Hydra is a distributed data processing and storage system originally developed at AddThis. It ingests streams of data (think log files) and builds trees that are aggregates, summaries, or transformations of the data. These trees can be used by humans to explore (tiny queries), as part of a machine learning pipeline (big queries), or to support live consoles on websites (lots of queries).
You can run hydra from the command line to slice and dice that Apache access log you have sitting around (or that gargantuan csv file). Or if terabytes per day is your cup of tea run a Hydra Cluster that supports your job with resource sharing, job management, distributed backups, data partitioning, and efficient bulk file transfer.
Assuming you have Apache Maven installed and configured:
mvn package
Should compile and build jars. All hydra dependencies should be available on maven central but hydra itself is not yet published.
Berkeley DB Java Edition is used for several core features. The sleepycat license has strong copyleft properties that do not match the rest of the project. It is set as a non-transitive dependency to avoid inadvertently pulling it into downstream projects. In the future hydra should have pluggable storage with multiple implementations.
The hydra-uber
module builds an exec
jar containing hydra and all
of it's dependencies. To include BDB JE when building with mvn package
use -P bdbje
. The main class of the exec
jar launches
the various components of a hydra cluster by name.
JDK 7 is required. Hydra has been developed on Linux (Centos 6) and should work on any modern Linux distro. Other unix-like systems should work with minor changes but have not been tested. Mac OSX should work for building and running local-stack (see below).
Hydra uses rabbitmq for low volume
command and control message exchange. On a modern Linux systems
apt-get install rabbitmq-server
and running with the default
settings is adequate in most cases.
To run efficiently Hydra needs a mechanism to take copy on write
backups of the output of jobs. The is currently accomplished by
adding the fl-cow library to
LD_PRELOAD
. Experimenting with other approaches such as ZFS or cp --reflink
are under consideration.
Many components assume that there is a local user called hydra
and
that all minion nodes can ssh as that user to each other. This is
used most prominently for rsync
based replicas.
While hydra can be used for ad-hoc analysis of csv and other local files, it's most commonly used in a distributed cluster. In that case the following components are involved:
- ZooKeeper
- Spawn: Job control and execution
- Minion: Task runner
- QueryMaster: Handler for queries
- QueryWorker: Handle scatter-gather requests from QueryMaster
- Meshy: File server
A typical configuration is to have a cluster head with Spawn & QueryMaster backed by a homogeneous clusters of nodes running Minion, QueryWorker, and Meshy.
For local development all of the above components can run together in
a single stack run out of hydra-local
. There is a local-stack.sh
script to assist with this. To run the local stack:
- You must be able to build hydra
- Have rabbitmq installed
- Allow your current user to ssh to itself
The first time the script is run a hydra-local
directory will be created.
./hydra-uber/bin/local-stack.sh start
- start ZooKeeper./hydra-uber/bin/local-stack.sh start
- start spawn, querymaster etc../hydra-uber/bin/local-stack.sh seed
- add some sample data
You can then navigate to http://localhost:5052/ and you should see the spawn web interface.
When done ./hydra-uber/bin/local-stack.sh stop
will stop everything
except ZooKeeper, and running stop
a second time will bring that
process down as well.
There are sample job configurations located in hydra-uber/local/sample/
Mailing list: http://groups.google.com/group/hydra-oss
Freenode channel: #hydra
It's x.y.z where:
- x: Something Big Happened
- y: next release
- z: strive for bug fix only
hydra is released under the Apache License Version 2.0. See Apache or the LICENSE file in this distribution for details.