Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Build Status Clojars Project


A queue which schedules fairly by key.


We developed this library with a program in mind that requires a central in-memory queue. The queue must allow the program to serve active users in a timely manner while still ensuring that users with massive traffic get their job done eventually.

We considered other options like Clojure's core.async, clojure.lang.PersistentQueue, and Java's java.util.PriorityQueue but none met the requirements.

So we wrote queue-by which does what we need and has a nice name, too.


To use this library in your project, add the following to your :dependencies in your project.clj or build.boot:

[com.acrolinx.clj-queue-by "0.1.1"]

To create a queue, require the com.acrolinx.clj-queue-by namespace and call queue-by:

(ns test-the-queue.core
   (:require [com.acrolinx.clj-queue-by :as q]))

(def queue (q/queue-by :name))

Here we create the queue with a key-fn :name, so items will get a dedicated queue per :name. You can store the queue in an atom, too. Alternatively, create it as a local variable and pass it around.

For a detailed description of the scheduling algorithm per key, see The Scheduling Mechanism below.

The queue can have a maximum size. In the first example, we stick to the default maximum of 128:

(def queue (q/queue-by :name))

If you want to have a different limit, call it with a second argument:

(def queue (q/queue-by :name 1000))

If the second argument is an explicit nil the queue is unbounded and no size checks are performed. Use at your own risk.

(def queue (q/queue-by :name nil))

Add an item to the queue by calling it with the item as the argument:

(queue {:name "alice" :a 1})

When you try to push more items than the limit, an exception is thrown.

(def queue (q/queue-by :id 1))
(queue {:id 1})
(try (queue {:id 2})
       (catch clojure.lang.ExceptionInfo e
         (ex-data e)))
;=> {:item {:id 2}, :current-size 1}

Calling the queue like a function works, because it implements the IFn interface which is used when calling functions in Clojure.

How many items are in the queue?

(count queue)

The queue implements Counted, the interface behind the count function. Performance guarantees are inherited from Clojure's hash map and clojure.lang.PersistentQueue which are used under the hood.

What's inside the queue?

(deref queue)

Dereferencing the queue returns a two-element vector: first the current snapshot queue (a clojure.lang.PersistentQueue), second a hash-map with the per-key queues (again clojure.lang.PersistentQueue). Works by implementing IDeref. The dereferenced information can be used for monitoring.

Finally, read an item from the queue by calling it without an argument.


Reading from the queue returns nil when no item is in the queue.


The queue allows you to add nil items but you won't be able to distinguish at the receiving end if nil was in the queue or the queue was empty.

Also, nil gets its own queue when the key-fn returns nil just as any other value.

Comparison with core.async

  • The core.async library is much more sophisticated and much more powerful. At the same time, it is also harder to use.
  • clj-queue-by doesn't support transducers while core.async does.
  • The buffers in core.async which back the channels are surprisingly intransparent. You can't look into them or log when things are being dropped. Also, channels do not support derefing and counting. Probably for good reasons, but the use-case that triggered the development of clj-queue-by required more introspection and transparency. In core.async, you can overcome all these limitations if you implement your own buffer to back a channel. In fact, we did this previously.
  • core.async is battle-proven and has shown that it runs well in production. clj-queue-by is just beginning to show it.
  • Channels in core.async are meant to be used a lot. You can easily create tens or hundreds of them. In contrast, clj-queue-by was developed to be the central in-memory queue for a program.

The Scheduling Mechanism

On the sending and the receiving end, the queue behaves just like any other queue. Internally though, items are put into separate queues given by the key-fn you define.

The scheduling mechanism is somewhat related to the Completely Fair Scheduling algorithm used in the Linux Kernel and the Fair Scheduler used in Apache Hadoop.

Imagine having hash-map items with a :name key in it:

 {:name "alice"
  :data 1}

If you push several items with different names into the queue, it creates separate queues per key.

"alice": item1, item2
"bob":   item3

Now, if Alice puts many items into a common FIFO queue and the consumption of the items takes a while, Bob would have to wait a long time for his thing to happen.

To solve this problem, this queue implementation always takes off the leading items of the queue per user and delivers them.

  1. Push {:name "alice" :data 1}
  2. Push {:name "alice" :data 2}
  3. Push {:name "alice" :data 3}
  4. Push {:name "bob" :data "x"}

After these pushes, the queue looks like this internally:

alice: {:name "alice" :data 1},
       {:name "alice" :data 2},
       {:name "alice" :data 3},
bob:   {:name "bob"   :data "x"}

When you start pulling, it will deliver the items in this order:

  1. Pull {:name "alice" :data 1}
  2. Pull {:name "bob" :data "x"}
  3. Pull {:name "alice" :data 2}
  4. Pull {:name "alice" :data 3}

Note, how Bob's item got delivered before Alice's second item.

Think of it as taking a snapshot of the heads of the queue when polling and then delivering this snapshot. When it's empty, a new snapshot is taken. Actually, this is exactly what happens behind the scenes.

Another example:

  1. Push {:name "alice" :data 1}
  2. Push {:name "bob" :data "x"}
  3. Pull {:name "alice" :data 1}. This takes a snapshot and delivers the oldest item. The item from Bob is now the head of the snapshot.
  4. Push {:name "alice" :data 2}. This adds the new item after the snapshot.
  5. Push {:name "alice" :data 3}
  6. Pull {:name "bob" :data "x"}
  7. Pull {:name "alice" :data 2}
  8. Pull {:name "alice" :data 3}

A last example:

  1. Push {:name "alice" :data 1}
  2. Push {:name "bob" :data "x"}
  3. Push {:name "alice" :data 2}
  4. Pull {:name "alice" :data 1}. This takes a snapshot and delivers the oldest item. The item from Bob is now the head of the snapshot. Alice's second item stays on her dedicated queue.
  5. Push {:name "alice" :data 3}
  6. Push {:name "alice" :data 4}
  7. Pull {:name "bob" :data "x"}. Was head of the snapshot.
  8. Push {:name "bob" :data "y"}
  9. Pull {:name "alice" :data 2}. A new snapshot is created with the head items of both Alice and Bob added. Thus, Bob's item overtakes Alice's larger queue of items.
  10. Pull {:name "bob" :data "y"}
  11. Pull {:name "alice" :data 3}
  12. Pull {:name "alice" :data 4}

If we reduce the items to just the :data, the queue in this example went through the following internal states (empty queues suppressed):

alice: data 1

alice: data 1
bob:   data x

alice: data 1, data 2
bob:   data x

SNAPSHOT:    bob/data x
alice:       data 2
-> Returned: alice/data 1

SNAPSHOT:    bob/data x
alice:       data 2, data 3

SNAPSHOT:    bob/data x
alice:       data 2, data 3, data 4

alice:       data 2, data 3, data 4
-> Returned: bob/data x

alice:       data 2, data 3, data 4
bob:         data y

SNAPSHOT:    bob/data y
alice:       data 3, data 4
-> Returned: alice/data 2

alice:       data 3, data 4
-> Returned: bob/data y

alice:       data 4
-> Returned: alice/data 3

-> Returned: alice/data 4


This project accepts pull request from registered authors. We ask you to sign our Contributor License Agreement first. Please reach out to us at so that we can prepare everything.

Don't panic. It is just some easy legalese that helps us maintain the availability of this software for everyone.


  • @replrep πŸ’» πŸ”
  • @nblumoe πŸ’» πŸ”

Type of contribution:

  • πŸ” Feedback, review
  • πŸ’» Code, pull request


Copyright Β© 2017-2019 Acrolinx GmbH

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.