Libkestrel is a library for scala/java containing:
-
ConcurrentBlockingQueue - a lock-free queue that allows readers to block (wait) for items
-
JournaledQueue - a queue with potentially many read-pointers, where state of the queue and its readers is saved to disk
These are variants and improvements of the building blocks of the kestrel distributed queue server project. The intent (as of November 2011) is to make kestrel use this library in its next major release.
$ sbt clean update package-dist
SBT must presently be version 0.11.2 or you'll get build errors (due to SBT plugin versioning). Note also that if the script you use to run SBT ("sbt" in the example) enables assertions (via java -ea or java -enableassertions) the build will fail due to a Scala compiler bug.
Come talk to us on the kestrel mailing list!
http://groups.google.com/group/kestrel-talk
ConcurrentBlockingQueue extends the idea of java's ConcurrentLinkedQueue to
allow consumers to block, indefinitely or with a timeout, until items arrive.
It works by having one ConcurrentLinkedQueue for the queue itself, and
another one to track waiting consumers. Each time an item is put into the
queue, it's handed off to the next waiting consumer, like an airport taxi
line.
The handoff occurs in a serialized block (like the Serialized trait in
util-core), so when there's no contention, a new item is handed directly from
the producer to the consumer. When there is contention, producers increment a
counter representing how many pent-up handoffs there are, and the producer
that got into the serialized block first will do each handoff in order until
the count is zero again. This way, no producer is ever blocked.
Consumers receive a future that will eventually be fulfilled either with
Some(item) if an item arrived before the requested timeout, or None if the
request timed out. If an item was available immediately, the future will be
fulfilled before the consumer receives it. This way, no consumer is ever
blocked.
A JournaledQueue is a journaled queue that may have multiple "readers", each of which may have multiple consumers.
When an item is added to a queue, it's journaled and notifiation is passed on
to any readers. There is always at least one reader, and the reader knows its
current location in the memory-mapped journal file. If there are multiple
readers, they behave as multiple independent queues, each receiving a copy of
each item added to the JournaledQueue, but sharing a single journal. They may
have different policies on queue size limits, item expiration, and error
handling.
Items are read only from readers. When an item is available, it's set aside as an "open read", but not committed to the journal. A separate call is made to either commit the item or abort it. Aborting an item returns it to the head of the queue to be given to the next consumer.
Periodically each reader records its state in a separate checkpoint journal. When initialized, if a journal already exists for a queue and its readers, each reader restores itself from this saved state. If the queues were not shutdown cleanly, the state files may be out of date and items may be replayed. Care is taken never to let any of the journal files be corrupted or in a non-recoverable state. In case of error, the choice is always made to possibly replay items instead of losing them.
The writer journal is treated as an "infinite scroll" of put operations. Each journal file is created with the current timestamp in its filename, and after it gets "full" (configurable, but 16MB by default), that file is closed and a new one is opened. If no readers ever consumed items from the queue, these files would sit around forever.
Once all readers have moved their read-pointer (the head of their queue) past the end of journal file, that file is archived. By default, that just means the file is deleted, but one of the configuration parameters allows you to have the dead files moved to a different folder instead.
There are several advantages to splitting the journals into a single (multi-file) writer journal and several reader checkpoint files:
-
Fan-out queues (multiple read-pointers into the same queue) are free, and all the readers share a single journal, saving disk space and bandwidth. Disk bandwidth is now almost entirely based on write throughput, not the number of readers.
-
The journals never have to be "packed" to save disk space, the way they did in kestrel 2.x. Packing creates more disk I/O at the very time a server might be struggling to keep up with existing load.
-
Archiving old queue items is trivial, and allows you to do some meta-analysis of load offline.
Some load tests are included. You can run them with
$ ./dist/libkestrel/scripts/qtest
which will list the available tests. Each test responds to "--help".
-
BlockingQueue - interface for any blocking queue
-
SimpleBlockingQueue - a simple queue using "synchronized", based on the one in kestrel 2.1
-
ConcurrentBlockingQueue - a lock-free BlockingQueue (see above)
-
PeriodicSyncFile - a writable file that
fsyncs on a schedule -
ItemIdList - a simple Set[Long] implementation optimized for small sets
-
QueueItem - an item plus associated metadata (item ID, expiration time)
-
JournalFile - a single on-disk journal file, encapsulating reading & writing journal records
-
Journal - representation of a collection of files (the writer files and a file for each reader)
-
JournaledQueue - a
Journalbased queue implementation -
JournaledBlockingQueue - JournaledQueue wrappers that provide a simplified interface for users that only use a single reader (with or without transactions)
-
Trustin suggested that the read-pointer files could be, instead of an id, a filename and position. That would save us from reading the first half of a journal file on startup (to find the id).
-
Nick suggested that writing all of the readers into the same file could reduce disk I/O by writing fewer blocks during reader checkpointing.