Querying the table was fast. What I didnít do was test the table under load, nor did I try this in a multi-node cluster.
As the number of columns in a row increases so does the size of the column index which is read as part of the read path. 

For background and comparisons of latency see http://thelastpickle.com/blog/2011/07/04/Cassandra-Query-Plans.html  or my talk on performance at the SF summit last year http://thelastpickle.com/speaking/2012/08/08/Cassandra-Summit-SF.html While the column index has been lifted to the -Index.db component AFAIK it must still be fully loaded.

Larger rows take longer to go through compaction, tend to cause more JVM GC and have issue during repair. See the in_memory_compaction_limit_in_mb comments in the yaml file. During repair we detect differences in ranges of rows and stream them between the nodes. If you have wide rows and a single column is our of sync we will create a new copy of that row on the node, which must then be compacted. Iíve seen the load on nodes with very wide rows go down by 150GB just by reducing the compaction settings. 

IMHO all things been equal rows in the few 10ís of MB work better. 

Cheers

-----------------
Aaron Morton
New Zealand
@aaronmorton

Co-Founder & Principal Consultant
Apache Cassandra Consulting
http://www.thelastpickle.com

On 11/12/2013, at 2:41 am, Robert Wille <rwille@fold3.com> wrote:

I have a question about this statement:

When rows get above a few 10ís  of MB things can slow down, when they get above 50 MB they can be a pain, when they get above 100MB itís a warning sign. And when they get above 1GB, well you you donít want to know what happens then. 

I tested a data model that I created. Hereís the schema for the table in question:

CREATE TABLE bdn_index_pub (
tree INT,
pord INT,
hpath VARCHAR,
PRIMARY KEY (tree, pord)
);

As a test, I inserted 100 million records. tree had the same value for every record, and I had 100 million values for pord. hpath averaged about 50 characters in length. My understanding is that all 100 million strings would have been stored in a single row, since they all had the same value for the first component of the primary key. I didnít look at the size of the table, but it had to be several gigs (uncompressed). Contrary to what Aaron says, I do want to know what happens, because I didnít experience any issues with this table during my test. Inserting was fast. The last batch of records inserted in approximately the same amount of time as the first batch. Querying the table was fast. What I didnít do was test the table under load, nor did I try this in a multi-node cluster.

If this is bad, can somebody suggest a better pattern? This table was designed to support a query like this: select hpath from bdn_index_pub where tree = :tree and pord >= :start and pord <= :end. In my application, most trees will have less than a million records. A handful will have 10ís of millions, and one of them will have 100 million.

If I need to break up my rows, my first instinct would be to divide each tree into blocks of say 10,000 and change tree to a string that contains the tree and the block number. Something like this:

17:0, 0, Ď/í
Ö
17:0, 9999, í/a/b/cí
17:1,10000, Ď/a/b/dí
Ö

Iíd then need to issue an extra query for ranges that crossed block boundaries.

Any suggestions on a better pattern?

Thanks

Robert

From: Aaron Morton <aaron@thelastpickle.com>
Reply-To: <user@cassandra.apache.org>
Date: Tuesday, December 10, 2013 at 12:33 AM
To: Cassandra User <user@cassandra.apache.org>
Subject: Re: Exactly one wide row per node for a given CF?

But this becomes troublesome if I add or remove nodes. What effectively I want is to partition on the unique id of the record modulus N (id % N; where N is the number of nodes).
This is exactly the problem consistent hashing (used by cassandra) is designed to solve. If you hash the key and modulo the number of nodes, adding and removing nodes requires a lot of data to move. 

I want to be able to randomly distribute a large set of records but keep them clustered in one wide row per node.
Sounds like you should revisit your data modelling, this is a pretty well known anti pattern. 

When rows get above a few 10ís  of MB things can slow down, when they get above 50 MB they can be a pain, when they get above 100MB itís a warning sign. And when they get above 1GB, well you you donít want to know what happens then. 

Itís a bad idea and you should take another look at the data model. If you have to do it, you can try the ByteOrderedPartitioner which uses the row key as a token, given you total control of the row placement. 

Cheers


-----------------
Aaron Morton
New Zealand
@aaronmorton

Co-Founder & Principal Consultant
Apache Cassandra Consulting

On 4/12/2013, at 8:32 pm, Vivek Mishra <mishra.vivs@gmail.com> wrote:

So Basically you want to create a cluster of multiple unique keys, but data which belongs to one unique should be colocated. correct?

-Vivek


On Tue, Dec 3, 2013 at 10:39 AM, onlinespending <onlinespending@gmail.com> wrote:
Subject says it all. I want to be able to randomly distribute a large set of records but keep them clustered in one wide row per node.

As an example, lets say Iíve got a collection of about 1 million records each with a unique id. If I just go ahead and set the primary key (and therefore the partition key) as the unique id, Iíll get very good random distribution across my server cluster. However, each record will be its own row. Iíd like to have each record belong to one large wide row (per server node) so I can have them sorted or clustered on some other column.

If I say have 5 nodes in my cluster, I could randomly assign a value of 1 - 5 at the time of creation and have the partition key set to this value. But this becomes troublesome if I add or remove nodes. What effectively I want is to partition on the unique id of the record modulus N (id % N; where N is the number of nodes).

I have to imagine thereís a mechanism in Cassandra to simply randomize the partitioning without even using a key (and then clustering on some column).

Thanks for any help.