taoensso.carmine.tundra :as tundra :refer (ensure-ks dirty)]
          [taoensso.carmine.tundra.s3]) ; Add to ns

(def my-tundra-store
    ;; A datastore that implements the necessary (easily-extendable) protocol:
    (taoensso.carmine.tundra.s3/s3-datastore {:access-key "" :secret-key ""}

;; Now we have access to the Tundra API:
 (worker    my-tundra-store {} {}) ; Create a replication worker
 (dirty     my-tundra-store "foo:bar1" "foo:bar2" ...) ; Queue for replication
 (ensure-ks my-tundra-store "foo:bar1" "foo:bar2" ...) ; Fetch from replica when necessary

Note that this API makes it convenient to use several different datastores simultaneously (perhaps for different purposes with different latency requirements).

See the relevant docstrings for details.


Redis is probably most famous for being fast. Carmine does what it can to hold up its end and currently performs well:

Performance comparison chart

Detailed benchmark information is available on Google Docs. Note that these numbers are for unpipelined requests: you could do a lot more with pipelining.

In principle it should be possible to get close to the theoretical maximum performance of a JVM-based client. This will be an ongoing effort but please note that my first concern for Carmine is performance-per-unit-power rather than absolute performance. For example Carmine willingly pays a small throughput penalty to support binary-safe arguments and again for composable commands.

Likewise, I'll happily trade a little less throughput for simpler code.


Carmine was developed with the help of the YourKit Java Profiler. YourKit, LLC kindly supports open source projects by offering an open source license. They also make the YourKit .NET Profiler.

This project supports the CDS and ClojureWerkz goals

  • CDS, the Clojure Documentation Site, is a contributer-friendly community project aimed at producing top-notch, beginner-friendly Clojure tutorials and documentation. Awesome resource.

  • ClojureWerkz is a growing collection of open-source, batteries-included Clojure libraries that emphasise modern targets, great documentation, and thorough testing. They've got a ton of great stuff, check 'em out!

Contact & contributing

lein start-dev to get a (headless) development repl that you can connect to with Cider (Emacs) or your IDE.

Please use the project's GitHub issues page for project questions/comments/suggestions/whatever (pull requests welcome!). Am very open to ideas if you have any!

Otherwise reach me (Peter Taoussanis) at or on Twitter. Cheers!


Copyright © 2012-2015 Peter Taoussanis. Distributed under the Eclipse Public License, the same as Clojure.