twitter-archive/kestrel

Kestrel

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 of 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 has a mailing list here: [email protected]

Author's address: Robey Pointer <[email protected]>

Status

We've deprecated Kestrel because internally we've shifted our attention to an alternative project based on DistributedLog, and we no longer have the resources to contribute fixes or accept pull requests. While Kestrel is a great solution up to a certain point (simple, fast, durable, and easy to deploy), it hasn't been able to cope with Twitter's massive scale (in terms of number of tenants, QPS, operability, diversity of workloads etc.) or operating environment (an Aurora cluster without persistent storage).

Features

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 2500 lines of scala, because it relies on Netty (a rough equivalent of Danger's ziggurat or Ruby's EventMachine) -- and 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.

Anti-Features

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 grouping multiple operations into an atomic unit.

Downloading it

The latest release is always on the homepage here:

• http://robey.github.io/kestrel/

Or the latest development versions & branches are on github:

• http://gitub.com/robey/kestrel

Building it

Kestrel requires java 6 and sbt 0.11.2. Presently some sbt plugins used by kestrel depend on that exact version of sbt. On OS X 10.5, 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:

$ sbt clean update package-dist

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

Running it

You can run kestrel by hand, in development mode, via:

$ ./dist/kestrel-VERSION/scripts/devel.sh

Like all ostrich-based servers, it uses the "stage" property to determine which config file to load, so devel.sh sets -Dstage=development.

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.zip 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. The current startup script, however, assumes that kestrel has been deployed to /usr/local/kestrel/current (e.g., as if by capistrano), and the startup script loads kestrel from that path.

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.

Configuration

Queue configuration is described in detail in docs/guide.md (an operational guide). Scala docs for the config variables are here.

Performance

Several performance tests are included. To run them, first start up a kestrel instance locally.

$ sbt clean update package-dist
$ ./dist/kestrel-VERSION/scripts/devel.sh

Put-many

This test just spams a kestrel server with "put" operations, to see how quickly it can absorb and journal them.

A sample run on a 2010 MacBook Pro:

$ ./dist/kestrel/scripts/load/put-many -n 100000
Put 100000 items of 1024 bytes to localhost:22133 in 1 queues named spam
  using 100 clients.
Finished in 6137 msec (61.4 usec/put throughput).
Transactions: min=71.00; max=472279.00 472160.00 469075.00;
  median=3355.00; average=5494.69 usec
Transactions distribution: 5.00%=485.00 10.00%=1123.00 25.00%=2358.00
  50.00%=3355.00 75.00%=4921.00 90.00%=7291.00 95.00%=9729.00
  99.00%=50929.00 99.90%=384638.00 99.99%=467899.00

Many-clients

This test has one producer that trickles out one item at a time, and a pile of consumers fighting for each item. It usually takes exactly as long as the number of items times the delay, but is useful as a validation test to make sure kestrel works as advertised without blowing up.

A sample run on a 2010 MacBook Pro:

$ ./dist/kestrel/scripts/load/many-clients
many-clients: 100 items to localhost using 100 clients, kill rate 0%,
  at 100 msec/item
Received 100 items in 11046 msec.

This test always takes about 11 seconds -- it's a load test instead of a speed test.

Flood

This test starts up one producer and one consumer, and just floods items through kestrel as fast as it can.

A sample run on a 2010 MacBook Pro:

$ ./dist/kestrel/scripts/load/flood
flood: 1 threads each sending 10000 items of 1kB through spam
Finished in 1563 msec (156.3 usec/put throughput).
Consumer(s) spun 0 times in misses.

Packing

This test starts up one producer and one consumer, seeds the queue with a bunch of items to cause it to fall behind, then does cycles of flooding items through the queue, separated by pauses. It's meant to test kestrel's behavior with a queue that's fallen behind and stays behind indefinitely, to make sure the journal files are packed periodically without affecting performance too badly.

A sample run on a 2010 MacBook Pro:

$ ./dist/kestrel/scripts/load/packing -c 10 -q small
packing: 25000 items of 1kB with 1 second pauses
Wrote 25000 items starting at 0.
cycle: 1
Wrote 25000 items starting at 25000.
Read 25000 items in 5279 msec. Consumer spun 0 times in misses.
cycle: 2
Wrote 25000 items starting at 50000.
Read 25000 items in 4931 msec. Consumer spun 0 times in misses.
...
cycle: 10
Wrote 25000 items starting at 250000.
Read 25000 items in 5304 msec. Consumer spun 0 times in misses.
Read 25000 items in 3370 msec. Consumer spun 0 times in misses.

You can see the journals being packed in the kestrel log. Like "many-clients", this test is a load test instead of a speed test.

Leaky-reader

This test starts a producer and several consumers, with the consumers occasionally "forgetting" to acknowledge an item that they've read. It verifies that the un-acknowledged items are eventually handed off to another consmer.

A sample run:

$ ./dist/kestrel/scripts/load/leaky-reader -n 100000 -t 10
leaky-reader: 10 threads each sending 100000 items through spam
Flushing queues first.
1000
2000
100000
Finished in 40220 msec (40.2 usec/put throughput).
Completed all reads

Like "many-clients", it's just a load test.