比较 Cassandra 结构与关系数据库

Question

A few days ago I read about wide-column stored type of NoSQL and exclusively Apache-Cassandra.

What I understand is that Cassandra consist of:

A keyspace(like database in relational databases) and supporting many column families or tables (Same as table in relational databases) and unlimited rows.

From Stackoverflow tags:

A wide column store is a type of key-value database. It uses tables, rows, and columns, but unlike a relational database, the names and format of the columns can vary from row to row in the same table.

In Cassandra all of the rows (in a table) should have a row key then each row key can have multiple columns. I read about differences in implementation and storing data of Relational database and NoSQL (Cassandra).

But I don't understand the difference between structure:

Imagine a scenario which I have a table (or column family in Cassandra):

When I execute a query (CQL) like this :

select * from users;

It gives me the result as you can see :

lastname  | age  | city          | email               
----------+------+---------------+----------------------
      Doe |   36 | Beverly Hills | janedoe@email.com       
    Jones |   35 |        Austin | bob@example.com        
    Byrne |   24 |     San Diego | robbyrne@email.com         
    Smith |   46 |    Sacramento | null                    
   Jones2 | null |        Austin | bob@example.com       

So I perform the above scenario in relational database (MS SQL) with the following query:

select * from [users] 

And the result is:

lastname  | age  | city          | email               
----------+------+---------------+----------------------
      Doe |   36 | Beverly Hills | janedoe@email.com       
    Jones |   35 |        Austin | bob@example.com        
    Byrne |   24 |     San Diego | robbyrne@email.com         
    Smith |   46 |    Sacramento | NULL                    
   Jones2 | NULL |        Austin | bob@example.com       

I know that Cassandra supports dynamic column and I can perform this by using sth like:

ALTER TABLE users ADD website varchar;

But it is available in relational model for example in mssql the above code can be implemented too. Something like:

ALTER TABLE users ADD website varchar(MAX);

What I see is that the first select and second select result is the same. In Cassandra , they just give a row key (lastname) as a standalone object but it is same as a unique field (like ID or a text) in mssql (and all relational databases) and I see the type of column in Cassandra is static (in my example varchar ) unlike what it describes in Stackoverflow tag.

So my questions is:

1. Is there any misunderstanding in my imagination about Cassandra?!

2. So what is different between two structure ?! I show you the result is same.

3. Is there any special scenarios (JSON like) that cannot be implemented in relational databases but Cassandra supports? (For example I know that nested column doesn't support in Cassandra.)

Thank you for reading.

Answer 1

We have to look at more complex example to see the differences :)

For start:

• column family term was used in older Thrift API
• in newer CQL API, the term table is used

Table is defined as "two-dimensional view of a multi-dimensional column family".

The term "wide-rows" was related mainly to the Thrift API. In cql it is defined a bit differently, but underneath looks the same.

Comparing SQL and CQL. In SQL table is a set of rows. In simple example it looks like in CQL it is the same, but it is not. CQL table is a set of partitions, where each partition can be just a single row (eg when you don't have a clustering key) or multiple rows. Partition containing multiple rows is in Thrift therminology named "wide-row". To see how it is stored underneath, please read eg part about composite-keys from here .

There are more differences:

• CQL can have static columns which are stored on partition level - it seems that every row in partition have a common value, but really it is a single value stored on upper level. It can be used also to model 1:N relations
• In CQL you can have collection type columns - set, list, map
• Column can contain a user defined type (you can define eg address as type, and reuse this type in many places), or collection can be a collection of user defined types
• But also CQL does not support JOINs which are available in SQL, and you have to structure your tables very carefully, since they have to be strictly query oriented (in cassandra you can't query data by any column value, secondary indexes also have many limitations). It is usually said that in relational model you model tables clearly basing on data, when in cassandra you model basing on queries.

I hope I was able to make it a bit more clear for you. I recommend watching some vidoes (or reading slides) from Datastax Core Concepts Course as solid introduction to Cassandra.

Answer 2

In my experience CQL misleads a lot of people. First of all you would never want to do:

SELECT * FROM a_table_here; 

On a production Cassandra cluster, since you are putting a huge load on your Coordinator node to aggregate all of the data from all of the other nodes. Also by default, you will be given back a maximum of 10000 "rows".

To understand how Cassandra stores your data, we need to establish a few terms first:

There's the Primary Key, in your case lastname , this is hashed to determine which node in the cluster owns this range, and it's stored there (plus any replica nodes).

Next there's Cluster Columns, I don't know if you have any in your example, but you define them like PRIMARY KEY ((lastname),age, city) . In that example you are clustering by age first then city, this is ORDERED.

Now for a simplistic high-level view of Cassandra for your use case, it stores the data as a Map to an ordered Multimap:

Doe -> 36:Beverly Hills -> janedoe@email.com

Where 'Doe' is the Primary Key, which tells you which node(s) have that row of data. And 36:Beverly Hills is the Ordered Clustering Keys (part of the ordered multimap key). Lastly janedoe@email.com is the final value (can be multiple mind you) for the Map to a Multimap.

There's a lot of nuisances that I left out to make the example simple, for a more in-depth I would highly suggest reading: http://www.planetcassandra.org/making-the-change-from-thrift-to-cql/