Archive for March, 2010

Thrift vs. Protocol Buffers

Posted on March 21, 2010. Filed under: Computer Science, Programming | Tags: , , , |

Google recently released its Protocol Buffers as open source. About a year ago, Facebook released a similar product called Thrift. I’ve been comparing them; here’s what I’ve found:

Thrift Protocol Buffers
Backers Facebook, Apache (accepted for incubation) Google
Bindings C++, Java, Python, PHP, XSD, Ruby, C#, Perl, Objective C, Erlang, Smalltalk, OCaml, and Haskell C++, Java, Python
(Perl, Ruby, and C# under discussion)
Output Formats Binary, JSON Binary
Primitive Types bool
byte
16/32/64-bit integersdouble
string
byte sequence
map<t1,t2>
list<t>
set<t>
bool32/64-bit integers
float
double
string
byte sequence

“repeated” properties act like lists

Enumerations Yes Yes
Constants Yes No
Composite Type struct message
Exception Type Yes No
Documentation So-so Good
License Apache BSD-style
Compiler Language C++ C++
RPC Interfaces Yes Yes
RPC Implementation Yes No
Composite Type Extensions No Yes

Overall, I think Thrift wins on features and Protocol Buffers win on
documentation. Implementation-wise, they’re quite similar. Both use
integer tags to identify fields, so you can add and remove fields
without breaking existing code. Protocol Buffers support
variable-width encoding of integers, which saves a few bytes. (Thrift
has an experimental output format with variable-width ints.)

The major difference is that Thrift provides a full client/server RPC
implementation, whereas Protocol Buffers only generate stubs to use in
your own RPC system.

Update July 12, 2008: I haven’t tested for speed, but from a cursory examination it seems that, at the binary level, Thrift and Protocol Buffers are very similar. I think Thrift will develop a more coherent community now that it’s under Apache incubation. It just moved to a new web site and mailing list, and the issue tracker is active.

Reference:

http://stuartsierra.com/2008/07/10/thrift-vs-protocol-buffers (Original Site)

Read Full Post | Make a Comment ( None so far )

up and running with Cassandra

Posted on March 21, 2010. Filed under: Java, Linux, MySQL, PHP, Services | Tags: , , |

Cassandra is a hybrid non-relational database in the same class as Google’s BigTable. It is more featureful than a key/value store like Dynomite, but supports fewer query types than a document store like MongoDB.

Cassandra was started by Facebook and later transferred to the open-source community. It is an ideal runtime database for web-scale domains like social networks.

This post is both a tutorial and a “getting started” overview. You will learn about Cassandra’s features, data model, API, and operational requirements—everything you need to know to deploy a Cassandra-backed service.

Jan 8, 2010: post updated for Cassandra gem 0.7 and Cassandra version 0.5.

features

There are a number of reasons to choose Cassandra for your website. Compared to other databases, three big features stand out:

  • Flexible schema: with Cassandra, like a document store, you don’t have to decide what fields you need in your records ahead of time. You can add and remove arbitrary fields on the fly. This is an incredible productivity boost, especially in large deployments.
  • True scalability: Cassandra scales horizontally in the purest sense. To add more capacity to a cluster, turn on another machine. You don’t have restart any processes, change your application queries, or manually relocate any data.
  • Multi-datacenter awareness: you can adjust your node layout to ensure that if one datacenter burns in a fire, an alternative datacenter will have at least one full copy of every record.

Some other features that help put Cassandra above the competition :

  • Range queries: unlike most key/value stores, you can query for ordered ranges of keys.
  • List datastructures: super columns add a 5th dimension to the hybrid model, turning columns into lists. This is very handy for things like per-user indexes.
  • Distributed writes: you can read and write any data to anywhere in the cluster at any time. There is never any single point of failure.

installation

You need a Unix system. If you are using Mac OS 10.5, all you need is Git. Otherwise, you need to install Java 1.6, Git 1.6, Ruby, and Rubygems in some reasonable way.

Start a terminal and run:

sudo gem install cassandra

If you are using Mac OS, you need to export the following environment variables:

export JAVA_HOME="/System/Library/Frameworks/JavaVM.framework/Versions/1.6/Home"
export PATH="/System/Library/Frameworks/JavaVM.framework/Versions/1.6/Home/bin:$PATH"

Now you can build and start a test server with cassandra_helper:

cassandra_helper cassandra

It runs!

live demo

The above script boots the server with a schema that we can interact with. Open another terminal window and start irb, the Ruby shell:

irb

In the irb prompt, require the library:

require 'rubygems'
require 'cassandra'
include Cassandra::Constants

Now instantiate a client object:

twitter = Cassandra.new('Twitter')

Let’s insert a few things:

user = {'screen_name' => 'buttonscat'}
twitter.insert(:Users, '5', user)  

tweet1 = {'text' => 'Nom nom nom nom nom.', 'user_id' => '5'}
twitter.insert(:Statuses, '1', tweet1)

tweet2 = {'text' => '@evan Zzzz....', 'user_id' => '5', 'reply_to_id' => '8'}
twitter.insert(:Statuses, '2', tweet2)

Notice that the two status records do not have all the same columns. Let’s go ahead and connect them to our user record:

twitter.insert(:UserRelationships, '5', {'user_timeline' => {UUID.new => '1'}})
twitter.insert(:UserRelationships, '5', {'user_timeline' => {UUID.new => '2'}})

The UUID.new call creates a collation key based on the current time; our tweet ids are stored in the values.

Now we can query our user’s tweets:

timeline = twitter.get(:UserRelationships, '5', 'user_timeline', :reversed => true)
timeline.map { |time, id| twitter.get(:Statuses, id, 'text') }
# => ["@evan Zzzz....", "Nom nom nom nom nom."]

Two tweet bodies, returned in recency order—not bad at all. In a similar fashion, each time a user tweets, we could loop through their followers and insert the status key into their follower’s home_timeline relationship, for handling general status delivery.

the data model

Cassandra is best thought of as a 4 or 5 dimensional hash. The usual way to refer to a piece of data is as follows: a keyspace, a column family, a key, an optional super column, and a column. At the end of that chain lies a single, lonely value.

Let’s break down what these layers mean.

  • Keyspace (also confusingly called “table”): the outer-most level of organization. This is usually the name of the application. For example, 'Twitter' and 'Wordpress' are both good keyspaces. Keyspaces must be defined at startup in the storage-conf.xml file.
  • Column family: a slice of data corresponding to a particular key. Each column family is stored in a separate file on disk, so it can be useful to put frequently accessed data in one column family, and rarely accessed data in another. Some good column family names might be :Posts, :Users and :UserAudits. Column families must be defined at startup.
  • Key: the permanent name of the record. You can query over ranges of keys in a column family, like :start => '10050', :finish => '10070'—this is the only index Cassandra provides for free. Keys are defined on the fly.

After the column family level, the organization can diverge—this is a feature unique to Cassandra. You can choose either:

  • A column: this is a tuple with a name and a value. Good columns might be 'screen_name' => 'lisa4718' or 'Google' => 'http://google.com'.It is common to not specify a particular column name when requesting a key; the response will then be an ordered hash of all columns. For example, querying for (:Users, '174927') might return:
    {'name' => 'Lisa Jones',
     'gender' => 'f',
     'screen_name' => 'lisa4718'}

    In this case, name, gender, and screen_name are all column names. Columns are defined on the fly, and different records can have different sets of column names, even in the same keyspace and column family. This lets you use the column name itself as either structure or data. Columns can be stored in recency order, or alphabetical by name, and all columns keep a timestamp.

  • A super column: this is a named list. It contains standard columns, stored in recency order.Say Lisa Jones has bookmarks in several categories. Querying (:UserBookmarks, '174927') might return:
    {'work' => {
        'Google' => 'http://google.com',
        'IBM' => 'http://ibm.com'},
     'todo': {...},
     'cooking': {...}}

    Here, work, todo, and cooking are all super column names. They are defined on the fly, and there can be any number of them per row. :UserBookmarks is the name of the super column family. Super columns are stored in alphabetical order, with their sub columns physically adjacent on the disk.

Super columns and standard columns cannot be mixed at the same (4th) level of dimensionality. You must define at startup which column families contain standard columns, and which contain super columns with standard columns inside them.

Super columns are a great way to store one-to-many indexes to other records: make the sub column names TimeUUIDs (or whatever you’d like to use to sort the index), and have the values be the foreign key. We saw an example of this strategy in the demo, above.

If this is confusing, don’t worry. We’ll now look at two example schemas in depth.

twitter schema

Here is the schema definition we used for the demo, above. It is based on Eric Florenzano’s Twissandra:

<Keyspace Name="Twitter">
  <ColumnFamily CompareWith="UTF8Type" Name="Statuses" />
  <ColumnFamily CompareWith="UTF8Type" Name="StatusAudits" />
  <ColumnFamily CompareWith="UTF8Type" Name="StatusRelationships"
    CompareSubcolumnsWith="TimeUUIDType" ColumnType="Super" />
  <ColumnFamily CompareWith="UTF8Type" Name="Users" />
  <ColumnFamily CompareWith="UTF8Type" Name="UserRelationships"
    CompareSubcolumnsWith="TimeUUIDType" ColumnType="Super" />
</Keyspace>

What could be in StatusRelationships? Maybe a list of users who favorited the tweet? Having a super column family for both record types lets us index each direction of whatever many-to-many relationships we come up with.

Here’s how the data is organized:

Click to enlarge

Cassandra lets you distribute the keys across the cluster either randomly, or in order, via the Partitioner option in the storage-conf.xml file.

For the Twitter application, if we were using the order-preserving partitioner, all recent statuses would be stored on the same node. This would cause hotspots. Instead, we should use the random partitioner.

Alternatively, we could preface the status keys with the user key, which has less temporal locality. If we used user_id:status_id as the status key, we could do range queries on the user fragment to get tweets-by-user, avoiding the need for a user_timeline super column.

multi-blog schema

Here’s a another schema, suggested to me by Jonathan Ellis, the primary Cassandra maintainer. It’s for a multi-tenancy blog platform:

<Keyspace Name="Multiblog">
  <ColumnFamily CompareWith="TimeUUIDType" Name="Blogs" />
  <ColumnFamily CompareWith="TimeUUIDType" Name="Comments"/>
</Keyspace>

Imagine we have a blog named ‘The Cutest Kittens’. We will insert a row when the first post is made as follows:

require 'rubygems'
require 'cassandra'
include Cassandra::Constants

multiblog = Cassandra.new('Multiblog')

multiblog.insert(:Blogs, 'The Cutest Kittens',
  { UUID.new =>
    '{"title":"Say Hello to Buttons Cat","body":"Buttons is a cute cat."}' })

UUID.new generates a unique, sortable column name, and the JSON hash contains the post details. Let’s insert another:

multiblog.insert(:Blogs, 'The Cutest Kittens',
  { UUID.new =>
    '{"title":"Introducing Commie Cat","body":"Commie is also a cute cat"}' })

Now we can find the latest post with the following query:

post = multiblog.get(:Blogs, 'The Cutest Kittens', :reversed => true).to_a.first

On our website, we can build links based on the readable representation of the UUID:

guid = post.first.to_guid
# => "b06e80b0-8c61-11de-8287-c1fa647fd821"

If the user clicks this string in a permalink, our app can find the post directly via:

multiblog.get(:Blogs, 'The Cutest Kittens', :start => UUID.new(guid), :count => 1)

For comments, we’ll use the post UUID as the outermost key:

multiblog.insert(:Comments, guid,
  {UUID.new => 'I like this cat. - Evan'})
multiblog.insert(:Comments, guid,
  {UUID.new => 'I am cuter. - Buttons'})

Now we can get all comments (oldest first) for a post by calling:

multiblog.get(:Comments, guid)

We could paginate them by passing :start with a UUID. See this presentation to learn more about token-based pagination.

We have sidestepped two problems with this data model: we don’t have to maintain separate indexes for any lookups, and the posts and comments are stored in separate files, where they don’t cause as much write contention. Note that we didn’t need to use any super columns, either.

storage layout and api comparison

The storage strategy for Cassandra’s standard model is the same as BigTable’s. Here’s a comparison chart:

multi-file per-file intra-file
Relational server database table* primary key column value
BigTable cluster table column family key column name column value
Cassandra, standard model cluster keyspace column family key column name column value
Cassandra, super column model cluster keyspace column family key super column name column name column value

* With fixed column names.

Column families are stored in column-major order, which is why people call BigTable a column-oriented database. This is not the same as a column-oriented OLAP database like Sybase IQ—it depends on what you use the column names for.

Click to enlarge

In row-orientation, the column names are the structure, and you think of the column families as containing keys. This is the convention in relational databases.

Click to enlarge

In column-orientation, the column names are the data, and the column families are the structure. You think of the key as containing the column family, which is the convention in BigTable. (In Cassandra, super columns are also stored in column-major order—all the sub columns are together.)

In Cassandra’s Ruby API, parameters are expressed in storage order, for clarity:

Relational SELECT `column` FROM `database`.`table` WHERE `id` = key;
BigTable table.get(key, "column_family:column")
Cassandra: standard model keyspace.get("column_family", key, "column")
Cassandra: super column model keyspace.get("column_family", key, "super_column", "column")

Note that Cassandra’s internal Thrift interface mimics BigTable in some ways, but this is being changed.

going to production

Cassandra is an alpha product and could, theoretically, lose your data. In particular, if you change the schema specified in the storage-conf.xml file, you must follow these instructions carefully, or corruption will occur (this is going to be fixed). Also, the on-disk storage format is expected to change in version 0.4.0. After that things will be a bit more stable.

The biggest deployment is at Facebook, where hundreds of terabytes of token indexes are kept in about a hundred Cassandra nodes. However, their use case allows the data to be rebuilt if something goes wrong. Currently there are no known deployments of non-transient data. Proceed carefully, keep a backup in an unrelated storage engine…and submit patches if things go wrong.

That aside, here is a guide for deploying a production cluster:

  • Hardware: get a handful of commodity Linux servers. 16GB memory is good; Cassandra likes a big filesystem buffer. You don’t need RAID. If you put the commitlog file and the data files on separate physical disks, things will go faster. Don’t use EC2 or friends except for testing; the virtualized I/O is too slow.
  • Configuration: in the storage-conf.xml schema file, set the replication factor to 3. List the IP address of one of the nodes as the seed. Set the listen address to the empty string, so the hosts will resolve their own IPs. Now, adjust the contents of cassandra.in.sh for your various paths and JVM options—for a 16GB node, set the JVM heap to 4GB.
  • Deployment: build a package of Cassandra itself and your configuration files, and deliver it to all your servers (I use Capistrano for this). Start the servers by setting CASSANDRA_INCLUDE in the environment to point to your cassandra.in.sh file, and run bin/cassandra. At this point, you should see join notices in the Cassandra logs:
    Cassandra starting up...
    Node 10.224.17.13:7001 has now joined.
    Node 10.224.17.14:7001 has now joined.

    Congratulations! You have a cluster. Don’t forget to turn off debug logging in the log4j.properties file.

  • Visibility: you can get a little more information about your cluster via the tool bin/nodeprobe, included:
    $ bin/nodeprobe --host 10.224.17.13 ring
    Token(124007023942663924846758258675932114665)  3 10.224.17.13  |<--|
    Token(106858063638814585506848525974047690568)  3 10.224.17.19  |   ^
    Token(141130545721235451315477340120224986045)  3 10.224.17.14  |-->|

    Cassandra also exposes various statistics over JMX.

Note that your client machines (not servers!) must have accurate clocks for Cassandra to resolve write conflicts properly. Use NTP.

conclusion

There is a misperception that if someone advocates a non-relational database, they either don’t understand SQL optimization, or they are generally a hater. This is not the case.

It is reasonable to seek a new tool for a new problem, and database problems have changed with the rise of web-scale distributed systems. This does not mean that SQL as a general-purpose runtime and reporting tool is going away. However, at web-scale, it is more flexible to separate the concerns. Runtime object lookups can be handled by a low-latency, strict, self-managed system like Cassandra. Asynchronous analytics and reporting can be handled by a high-latency, flexible, un-managed system like Hadoop. And in neither case does SQL lend itself to sharding.

I think that Cassandra is the most promising current implementation of a runtime distributed database, but much work remains to be done. We’re beginning to use Cassandra at Twitter, and here’s what I would like to happen real-soon-now:

  • Interface cleanup: the Thrift API for Cassandra is incomplete and inconsistent, which makes writing clients very irritating.
    Done!
  • Online migrations: restarting the cluster 3 times to add a column family is silly.
  • ActiveModel or DataMapper adapter: for interaction with business objects in Ruby.
    Done! Michael Koziarski on the Rails core team wrote an ActiveModel adapter.
  • Scala client: for interoperability with JVM middleware.

Go ahead and jump on any of those projects—it’s a chance to get in on the ground floor.

Cassandra has excellent performance. There some benchmark results for version 0.5 at the end of the Yahoo performance study.

further resources

Reference(cited):
Read Full Post | Make a Comment ( 1 so far )

Google Protocol Buffers and other data interchange formats

Posted on March 20, 2010. Filed under: Computer Science, Programming, Services | Tags: , , , , , |

We’ve been planning on moving to a new messaging protocol for a while. We’ve looked at a lot of different solutions but had enough issues with every proposed solution to date that we haven’t made a decision. JR Boyens pointed us to Google’s announcement Protocol Buffers: Google’s Data Interchange Format in July. Glanced at it but then it got lost in the everyday noise. Recent work on a project caused it to get more attention. I like what I see.

As part of a new offering we decided to add in our new messaging direction. We’re processing realtime voice conversations. Some of our major considerations are:

  1. Latency and Performance – Latency matters to us. A LOT. I’m including not only network transport but also memory and CPU. The total time it takes for a message to get from it’s native format in the sender to it’s native format in the receiver. We’re dealing with real time voice communications, too much latency and best case is the callers experience suffers. Our labor model is also sensitive to even small changes in latency. The smaller the latency the more efficient we are, the happier our client’s customers are and the more money we make. As greedy capitalist we see that as a good thing.
  2. Versioning – Our current system has no versioning. Yeah, short sighted on my part. We have to fix it so it’s required for any new message protocol. Protobuf fits our needs on this nicely. Different versions have to coexist and interoperate. We could do this on a different layer than the messaging but it makes sense to me to keep it at this level.
  3. Java and C++ – Language independence is cool and all but in practice if the protocol support Java and C++ we’re good to go. Maybe I’m being a bit myopic but my feeling is the likely hood that whatever we choose will expand to support more languages in the future is very high if it supports several today.
  4. Internal – We control the end points. I don’t care if the schema is external to the data package. In fact, for our use case that’s a plus. For any external services we’ll still expose those using the usual standards. Internally our applications will be using PB for their messaging format.
In short, we’re all about high volume low latency messages.

Protocol buffers are a flexible, efficient, automated mechanism for serializing structured data – think XML, but smaller, faster, and simpler. You define how you want your data to be structured once, then you can use special generated source code to easily write and read your structured data to and from a variety of data streams and using a variety of languages. You can even update your data structure without breaking deployed programs that are compiled against the “old” format.

Protocol buffers have many advantages over XML for serializing structured data. Protocol buffers:

  • are simpler
  • are 3 to 10 times smaller
  • are 20 to 100 times faster
  • are less ambiguous
  • generate data access classes that are easier to use programmatically
Seems like a decent fit. OK, actually an awesome fit. One of our developers has been doing some testing. It’s impressive.

Alternatives

JSON

To me protobuf feels like compiled JSON. They are very similar.  The main difference being JSON sends data over the wire in text format verses protobuf’s binary format. The latter has the advantage of a smaller size and being faster for a computer to parse.

ASN.1

Why not ASN.1? Seems like one of the best choices. Well understood and widely used. Sure the full ASN.1 specification is complex but we’d only need a small subset. I’m still struggling with this one a bit. Tool support seems a bit better in protobuf and it’s definitely simpler.

Thrift

Facebook’s Thrift is very similar to protobuf. Not surprising since the main author interned at Google. It’s a strong offering and recently became an Apache project. Nice stuff. Stuart Sierra has a nice comparison on his blog, Thrift vs. Protocol Buffers. Another worthy contender but not a big enough advantage to stop the internal momentum protobuf already has.

HDF5

The HDF wiki has an entry Google Protocol Buffers and HDF5 that concludes:

In summary, Protocol Buffers and HDF5 were designed to serve different kinds of data intensive applications: a network based transient message system, and a high performance data storage system for very large datasets such as multi-dimensional images, respectively. That said, both 1) offer open source technologies that can reduce data management headaches for individual developers and projects, 2) increase the ability to share data through the use of well-defined binary formats and supporting libraries that run on a variety of platforms, and 3) provide the ability to access data stored with “older” versions of the data structures.

Different design goals. HDF5 doesn’t fit our needs as well.

Hessian

The Hessian protocol has the following design goals:

  • It must not require external IDL or schema definitions, i.e. the protocol should be invisible to application code.
  • It must be language-independent.
  • It must be simple so it can be effectively tested and implemented.
  • It must be as fast as possible.
  • It must be as compact as possible.
  • It must support Unicode strings.
  • It must support 8-bit binary data (i.e. without encoding or using attachments.)
  • It must support encryption, compression, signature, and transaction context envelopes.

I still haven’t figured out how/if you can version your messages. Can you add and remove fields and still have compatibility (backwards and forwards)?  Cool effort, still feels very rough in places. Another worthy effort to consider.

ZeroC ICE

At it’s core it’s

The Ice core library. Among many other features, the Ice core library manages all the communication tasks using a highly efficient protocol (including protocol compression and support for both TCP and UDP), provides a flexible thread pool for multi-threaded servers, and offers additional functionality that supports extreme scalability with potentially millions of Ice objects.

ICE is a comprehensive middleware system. It can even use PB as it’s messaging layer. It’s messaging layer doesn’t handle adding or removing fields as well as PB. We don’t need the RPC side of ICE. Just not a good fit for us.

SDO

Service Data Objects provides a rather ambitious messaging architecture. It’s concerns aren’t speed and efficiency. The SDO V2.1 White Paper states

SDO is intended to create a uniform data access layer that provides a data access solution for heterogeneous data sources in an easy-to-use manner that is amenable to tooling and frameworks.

Interesting, not a fit.

Cisco Etch

Primary focus is an RPC implementation, not a messaging protocol. Steve Vinoski summarized it nicely in Just What We Need: Another RPC Package. In fairness Steve had some negative thoughts on PB also in Protocol Buffers: Leaky RPC. However, his concerns are around the undefined RPC features Google put in PB, not the IDL type aspects of PB.

Some other XML based protocol

Yeah, I know the problem with XML isn’t XML it’s with the parsers. Cute argument. Getting my message from native format on one system to native format on another as fast as possible is what matters to me. So oddly enough parsers are part of the equation. Yeah, jaxb is fast but just how fast?

Remember, we’re all about high volume low latency messages. It’s not a focus for XML. Yep, no one will take issue with that statement!

Binary XML

Enough said. Next.

Corba/IIOP

Well defined IDL, a bit complicated (because it addresses a wide range of issues).  Built in to the JDK! Not designed for speed or efficiency. Bad fit.

Conclusion

Several good choices. I’m sure there’s others I missed. We’re going with protobuf. Early tests by our developers have been very impressive. Google fan bois can rejoice and the Google haters gripe. In the meantime we’ve got a job to do.

Reference:

http://dewpoint.snagdata.com/2008/10/21/google-protocol-buffers/ (Original)

Read Full Post | Make a Comment ( 1 so far )

python thread.error: can’t start new thread

Posted on March 5, 2010. Filed under: Python | Tags: , , , |

The python has thread.error such as:

File "/usr/lib/python2.5/threading.py", line 440, in start
_start_new_thread(self.__bootstrap, ())
thread.error: can't start new thread

The “can’t start new thread” error almost certainly due to the fact that you have already have too many threads running within your python process, and due to a resource limit of some kind the request to create a new thread is refused.

You should probably look at the number of threads you’re creating(maybe in the /proc/pid/); the maximum number you will be able to create will be determined by your environment, but it should be in the order of hundreds at least. (Can try ulimit to solve this issue)

It would probably be a good idea to re-think your architecture here; seeing as this is running asynchronously anyhow, perhaps you could use a pool of threads to fetch resources from another site instead of always starting up a thread for every request.

Another improvement to consider is your use of Thread.join and Thread.stop; this would probably be better accomplished by providing a timeout value to the constructor.

Reference:

http://stackoverflow.com/questions/1834919/error-cant-start-new-thread

http://adywicaksono.wordpress.com/2007/07/10/i-can-not-create-more-than-255-threads-on-linux-what-is-the-solutions/

http://www.afnog.org/archives/2008-September/004535.html

http://rcsg.rice.edu/rcsg/shared/ulimit.html

http://answers.google.com/answers/threadview/id/311442.html

http://ubuntuforums.org/archive/index.php/t-114071.html

Read Full Post | Make a Comment ( 2 so far )

Liked it here?
Why not try sites on the blogroll...