Kestrel is based on Blaine Cook's "starling" simple, distributed message queue, with added features and bulletproofing, as well as the scalability offered by actors and the JVM.

Each server handles a set reliable, ordered message queues. When you put a cluster of these servers together, with no cross communication, and pick a server at random whenever you do a set or get, you end up with a reliable, loosely ordered message queue.

In many situations, loose ordering is sufficient. Dropping the requirement on cross communication makes it horizontally scale to infinity and beyond: no multicast, no clustering, no "elections", no coordination at all. No talking! Shhh!

For more information about what it is and how to use it, check out the included guide.

Kestrel is:

• fast

  It runs on the JVM so it can take advantage of the hard work people have put into java performance.

• small

  Currently about 2K lines of scala (including comments), because it relies on Apache Mina (a rough equivalent of Danger's ziggurat or Ruby's EventMachine) and actors -- and frankly because Scala is extremely expressive.

• durable

  Queues are stored in memory for speed, but logged into a journal on disk so that servers can be shutdown or moved without losing any data.

• reliable

  A client can ask to "tentatively" fetch an item from a queue, and if that client disconnects from kestrel before confirming ownership of the item, the item is handed to another client. In this way, crashing clients don't cause lost messages.

Kestrel is not:

• strongly ordered

  While each queue is strongly ordered on each machine, a cluster will appear "loosely ordered" because clients pick a machine at random for each operation. The end result should be "mostly fair".

• transactional

  This is not a database. Item ownership is transferred with acknowledgement, but kestrel does not support multiple outstanding operations, and treats each enqueued item as an atomic unit.

Kestrel requires java 6 (for JMX support) and ant 1.7. If you see an error about missing JMX classes, it usually means you're building with java 5. On a mac, you may have to hard-code an annoying JAVA_HOME to use java 6:

$ export JAVA_HOME=/System/Library/Frameworks/JavaVM.framework/Versions/1.6/Home

Building from source is easy:

$ ant

Scala libraries and dependencies will be downloaded from maven repositories the first time you do a build. The finished distribution will be in dist.

You can run kestrel by hand via:

$ java -jar ./dist/kestrel-VERSION/kestrel-VERSION.jar

To run in development mode (using development.conf instead of production.conf), add a stage variable:

$ java -Dstage=development -jar ./dist/kestrel-VERSION/kestrel-VERSION.jar

When running it as a server, a startup script is provided in dist/kestrel-VERSION/scripts/kestrel.sh. The script assumes you have daemon, a standard daemonizer for Linux, but also available here for all common unix platforms.

The created archive kestrel-VERSION.tar.bz2 can be expanded into a place like /usr/local (or wherever you like) and executed within its own folder as a self-contained package. All dependent jars are included, and the startup script loads things from relative paths.

The default configuration puts logfiles into /var/log/kestrel/ and queue journal files into /var/spool/kestrel/.

The startup script logs extensive GC information to a file named stdout in the log folder. If kestrel has problems starting up (before it can initialize logging), it will usually appear in error in the same folder.

Queue configuration is described in detail in docs/guide.md (an operational guide). There are a few global config options that should be self-explanatory:

• host

  Host to accept connections on.

• port

  Port to listen on. 22133 is the standard.

• timeout

  Seconds after which an idle client is disconnected, or 0 to have no idle timeout.

• queue_path

  The folder to store queue journal files in. Each queue (and each client of a fanout queue) gets its own file here. For saving to Hadoop File System, the queue_path should be prefixed with "hdfs://hostname:port" (typical HDFS scheme.)

• log

  Logfile configuration, as described in configgy.

All of the below timings are on my 2GHz 2006-model macbook pro.

Since starling uses eventmachine in a single-thread single-process form, it has similar results for all access types (and will never use more than one core).

=========  =================  ==========
# Clients  Pushes per client  Total time
=========  =================  ==========
        1             10,000        3.8s
       10              1,000        2.9s
      100                100        3.1s
=========  =================  ==========

Kestrel uses N+1 I/O processor threads (where N = the number of available CPU cores), and a pool of worker threads for handling actor events. Therefore it handles more poorly for small numbers of heavy-use clients, and better for large numbers of clients.

=========  =================  ==========
# Clients  Pushes per client  Total time
=========  =================  ==========
        1             10,000        3.8s
       10              1,000        2.4s
      100                100        1.6s
=========  =================  ==========

A single-threaded set of 5 million puts gives a fair idea of throughput distribution, this time on a 2.5GHz 2008-model macbook pro:

$ ant -f tests.xml put-many-1 -Ditems=5000000
[java] Finished in 1137250 msec (227.5 usec/put throughput).
[java] Transactions: min=106.00; max=108581.00 91335.00 60721.00; median=153.00; average=201.14 usec
[java] Transactions distribution: 5.00%=129.00 10.00%=134.00 25.00%=140.00 50.00%=153.00 75.00%=177.00 90.00%=251.00 95.00%=345.00 99.00%=586.00 99.90%=5541.00 99.99%=26910.00

This works out to about 3.23MB/sec (over loopback) and about 4400 puts/sec.

Robey Pointer <[email protected]>