Entries in nosql (8)


Oracle tables vs Cassandra SuperColumns


In my last post,  I wrote some Java code to insert Oracle tables into Cassandra column families.  As much fun as this was for me, it was  fairly trivial and not a particularly useful exercise in terms of learning Cassandra. 

In Cassandra,  data modelling is very different from the relational models we are used to and one would rarely convert a complete Oracle schema from tables directly to ColumnFamilies .  Instead, Cassandra data modelling involves the creation of ColumnFamilies with SuperColumns to represent master-detail structures that are commonly referenced together

SuperColumns vs Relational schema


At the Cassandra Summit in August,  Eben Hewitt gave a presentation on Cassandra Data Modelling.   There’s a lot of nuance in that talk and in the topic, but a key point in Cassandra – as in many other NoSQL databases – is that you model data to match the queries you need to satisfy,  rather than to a more theoretically "pure" normalized form.   For relational guys, the process is most similar to radical denormalization in which you introduce redundancy to allow for efficient query processing.

For example, let’s consider the Oracle SH sample schema.  Amongst other things, it includes SALES, PRODUCTS and CUSTOMERS:


9-09-2010 3-35-32 PM Oracle sample schema

We could map each Oracle table to a Cassandra ColumnFamily, but because there are no foreign key indexes or joins,  such a Cassandra data model would not necessarily support the types of queries we want.  For instance, if we want to query sales totals by customer ID, we should create a column family keyed by customer id, which contains SuperColumns named for each product which in turn includes columns for sales totals.  It might look something like this:

ID CustomerDetails Product Name #1 Product Name #2 ………….. Product Name #N
First Name Last Name
Guy Harrison
Quantity Value
3 $100,020
Quantity Value
3 $130,000
First Name Last Name
Greg Cottman
Quantity Value
34 $10,080
Quantity Value
4 $99,000


Each customer “row” has super column for each product that contains the sales for that product.  Not all customers have all the supercolumns - each customer has supercolumns only for each product they have purchased.  The name of the SuperColumn is the name of the product.  

Giving the column the name of the product is a major departure from how we would do things in Oracle.  The name of a column or SuperColumn can be determined by the data, not by the schema - a concept completely alien to relational modelling.

Inserting into SuperColumns with Hector


To try and understand this,  I created a Cassandra columnfamily of the type “Super”.  Here’s my definition in the storage-conf.xml file:

<ColumnFamily Name="SalesByCustomer" 
Comment="Sales summary for each customer "/>

And here is some of my Hector Java program, which reads sales totals for each customer from the Oracle sample schema, and inserts them into the ColumnFamily:

   1: private static void insertSales(Connection oracleConn, Keyspace keyspace,
   2:         String cfName) throws SQLException {
   3:     int rows = 0;
   4:     ColumnPath cf = new ColumnPath(cfName);
   5:     Statement query = oracleConn.createStatement();
   7:     String sqlText = "SELECT cust_id, cust_first_name,  cust_last_name, prod_name, "
   8:             + "           SUM (amount_sold) sum_amount_sold,sum(quantity_sold) sum_quantity_sold "
   9:             + "          FROM sh.sales    "
  10:             + "          JOIN sh.customers USING (cust_id) "
  11:             + "          JOIN sh.products  USING (prod_id)  "
  12:             + "         GROUP BY cust_id, cust_first_name,  cust_last_name,  prod_name "
  13:             + "         ORDER BY cust_id, prod_name ";
  14:     ResultSet results = query.executeQuery(sqlText);
  15:     int rowCount = 0;
  16:     int lastCustId = -1;
  17:     while (results.next()) { // For each customer
  18:         Integer custId = results.getInt("CUST_ID");
  19:         String keyValue = custId.toString();
  21:         if (rowCount++ == 0 || custId != lastCustId) { // New Customer
  22:             String custFirstName = results.getString("CUST_FIRST_NAME");
  23:             String custLastName = results.getString("CUST_LAST_NAME");
  24:             System.out.printf("%s %s\n", custFirstName, custLastName);
  25:             //Create a supercolumn for customer details (first, lastname)     
  26:             cf.setSuper_column(StringUtils.bytes("CustomerDetails"));
  27:             cf.setColumn(StringUtils.bytes("customerFirstName"));
  28:             keyspace.insert(keyValue, cf, StringUtils.bytes(custFirstName));
  29:             cf.setColumn(StringUtils.bytes("customerLastName"));
  30:             keyspace.insert(keyValue, cf, StringUtils.bytes(custLastName));
  31:         }
  32:         //Insert product sales total for that customer 
  33:         String prodName = results.getString("PROD_NAME");
  34:         Float SumAmountSold = results.getFloat("SUM_AMOUNT_SOLD");
  35:         Float SumQuantitySold = results.getFloat("SUM_QUANTITY_SOLD");
  36:         //Supercolumn name is the product name 
  37:         cf.setSuper_column(StringUtils.bytes(prodName));
  38:         cf.setColumn(StringUtils.bytes("AmountSold"));
  39:         keyspace.insert(keyValue, cf, StringUtils.bytes(SumAmountSold.toString()));
  40:         cf.setColumn(StringUtils.bytes("QuantitySold"));
  41:         keyspace.insert(keyValue, cf, StringUtils.bytes(SumQuantitySold.toString()));
  43:         lastCustId = custId;
  44:         rows++;
  45:     }
  46:     System.out.println(rows + " rows loaded into " + cf.getColumn_family());
  47: }

This code is fairly straightforward,  but let’s step through it anyway:

Lines Description
7-14 Execute the Oracle SQL to get product summaries for each customer
17 Loop through each row returned (one row per product per customer)
21 Check to see if this is a completely new customer
26-30 If it is a new customer,  create the CustomerDetails SuperColumn for that customer.  The SuperColumn name is “CustomerDetails” and it contains columns for Firstname and Lastname.

Now we create a SuperColumn for a specfic product, still keyed to the customer.  The SuperColumn name is set to the name of the product (line 37).  Inside the supercolumn are placed columns “AmountSold” (lines 38-39) and “QuantitySold” (lines 40-41)

Querying SuperColumns


Inserting master detail relationships into a supercolumn column family was easy enough.  I had a lot more difficulty writing code to query the data.  The tricky part seems to be when you don’t know the name of the SuperColumn you want to read from.  There's no direct equivalent to the JDBC ResultMetaData object to query the SuperColumn names - instead you create a "SuperSlice" predictate that defines a range of SuperColumns that you want to retrieve.  It's a bit awkward to express the simple case in which you want to return all the SuperColumns. 

Below is a bit of code which retrieves sales totals for a specific customer id.  I suspect I've made a few newbie mistakes :-):

   1: public static void querySuperColumn(Keyspace keyspace, String cfName,
   2:         String keyValue) {
   4:     ColumnPath colFamily = new ColumnPath(cfName);
   5:     System.out.println("Details for customer id " + keyValue);
   7:     /* Get Customer Details */
   8:     colFamily.setSuper_column(StringUtils.bytes("CustomerDetails"));
   9:     SuperColumn custDetailsSc = keyspace
  10:             .getSuperColumn(keyValue, colFamily);
  11:     for (Column col : custDetailsSc.getColumns()) {
  12:         String colName = StringUtils.string(col.getName()); 
  13:         String colValue = StringUtils.string(col.getValue()); 
  14:         System.out.printf("\t%-20s:%-20s\n", colName, colValue);
  15:     }
  16:     /* Get dynamic columns -  */
  17:     ColumnParent colParent = new ColumnParent(cfName);
  18:     SliceRange sliceRange = new SliceRange(StringUtils.bytes(""), StringUtils
  19:             .bytes(""), false, 2 ^ 32); // TODO: what if there are > 2^32 ??                                             
  20:     SlicePredicate slicePredicate = new SlicePredicate();
  21:     slicePredicate.setSlice_range(sliceRange);
  22:     //TODO:  Surely there's an easier way to select all SC than the above??
  23:     List superSlice = keyspace.getSuperSlice(keyValue,
  24:             colParent, slicePredicate);
  25:     for (SuperColumn prodSuperCol : superSlice) {  //For each super column
  26:         String superColName = StringUtils.string(prodSuperCol.getName());
  27:         if (!superColName.equals("CustomerDetails")) { // Already displayed
  29:             System.out.printf("\n%50s:", superColName); // product Name 
  30:             List columns1 = prodSuperCol.getColumns();
  31:             for (Column col : columns1) {               // product data 
  32:                 String colName = StringUtils.string(col.getName()); 
  33:                 String colValue = StringUtils.string(col.getValue()); 
  34:                 System.out.printf("\t%20s:%-20s", colName, colValue);
  36:             }
  37:         }
  38:     }
  40: }
Lines Description
8-9 Set the superColumn to the “CustomerDetails” supercolumn
11-14 Retrieve the column values (firstname, surname) for the CustomerDetails supercolumn
17-21 Set up a “SlicePredicate” that defines the supercolumns to be queried.  I want to get all of the supercolumns (eg every product), so I set up an unbounded range (line 18) and supply that to the slice predicate (line 21)
23 Create a list of supercolumns.  This will include all the SuperColumns in the column family (including, unfortunately,  CustomerDetails)
27 Eliminate CustomerDetails from the result.  Here we only want product names
30-35 Iterate through the columns in each supercolumn.  THis will extract QuantitySold and AmountSold for each Product name


Here’s some output from the Java program.  It prints out customer Details and product sales totals for customer# 10100:

Details for customer id 101000
customerFirstName :Aidan
customerLastName :Wilbur

CD-RW, High Speed Pack of 5: AmountSold:11.99 QuantitySold:1.0
Keyboard Wrist Rest: AmountSold:11.99 QuantitySold:1.0
Multimedia speakers- 3" cones: AmountSold:44.99 QuantitySold:1.0

SuperColumns with Toad for Cloud Databases 


Toad for cloud databases now has Cassandra support, which makes querying SuperColumns s a lot easier.  SuperColumns that have dynamic names but uniform internal column structure (as in my example above) are represented by Toad for Cloud Databases as a detail table.  To put it another way,  Toad for Cloud Databases re-normalizes the data - displaying it in the format that we would typically use in an RDBMS. 

So when we point Toad for Cloud databases at our SalesByCustomer column family, it maps the column family to two tables:  one for CustomerDetails and the other - which by default it will call SalesByCustomersuper_column” – for product sales totals.  We can rename the subtable and subtable key during the mapping phase to make it clearer that it represents product details.

9-09-2010 1-56-19 PM map cassandra super col

Now if we want to extract product details for a particular customer, we can do a SQL join.  Below we build the join in the query builder, but of course we could simply code the SQL by hand as we would for any NoSQL or SQL database supported by Toad for Cloud Databases:

9-09-2010 3-49-36 PM cassandra supercol qry

And just to close the loop, here we can see that the Toad for Cloud databases query returns the same data as the Hector query:

9-09-2010 3-50-48 PM cassabdra supercol results




All NoSQL databases require that we change the way we think about data modelling, and Cassandra is no exception.  SuperColumns are an incredibly powerful construct, but I can’t say that I found them intuitive or easy.  Hopefully APIs and tooling will evolve to make life easier for those of us coming from the relational world.


Playing with Cassandra and Oracle

Cassandra  is one of the hottest of the NoSQL databases.  From a production DBAs perspective it’s not hard to see why:  while some of the other NoSQLs offer more programming bells and whistles for the developer, Cassandra is built from the ground up for total and transparency redundancy and scalability, close to the heart of every DBA.

However,  Cassandra involves some complex data modelling concepts – mainly around the notorious SuperColumn concept, and I don’t think I’ll ever understand it fully until I’ve played directly with some data.  To that end, I thought I’d start by trying to model some familiar Oracle sample schemas in Cassandra.

Toad for Cloud Databases is releasing support for Cassandra early next month (eg September 2010), so I’ve been using that – as well as Java of course – to try to get some initial data loaded.

For other NoSQL databases,  Toad for Cloud lets us create NoSQL tables from relational tables with a couple of clicks.  Unfortunately, we can’t do that with Cassandra, since you can’t create a ColumnFamily on the fly.  So my first Cassandra tasks was to write a simple program to take an Oracle table (or query) and create a matching column family.

Getting started

Getting started with Cassandra was surprisingly easy.  I followed the instructions in http://schabby.de/cassandra-installation-configuration/ to install Cassandra on my laptop, and installed the hector Java interface from http://prettyprint.me/2010/02/23/hector-a-java-cassandra-client/.

Terminology in NoSQL can be confusing, with each NoSQL database using terms differently from each other, and all of them using terms differently from RDBMS.  In Cassandra:

  • A Keyspace is like a schema
  • ColumnFamily is roughly like a table

Things get very funky when SuperColumns are introduced, but lets skip that for now.

To create a ColumnFamily in Cassandra 0.6, we have to add its name to the storage-conf.xml file which is in the Conf directory and then restart Cassandra.  In 0.7 there’ll be a way to do this without restarting the server.

Here is where I created a keyspace called “Guy” and created some ColumnFamilies to play with:

   1: "Guy">
   2:   "G_Employees" CompareWith="UTF8Type"/>
   3:   "G_Employees2" CompareWith="UTF8Type"/>
   4:   "G_Employees3" CompareWith="UTF8Type"/>
   5:   org.apache.cassandra.locator.RackUnawareStrategy
   6:   1
   7:   org.apache.cassandra.locator.EndPointSnitch


Loading data


I wrote some Java code that takes a SQL statement, and loads the result set directly into a column family.  Here’s the critical method (the complete java program with command line interface is here):

   1: private static void oracle2Cassandra(Connection oracleConn,
   2:         Keyspace keyspace, String cfName, String sqlText)
   3:         throws SQLException {
   4:     int rows = 0;
   5:     ColumnPath cf = new ColumnPath(cfName);
   6:     Statement oraQuery = oracleConn.createStatement();
   7:     ResultSet result = oraQuery.executeQuery(sqlText);
   8:     ResultSetMetaData rsmd = result.getMetaData();
   9:     while (result.next()) { // For each row in the output
  10:         // The first column in the result set must be the key value
  11:         String keyValue = result.getString(1);
  12:         // Iterate through the other columns in the result set
  13:         for (int colId = 2; colId <= rsmd.getColumnCount(); colId++) {
  14:             String columnName = rsmd.getColumnName(colId);
  15:             String columnValue = result.getString(colId);
  16:             if (!result.wasNull()) {
  17:             cf.setColumn(StringUtils.bytes(columnName));
  18:                 keyspace.insert(keyValue, cf, StringUtils
  19:                         .bytes(columnValue));
  20:             }
  21:         }
  22:         rows++;
  23:     }
  24:     System.out.println(rows + " rows loaded into " + cf.getColumn_family());
  25: }

The method take s a Oracle connection and a SQL statement, and pushes the data from that SQL into the Cassandra column family and keyspace specified.   The first column returned by the query is used on the key to the Cassandra data.

Lines 6-8 execute the statement and retrieve a ResultSet object – which contains the data – and a ResultSetMetaData object which contains the column names.  Lines 9-21 just iterate through the rows and columns and create entries in the Column Family that match.   We use the Hector setColumn methodto set the name of the column and the insert method to apply the column value.  Too easy!

Of course, I’d have no idea as to whether my job had worked if I didn’t have Toad for Cloud databases available.  Using TCD, I can map the Cassandra columnFamily to a TCD “table” and browse the table (eg Cassandra Column Family) to see the resulting data:


I can even use SQL to join the Cassandra data to the Oracle data to make absolutely certain that the data transfer went OK:



It’s surprisingly easy to get started with Cassandra.  Installation of a test cluster is a breeze, and the Hector Java API is straight forward.    Of course,  direct mapping of RDBMS tables to Cassandra ColumnFamilies doesn’t involve the complexities of advanced Cassandra data models using variable columns and SuperColumns.    Next, I’m going to try and map a more complex ColumnFamily which maps to multiple Oracle tables – hopefully won’t make my brain hurt too much!

Toad for Cloud Databases is introducing Cassandra support in the 1.1 release due out within the next two weeks.  Its a free download from toadforcloud.com


Consistency models in Non relational Databases

One of the most significant differences between the new generation of non-relational (AKA NoSQL) databases and the traditional RDBMS is the way in which consistency of data is handled.  In a traditional RDBMS, all users see a consistent view of the data.  Once a user commits a transaction, all subsequent queries will report that transaction and certainly no-one will see partial results of a transaction.

RDBMS transactions are typically described as “ACID” transactions.  That is, they are:

  • Atomic: The transaction is indivisible – either all the statements in the transaction are applied to the database, or none are.
  • Consistent: The database remains in a consistent state before and after transaction execution.
  • Isolated: While multiple transactions can be executed by one or more users simultaneously, one transaction should not see the effects of other concurrent transactions.
  • Durable: Once a transaction is saved to the database (an action referred to in database programming circles as a commit), its changes are expected to persist.

As databases become distributed across multiple hosts,  maintaining ACID consistency becomes increasingly difficult.  In a transaction that spans multiple independent databases, complex two-phase commit protocols must be employed.  In the case of a truly clustered distributed database even more complex protocols are required, since the state of data in memory and the state of data in various transaction logs and data files must be maintained in a consistent state (cache fusion in Oracle RAC for instance).

CAP Theorem:  You can’t have it all


In 2000,  Eric Brewer outlined the CAP (AKA Brewer’s) Theorem.   Simplistically,  CAP theorem says that in a distributed database system, you can only have at most two of the following three characteristics:

  • Consistency: All nodes in the cluster see exactly the same data at any point in time
  • Availability: Failure of a node does not render the database inoperative
  • Partition tolerance:  Nodes can still function when communication with other groups of nodes is lost

Text Box-Quest Blue

Interpretation and implementations of CAP theorem vary,  but most of the NoSQL database system architectures favour partition tolerance and availability over strong consistency.

Eventual Consistency

A compromise between eventual consistency and weak (no guarantees) consistency is Eventual Consistency.

The core of the eventual consistency concept is that although the database may have some inconsistencies at a point in time, it will eventually become consistent should all updates cease.  That is,  inconsistencies are transitory:  eventually all nodes will receive the latest consistent updates.

BASE – Basically Available Soft-state Eventually consistent is an acronym used to contrast this approach with the RDBMS ACID transactions described above.

Not all implementations of eventually consistent are equal.     In particular, an eventually consistent database may also elect to provide the following:

  • Causal consistency:  This involves a signal being sent from between application sessions indicating that a change has occurred.  From that point on the receiving session will always see the updated value.
  • Read your own writes:  In this mode of consistency, a session that performs a change to the database will immediately see that change, even if other sessions experience a delay.
  • Monotonic consistency:  In this mode, A session will never see data revert to an earlier point in time.   Once we read a value, we will never see an earlier value.   


The NRW notation

NRW notation describes at a high level how a distributed database will trade off consistency, read performance and write performance.  NRW stands for:

  • N: the number of copies of each data item that the database will maintain. 
  • R: the number of copies that the application will access when reading the data item 
  • W: the number of copies of the data item that must be written before the write can complete.  

When N=W then the database will always write every copy before returning control to the client – this is more or less what traditional databases do when implementing synchronous replication.   If you are more concerned about write performance than read performance, then you can set W=1, R=N.  Then each read must access all copies to determine which is correct, but each write only has to touch a single copy of the data.

Most NoSQL databases use N>W>1:  more than one write must complete, but not all nodes need to be updated immediately.   You can increase the level of consistency in roughly three stages:

  1. If R=1, then the database will accept whatever value it reads first.  This might be out of date if not all updates have propagated through the system.   
  2. If R>1 then the database will more than one value and pick either the most recent (or “correct”) value.
  3. If W+R>N, then a read will always retrieve the latest value,  although it may be mixed in with “older” values.  In other words, the number of copies you write and the number of copies you read is high enough to guarantee that you’ll always have at least one copy of the latest version in your read set.   This is sometimes referred to as quorum assembly. 
NRW configuration Outcome
W=N  R=1 Read optimized strong consistency
W=1 R=N Write optimized strong consistency
W+R<=N Weak eventual consistency.  A read might not see latest update
W+R>N Strong consistency through quorum assembly.  A read will see at least one copy of the most recent update


NoSQL databases generally try hard to be as consistent as possible, even when configured for weaker consistency.  For instance, the read repair algorithm is often implemented to improve consistency when R=1.  Although the application does not wait for all the copies of a data item to be read,  the database will read all known copies in the background after responding to the initial request.  If the application asks for the data item again, it will therefore see the latest version. 

Vector clocks

NoSQL databases can seem simplistic in some respects, but there’s a lot of really clever algorithms going on behind the scenes.   For example,  the vector clock algorithm can be used to ensure that updates are processed in order (monotonic consistency).

With vector clocks,  each node participating in the cluster maintains an change number (or event count) similar to the System Change Number used in some RDBMSs.  The “vector” is a list including the current node's change number as well as the change numbers that have been received from other nodes.  When an update is transmitted, the vector is included with the update and the receiving node compares that vector with other vectors that have been received to determine if updates are being received out of sequence.    Out of sequence updates can be held until the preceding updates appear.

I found vector clocks hard to understand until I read the description in Operating Systems: Concurrent and Distributed Software Design  by Jean Bacon and Tim Harris (Addison-Wesley).

Amazon’s Dynamo

A lot of the eeventually consistent concepts were best articulated by Amazon in Verner Vogels’ Eventually Consistent paper and in Amazon’s paper on the Dynamo eeventually consistent key-value store.   Dynamo implements most of the ideas above, and is the inspiration for several well known NoSQL datastores including Voldemort and – together with Google’s BigTable specification – Cassandra.





Page 1 2