Data Migration: Handling Data Types

Handling the structure, and deciding which fields, data types, or data points, relate to which fields or target elements within the destination database are one thing. Normally you can make an intelligent decision about the information that is being transferred and how that structural information should be handled.

The actual data can be a completely different problem. There are so many potential problems here with simply parsing and understanding the data that we need to look at some of the more common issues and how they should be addressed.

For most cases, the content and structure of the individual types is about understanding the two dimensions of the problem

  • Supported types – that is, whether the target database understands, or even identifies the underlying type. For example, within most RDBMS environments, most data types are very strictly enforced, but in NoSQL or datawarehouse environments they corresponding engines may either not care, or only care in specific situations.
  • Parseable information – the information may be readable or storable, but not to the same precision or definition. For example, with time values, some store hours, minutes and seconds, some store hours and minutes, and a whole variety store different lengths of fractions of seconds, to one or even 15 points of precision. At the end of the same spectrum, some dates are represented by seconds, others by a string in various formats.

Let’s look at the finer details of understanding these different types and how they can be stored, exchanged and ultimately parsed.

Basic Types

There are four basic types that make up all data and that, on the whole, get represented within a database system. Identifying these four types will help us to set the groundwork for how we handle, identify and treat the rest of the data:

  • Numeric – basically any kind of number or numeric value, from integers through to floating point and even ‘Big’ numbers.
  • Strings – Strings seem like a very simple list of characters, but they aren’t as straightforward as you might think.
  • Dates – Dates are a special type all of their own. Although they look like the worst parts of numbers and strings combined, ensuring that they are identifiable as a date in the target database relies on understanding the different formats, separators and structures.
  • Binary – any type of data that does not fall into the previous three groups is binary data. Here the issue is one of identity and the ability to interpret the content once it reachs the destination database.

Remember that in all of these situations, we are not only talking about one time formatting of the information. We could be dealing with frequent and regular exchanges of this information between the databases, and even need to perform these changes and differences regularly if the data is integrated across multiple database environments.

When combining, for example, MongoDB data with Oracle information for the processes of reporting, you need to do more than change the format once. It needs to be in a common representable format for both databases throughout the life of the data, while simultaneously ensuring that the information is best stored within each database to get the performance you need.

Strict and Relaxed Translation

Whenever you are moving data between different databases you need to determine whether the type and structure of information is important enough, or critical enough, that it must be represented in its native format.

That may sound like a completely nonsensical approach to data – surely the quality of the data is absolutely critical and should be represented as such everywhere? In theory it should, but different database storage environments treat and handle the data in different ways according to the basic type in question.

To understand why this is important, we need to look back both historically and technically why information was stored in the strict formats we described in the last section.

In any old, and particularly RDBMS-based database solution, data was stored into fixed types and with fixed lengths so that the record could be manipulated as efficient as possible. We saw some examples of this in Chapter 1. For numerical values, it is much more efficient to store a 32-bit integer as just 4 bytes of data than it is to store the string 2147483647 (which would take 9 bytes).

Similarly, with string types, the primary consideration has always been to minimize the amount of storage reserved for the string because handling bigger strings, or bigger potential blocks for strings, was more expensive in computing time, memory space, and disk space. Back when databases ran on machines with 512KB of RAM, devoting massive blocks of memory to non-usable space allocated but not used to store data just wasn’t an option. This is why 8 character filenames and two or three letter codes for a variety of databases and storage methods became common.

In modern systems of course, we actually have almost the opposite problem. Data sizes are now regularly so large that we need to be prepared to handle massive blocks of information whereas before that might have been impossible. This is fine when we are moving data from a traditional RDBMS to say Hadoop, because we move from a strict environment to a very relaxed one. But when moving in the opposite direction this is not true.

To make matters worse, in many Big Data environments, including most of the Hadoop database layers like Hive, the datatype is only significant at the time the data is queried. Within Hive you can load a CSV file that contains a variety of different types, but unless you explicitly tell Hive that the fifth column is a string or a number, Hive doesn’t really care, and let’s you export and query the data as normal.

For example, here’s a table with a mixture of different column types, here using the strict datatypes to define the actual content for each column:

hive> select * from stricttyping;
1      Hello World   2014-10-04 12:00:00   439857.34
1      Hello World   2014-10-04 12:00:00   157.3457
1      Hello World   2014-10-04 12:00:00   4.8945796E7

Now we can view the same data, this loaded into a table where each column has been defined as a string:

hive> select * from relaxedtyping;
1      Hello World   2014-10-04 12:00:00   439857.345
1      Hello World   2014-10-04 12:00:00   157.3457
1      Hello World   2014-10-04 12:00:00   48945797.3459845798475

The primary differences are in the handling of floating point values – the top strict table loses precision (the value was a FLOAT), and at the bottom the value is represented as a DOUBLE with a loss of precision digits. In fact, within Hive, the data is not parsed into the corresponding type until the data is used or displayed. If you examine the raw CSV file:

$ hadoop fs -cat stricttyping/sample_copy_1.csv
1,Hello World,2014-10-04 12:00:00,439857.345
1,Hello World,2014-10-04 12:00:00,157.3457
1,Hello World,2014-10-04 12:00:00,48945797.3459845798475

In fact, many people deliberately don’t explicitly load the data into fixed type columns; they define the column types as strings and then import the date and ultimately ignore the real type until they have to parse or understand it for some reason.

Similarly, in NoSQL environments, the data types may really only be for explicitly representation requirements, and have no effect on the ability to either store or display and query the information. Even in a traditional RDBMS, there is no requirement to explicitly store certain values in certain column types, but certain operations may be limited. For example, most RDBMSs will not perform a SUM() operation on a string column.

The bottom line is that you will need to think about whether to explicitly make use of these columns because you need them as specific types in the target database, or whether to ignore them completely.

  • Strict transformations – Use strict datatypes when the information you want to store must be correctly interpreted within the target database, and it provides some form of performance advantage, unless doing so reduces the validity or precision of the information.
  • Relaxed transformations – Use relaxed transformations whenever the processing or target system does not support the required precision, or in those cases where the processing of the information is irrelevant. Most t ransfers to NoSQL and Big Data environments fit this model automatically.

With this options to you in mind, let’s look at some of the more specific types available.

Handling Numeric Values

Simple, plain, integers are supported by nearly all databases as explicit and identifiable types. Even document databases such as MongoDB and Couchbase understand the significance of a numeric value over a string representation.

However, if you are transferring big integers, be conscious of the limitations of the target database. Some environments explicitly support very large integers. Hive, for example, supports the BIGDECIMAL datatype, which holds numbers with up to 10 to the power of 308. Others do not.

Floating Point Issues

The biggest problem with floating point values is one of precision and storage capability. There are large variations between the supported types, how much is stored and how precise it can be. Further more, some databases specifically differentiate between decimal and floating point values and have different rules for how these should be represented and stored, and the two are not necessarily compatible

For floating-point values, the main issues are:

  • Representation – float values are generally displayed as a decimal value, for example:
  • There are no specific rules for this, but many values are based on the support of the operating systems own data type support. On modern 32-bit (and 64-bit) systems, floating-point values tend to have 7 digits of precision after the decimal point. This is due to the nature of the structure used to store and define them internally. A double has twice the precision, up to 15 or even 16 digits past the decimal point.
  • Parsing – these values properly is critical if you are storing the data; unfortunately rounding-errors, both made when the data is output, and when it is parsed back, are notoriously difficult, and not always religiously honoured.
    For this reason, some database explicitly support a DECIMAL type. Unlike the float, the DECIMAL type works more like two integers either side of the decimal.

Processing these values reliably, and storing them in a target database may lead to problems if the target system doesn’t support the datatype size, precision, or structure properly. Moving the data may lose the precision or content. On a simple movement of the data in an export/import environment might parse or store it correctly, or it may lose or truncate the precision entirely.

If you are replicating and regularly exchanging data from one database to the other and back again, these precision errors can build up to translate and convert a number from one value to one statistically significant.  If the double type within the databases environment does not support the complexity or precision of the values involved, consider using one of the big integer types and a suitable multiplier.

Finally, if the target database does not support the precision and detail that you need, consider moving the data using relaxed methods, for example by importing the data into a string, rather than a numerical type so that it can be reliabily stored.

Base-N Numbers

If you are exchanging numbers in other than base 10, for example, octal, hexadecimal, or others, ensure that the target database supports the required number format. If an explicit number format is not supported, either translate the number to decimal and handle the output and formatting of the data as the corresponding type within the target database and application, or use the relaxed method and keep it as a string.

Strings and Character Encoding

More problems are created by strings than you might think, and the most significant is usually the character set used to store, transfer, and import the data. Character sets used to refer to the difference between the byte-level encoding for things like EBCDIC and ASCII. Today, they span a much wider array of issues as the number of character sets and the use of a wider range of languages, characters, and ideographs increases.

The best way to encode strings when moving the data between databases is to use either UTF-8 (which encodes Unicode character in 8-bit bytes) or one of the high-bitrate encodings if your data requires it. For example, if you are specifically storing foreign-language, katana, or Chinese characters, using UTF-16 or UTF-32 may be more reliable, if not necessarily more efficient. UTF-8 can be used for a very wide range of different Unicode characters and is rarely a hindrance.

Also be aware that some databases identify character encoding capabilities and data types differently. For example, the VARCHAR2 type within Oracle can be used to store strings with an optional size (byte or character) declaration, but the NVARCHAR2 type is the Unicode (byte) sized datatype. The definition of the column and size can also be different. In Amazon RedShift for example, the size of VARCHAR column is defined in bytes, but in MySQL it’s defined in characters, so a VARCHAR(20) in MySQL has to be a VARCHAR(80) in RedShift. Sneaky.

A secondary issue is one of storage space. Different database environments support different representations of the different character storage types, and sometimes have wildly different performance characteristics for these different types.

Within a NoSQL or Big Data environments, the length (or size) of the string is rarely a problem, as they don’t have fixed or strict datatypes. However, for most RDBMS environments there are specific lengths and limits. Oracle supports only 4000 bytes in VARCHAR2 for example; MySQL supports 255 bytes in a CHAR, or 65535 bytes in a VARCHAR.

Finally, when transferring the information you may need to pay attention to any delimiters. Using CSV, for example, and using quotes to define the field limits only works when there aren’t quotes in the field content.

Dates and Times

Of all the different data types that we have covered up to now there have been problems with understanding and parsing the values because of differences in the types, format, or structure of the data, but all of them were largely covered within some simple limits and structure.

Dates are another problem entirely. In particular:

  • Date precision and format
  • Time precision and format
  • Dates or Epochs?
  • Time Zones

All go together to make for one of the most complicated of the all the types supported when transferring data, because there are so many times where it can go wrong.


Epoch values are those where the data is represented as an integer counting, usually, the seconds from a specific reference point in time, from which the current date can be calculated. For example, Unix-based Epoch times are represented as the number of seconds that have elapsed since Jan 1st 1970 at 00:00:00 (12:00am) GMT. Counting forward from this enabels you to represent a date. For example, the value:


Is in fact 12th October 2014.

There are two issues that arise from Epoch dates, time drift and date limits.

Time drift occurs if the date has been stored as an epoch that is relative to the current timezone. This can actually happen more frequently than you realize if dates are reconstituted back to an Epoch from a local time based balue into an Epoch. For example, some libraries that parse a date without an explicit timezone will assume that the date is within the current timezone of the system.

This is a particularly thorny problem when you realize that epochs have no timezones of their own. This means that the Epoch value:


Is 15:56 BST, but 07:56 PST. If you now transfer a PST-based epoch to GMT and then use it without conversion, all your times will be out by 8 hours. If you ran batch jobs on the imported data at 1am, that time would actually refer to a completely different day.

If you must use epoch values, ensure that you either know what the timezone was, or adjust the value so that it is against GMT and you can translate to the appropriate timezone when you need to. Also see the secion on timezones below.

The secondary problem is date limits. Traditionally epoch values were stored as 32-bit integers, which limits the date between 1970 and 2038. While this is fine for current times (at least for the next 24 years or so), for any future dates, this can be an issue.

If you are porting epoch values to a target that only supports 32-bit dates, and the date is beyond 2038, don’t transfer it using the epoch value, translate it into an explicit date that can be parsed and stored in whatever local format is required for the target environment. For example, within MySQL you can use the FROM_UNIXTIME() function to translate your epoch date into something more usable.

Date Formats

When transferring dates, use a format that is unambiguous and supported by the target system. Different locations and systems have different ways of representing dates, including the different seaprators that are used, and the different orders of the components. Even the use of the prefix for some numbers differs between regions. Some examples are shown in the table below.

Location/Format Example
USA Month.Day.Year
Japan Year-Month-Day
Europe Day.Month.Year
UK Day/Month/Year

Different locations and date formats

The best format to use is usually the ISO format:


With a zero prefix added to each value to pad it to the correct number of characters. For example, the the 1st of January:


Or the year 1:


The ISO format is not only readable on just about every single platform, it also has the advantage of being sortable both numerically and by ASCII code, making a practical way of exporting and loading data in date order without having to explicitly order data by dates.

Time Formats

Time is usually restricted to a fairly obvious format, that of:


Or in some regions and standards:


Aside from the timezone issue, which we will look at next, the other problem is the level of precision. Some databases do not support any precision beyond seconds. For example, within Oracle you can store precision for eseconds up to 9 decimal points. Amazon RedShift supports only 6 digits of precision.

Also be aware that some environments may not support explicit date and time types, but only a unified datetime or timestamp type. In this case, the structure can be even more limited. For example, within Amazon RedShift, the timestamp datatype is actually formatted as follows:


With the date in ISO format but without explicit date separators.

Time Zones

Every time represented everywhere is actually a time within a specific timezone, even if that timezone is UTC (Universal Time Coordination). The problem with timezones is that the timezone must either be explicitly stored, shared, and represented, or it should be stated or understood between the two systems that the time is within a specific known timezone. Problems occur either when the timezone is correctly represented, or assumptions are made.

For example, the following time:

2014-09-03 17:14:53

Looks clear enough. But if this has come from the BST (British Summer Time) timezone and gets imported into a database running in the IST (India Timezone) then you start to get the time stored in the wrong format if the timezone is not explicitly specified.

Another issue is when there are timezone differences when data is transferred, not because of the physical time difference, but because of the effect of daylight savings time. Transferring data from, say, GMT to PST is not a problem if you know the timezone. Transfer the data over during a daylight savings time change, and you can hit a problem. This is especially true for timezones that have different dates for when daylight savings time changes.

Finally, be prepared for databases that simply do not support the notion of timezones at all. To keep these databases in synchronization with other databases with which you might be sharing information, the easiest method is to use GMT.

In general, the easiest solution for all timezone related data is to store, transfer, and exchange the data using the UTC timezone and let the target database handle any translation to a localized timezone. If you need to explicitly record the timezone – perhaps because the data refers to a specific timezone as part of the day – then use the time type that supports it, or store a second field that contains the timezone information.

Compound Types

We’ve already looked at some of the issues in terms of the structural impact of compound types. Even if you have the ability to represent a value as a compound structure within your target data, you need to understand the limitations and impact of compound types, as not all systems are the same. Some of these limitations and effects are database specific, others are implementation specific.

For example, within MySQL and PostgreSQL, the ENUM type enables you to store a list of fixed string-like values that can be chosen from a fixed list. For example:


The benefit of this from a database perspective is that internally each string can be represented by a single number, but only reconstituted into the string during output. For targets that do not support it, therefore, the solution is to translate what was an ENUM column in MySQL into a string in the target database.

MySQL also supports the SET type, which is similar to ENUM, except that the value can refer to muiltiple options. For example:


The SET type enables you to record not only the specific day, but maybe a group of days, for example:

INSERT INTO table VALUES (‘Mon,Wed,Fri’)

Again, interally this information is represented this time as a BIT, and so the actual data is implied and displayed as a compound type.

When translated to a database that doesn’t support the type, you may either want to create an additional column for each value to store it, or, if you are using a document database, you may want the set firled converted to an array or hash of values:

    ‘Mon’ => 1,
    ‘Wed’ => 1,
    ‘Fri’ => 1

Always consider how the data was used and will be searched on either side of the database transfer. In an SQL RDBMS queries normally take the form:

SELECT * FROM TABLE where FIND_IN_SET(‘Mon’,days)>0;

That is, return all the values where the field contains the value ‘Mon’. In a database that supports searching or indexing on individual values (MongoDB, Couchbase), the key-based transfer, where we effectively set the member of a hash to a meaningless value so that we can do key-based lookups. We’ll examine this in more detail when we examine the rnvironments of these databases.

Serialized and Embedded Formats

For a whole variety of different reasons, some people store more complex formats into their database so that they can bmore easily be manipulated and used within the depper element sof their application.

For example, serializing an internal structure, for example, a Perl object or a Java object so that it can be stored into a field or BLOB within the database is a good way of making use of complex internal structures and still have the ability to store and manipulate the the more complex data within the application environment.

If all you want is to transfer these the serialized format from one database to another, then the basics are unlikely to change. You may need to use the binary translation methods in the next section to realisitically get the data over into the new database reliably, but otherwise, the transfer should be straightforward.

However, there is also the possibility that you want to be able to query or extract dta that may have been embedded into the serialized object.

In this case, you need to change the way that you use and manipulate the information as part of the data migration process. In this case, you may want to take the information and either expand the data to expose the new fields as transferrable data.

Or, you may more simply want to change the content of the information from its serialized format into a more universal format, such as JSON.

Binary and Native Values

Binary data, that is, data that is explicitly stored and represented in a binary format is difficult to process when moving data.

  • Binary means that single-character delimiters become useless, even control characters. 0x01 is just as likely to come up in binary data as it is when used as a field separator.
  • Pure, native, binary data suffers from the problems of ‘endianess’, that is, the byte order of the individual bytes. Test and numerical translations don’t tend to suffer from this because systems know how to parse text. When exchanging binary data, the endianness of the data applies.

Binary data can also be affected by any translation or migration process that is expecting a string representation of information. For example, it is not uncommon for UTF8 to be used when reading binary data, which leads to interpretation and storage problems.

In general, the best advice for true binary information is for the data to be encoded into one of the many forms of binary-to-hex translation formats. This can include solutions such as raw hex conversion, where the data is quite literally expanded to a two-character hex string for each binary byte. For example, we can translate any byte strinf into hex values with tools like Perl:

$ perl -e "print unpack('H*','Hello World')"

Or use uunencode:

begin 666 HelloWorld

Or use the MIME64 standard that is employed in many modern email and Web environments for transferring attachments, as it ensures that even multi-byte cahracters are effectively transferred.

All of these solutions can be easily processed on the other side back into the binary format according to the endianess of the host involved.

Data Migration: Mapping the Data

When moving the data between different databases the primary considering is what that’s going to look like so that it can be used in the target environment. Later chapters are going to dig deeper into this topic, but let’s fly over some of the key considerations here.

Mapping Columns to Tables

If we were replicating this data from our existing RDBMS into another, the most obvious method is for us to simply move the tables wholesale from one environment to the other. If they both support table structure, then there is no reason not to duplicate this structure on the other side.


But, always be conscious of how the data is going to be handled over on the other side. If your target database does not support joins between tables, as some Hadoop alternatives do not, then you will need to determine whether you are better to merge the table data together, either into a bigger table and either duplicate the information, or hide it.

For example, if you wanted to analyse which recipe has the most contents made of chicken, then you could combine the tables together from the transactional side into a pre-joined table that contains the detail. For example, converting our three tables, ingredients, recipe ingredients, and recipes, into something like this:

|     3006 | Tabboule            | <olive oil>         |
|     3006 | Tabboule            | <salt>              |
|     3006 | Tabboule            | <onion>             |
|     3011 | Minted pea puree    | <olive oil>         |
|     3011 | Minted pea puree    | <frozen peas>       |
|     3011 | Minted pea puree    | <seasoning>         |
|     3011 | Minted pea puree    | <butter>            |
|     3012 | Lamb patties        | <ground coriander>  |
|     3012 | Lamb patties        | <ground cumin>      |
|     3012 | Lamb patties        | <ground turmeric>   |

Now we have the information from the ingredients table merged with the recipe ID and title from the souce database. Over in our destination store, performing a count or summation operation will now be an easier way to enable us to do the query. In fact, with a single table structure like this some operations are quicker and provide the information we want. For example, find all the recipes with frozen peas in them is a single (possibly indexed) table:

| recipeid | title              | description   |
|      984 | Waffle fish pie    | <frozen peas> |
|      633 | Vegetable korma    | <frozen peas> |
|       27 | Spicy tuna stew    | <frozen peas> |
|     1261 | Seafood paella     | <frozen peas> |
|      902 | Choux au gratin    | <frozen peas> |
|      866 | Tomato baked rice  | <frozen peas> |
|     1971 | Spicy risotto star | <frozen peas> |
|     2741 | Cheat's jambalaya  | <frozen peas> |
|     2750 | Spicy sausage rice | <frozen peas> |
|     2778 | Quick jambalaya    | <frozen peas> |
|     3011 | Minted pea puree   | <frozen peas> |

In a columnar store this can be orders of magnitude faster than a join across the three source tables, and still provides us with the core of information we want to display – the recipe id and title.

Mapping Columns to Documents

Moving the data over verbatim as tables is unlikely to work as efficiently as you think. For example, in a NoSQL database, joins are normally either impossible, or computationally expensive, and so expecting to be able to reconstitute the structure in the destination database. You also want to make sure that you are using the most efficient method for the database system. Simply putting a JSON representation of each row from an existing table or column store is probably not going to work, especially as JOINs are typically unavailable. A better solution is merge the data from multiple tables itno a single document that contains all of the information you need.

Doing this for table data to be inserted into a document store normally requires the composition of a document from the constituent elements, in the case of our recipe database, the recipe, method, ingredients, and nutritional information needs to be combined together into one big document. There are a variety of ways to do this, but an obvious solution is to logically group ‘objects’ together. That is, an object that might be represented by a collection of tables. Like this:


Within our recipe data, for example, in a document store the use case is for us to extract or remove the entire recipe – base data, ingredients, and methods – as a single document that contains all the information we need. This puts all of the information in one document, and makes it easy to update and format as that entire recipe at a time. We can actually see a sample of this, first by looking at the diagrammatic example, here with some dummy data, but you can see how the tables on the right can be mapped to fragments of the document on the left.

Document and Table Mapping

We can also look at a simple script that performs this operation for me, here collecting the recipe object (which queries the underlying database) and then converting that into a JSON structure for writing into the database:

use JSON;
use lib '';
use Foodware;
use Foodware::Public;
use Foodware::Recipe;

my $fw = Foodware->new();

my $recipes = $fw->{_dbh}->get_generic_multi('recipe','recipeid',{ active => 1});

foreach my $recipeid (keys %{$recipes})
    my $recipe = new Foodware::Recipe($fw,$recipeid,{ measgroup => 'Metric',
                                                 tempgroup => 'C',});

    my $id = $recipe->{title};
    $id =~ s/[ ',\(\)]//g;
    my $record = {
       _id => $id,
       title => $recipe->{title},
       subtitle => $recipe->{subtitle},
       servings => $recipe->{servings},
       cooktime => $recipe->{metadata_bytag}->{totalcooktime},
       preptime => $recipe->{metadata_bytag}->{totalpreptime},
       totaltime => $recipe->{metadata_bytag}->{totaltime},
       keywords => [keys %{$recipe->{keywordbytext}} ],
       method => $recipe->{method},

    foreach my $ingred (@{$recipe->{ingredients}})
               meastext => $ingred->{'measuretext'},
               ingredient => $ingred->{'ingredonly'},
               ingredtext => $ingred->{'ingredtext'},

    print to_json($record);

Finally, let’s look at how this is all translated into a full JSON representation of the same information:

   "subtitle" : "A good use for bananas when they're going soft.",
   "keywords" : [
      "occasion@kids' parties",
      "special collections@lunchbox",
      "meal type@cakes, biscuits, sweets",
      "special collections@classic recipe",
      "cook method.hob, oven, grill@oven",
      "special collections@store cupboard",
      "special collections@budget",
      "occasion@prepare-ahead entertaining",
      "main ingredient@fruit",
      "special collections@cheffy recommended"
   "preptime" : "20",
   "servings" : "8",
   "cooktime" : "45",
   "method" : [
         "_sort" : "4",
         "recipeid" : "2035",
         "step" : "4",
         "altgroup" : "0",
         "methodstep" : "Spoon into the loaf tin. Spoon the top. Bake for 45-50 min or until well risen and cooked through. ",
         "text_formatted" : "Spoon into the loaf tin. Spoon the top. Bake for 45-50 min or until well risen and cooked through. "
         "text_formatted" : "Slowly beat in the egg. Add the banana. Fold in the flour and bicarbonate of soda. Add the dried fruit. ",
         "methodstep" : "Slowly beat in the egg. Add the banana. Fold in the flour and bicarbonate of soda. Add the dried fruit. ",
         "_sort" : "3",
         "recipeid" : "2035",
         "altgroup" : "0",
         "step" : "3"
         "recipeid" : "2035",
         "_sort" : "2",
         "step" : "2",
         "altgroup" : "0",
         "text_formatted" : "Cream the butter and sugar together until pale and fluffy. ",
         "methodstep" : "Cream the butter and sugar together until pale and fluffy. "
         "recipeid" : "2035",
         "_sort" : "1",
         "altgroup" : "0",
         "step" : "1",
         "text_formatted" : "Preheat oven to 180&deg;C. Grease a 900 g loaf tin.",
         "methodstep" : "Preheat oven to 180.C. Grease a 900g loaf tin."
         "text_formatted" : "Turn out onto a wire tray. Leave to cool. ",
         "methodstep" : "Turn out onto a wire tray. Leave to cool. ",
         "_sort" : "5",
         "recipeid" : "2035",
         "altgroup" : "0",
         "step" : "5"
   "totaltime" : "65",
   "_id" : "Bananacake",
   "title" : "Banana cake",
   "ingredients" : [
         "ingredtext" : "butter",
         "meastext" : "75 g",
         "ingredient" : "butter"
         "ingredtext" : "bicarbonate of soda",
         "meastext" : "[sup]1[/sup]/[sub]2[/sub] tsp",
         "ingredient" : "bicarbonate of soda"
         "meastext" : "2",
         "ingredient" : "bananas",
         "ingredtext" : "ripe bananas, peeled and mashed"
         "ingredtext" : "white self-raising flour, sifted",
         "ingredient" : "white self-raising flour",
         "meastext" : "200 g"
         "ingredtext" : "salt",
         "meastext" : "1 pinch",
         "ingredient" : "salt"
         "ingredtext" : "egg, beaten",
         "ingredient" : "egg",
         "meastext" : "1"
         "ingredient" : "dried mixed fruit",
         "meastext" : "100 g",
         "ingredtext" : "dried mixed fruit"
         "ingredtext" : "caster sugar",
         "ingredient" : "caster sugar",
         "meastext" : "100 g"

Looking at the output, you can see how the structure of document has been merged together into something more usable. We have a block for ingredients, keywords, method, and you can see how the ‘ingredient’ field in the ingredient block could be used as a searchable element.

The primary questions about the format come will come down to how the data will be used. In CouchDB and Couchbase for example a map/reduce like process will be used to create an index on the information. Choosing the right structure is therefore about understanding the structure and how it will be used. When I come to build an index on this information, and I want to build an index so that I can query by ingredient, is this an effective method to format the data? Is processing all of that data to determine the occurrences of turkey the most efficient method?

How about if I realize that all my vegetarian recipes are really vegan, will that change the structure? We’ll return to these questions later in the book.

Handling Compound Assignments

While we’re talking about constructing documents from tables and vice versa, I want to consider those pesky compound values again. Compound values are hugely complex when it comes to data translations because there are differences within a database type as well as between types that should be considered. Always with a compound type you should start by asking three basic questions:

  • How will it be queried in the target database?
  • Can I have multiple values within the same field?
  • Is the order of those multiple values significant?

One translation you want to avoid is to convert this relatively structured format into something that ultimately becomes hard to process. For example, the temptation is to convert this ‘field’ from the source CouchDB environment into a format that looks similar to the original, for example, by using a comma-separated list:


There are so many dangers with this, it’s hard not to see how this is a bad idea. Some obvious problems are:

  • How do we cope with an ever-increasing number of potential options? We’re showing just four here, what if there were 20? 30? 60?
  • What happens if more data than the width of the column are defined? If it’s truncated we lose information. Using a wider column only works for so long.
  • Indexes become unmanageable from a database adminisitration perspective.
  • How is integrity enforced? Should the options be sorted alphabetically? What happens when they get updated?
  • What happens if we want to change a value? What happens if ‘red’ becomes ‘scarlet’, do we change every row?

The biggest problem with this approach is the usability once that value is used in a transactional or columnar store, the data now becomes difficult and potentially expensive to process. It takes a lot more computational effort for a database to search even with the use of indexes a text string that contains a sequence of letters, against either a join or boolean set of columns matching the same structure.

You can actually test this quite effectively by creating sample tables that emulate this basic functionality. MySQL is being used here to do the comparisons, but the general effects are quite clear. Here’s a search against a single text column using commas:

SELECT id,title FROM recipes WHERE keywords LIKE '%CHICKEN%';

Now here’s the same query where the colours are represented a boolean columns:

SELECT id,title FROM recipes WHERE kw_chicken = 1;

A boolean index in the latter example will be much quicker than the LIKE search in the former.

Adventures in recipes

When developing an application, and particularly the database that will support it, there comes a time when you realize you may not have planned and identified all of the problems you could, and that’s when you notice that something doesn’t quite work.

The use of free-text for ingredients in the recipe database was one such moment. Searching for a recipe that contains ‘Chicken’ is fine if you look at the ingredients, you get to pick up everything from ‘whole chicken’ to ‘chicken breasts’ within the search. Unfortunately, you also pick up ‘chicken stock’. When a user searches for chicken recipes, chicken stock is used in a surprising number of recipes that otherwise contain no chicken of any variety whatever.

When migrating data around, you can see the same principles at work; if you’d relied on using a text field to store the value over separate, stricter, fields, the quality of the data is ruined. Keep this in mind when moving data around.

The recommendations for how to work out which is the best method are actually comparatively straightforward:

  • If the list of possible values is small (for example, four or five different values) and there are few of them within the overall ‘table’ that they are been moved to, use fields/boolean values.
  • If the list is larger, and likely to grow and expand with new options, then use a lookup table.

When applying the same principles the other way round, you should use the most appropriate format for the corresponding target database in a way that makes the data usable. In document databases, for example, it’s efficient to translate a compound value into an array of sub-values, just as in our original.

Mapping Documents to Columns

When mapping data from a document-based structure back into columns, the process is significantly harder. You are specifically going from a flexible multi-field format where the number and type of fields could be entirely different into one where the format is rigid and inflexible.

There are two choices available:

  • Transform the data back into a table structure that as close as possible matches the overall data structure of the underlying information. This is great if the document is relatively flat. But if, as we’ve seen, we have compound data, or variable length compound information, this method doesn’t work as well.
  • Transform the data back out into a multi-table format that has the capability for joins. This is basically the reverse of the process we just examined for converting the table-based recipe data into documents. You must remember to use a unique identifier for the parent record so that it can be linked back properly when a JOIN is used.

The flatter the document, the easier the conversion.

Data Migration: Moving the Actual Data

There are two key elements to the exchange of any information between databases. One is the data structure used for the exchange, and the other is the transformation required to reach those structures.

Some of these are driven by the source database, others by the target database. For example, when moving data from RDBMS to NoSQL database generally requires constructing documents from what might be tabular, or joined-tabular data. This may involve both join elements on the relational side, as well as formatting on the NoSQL side. The eventual aim is to ensure that the data reaches the target database in both the right format, and without corruption, and also in a format that is most appropriate or efficient. That ultimately depends on what you are using the transferred data for.

The other aspect is the difference between source and target data types – that is the format and construction of the individual fields or components of the data. Document databases and Big Data stores tend not to care about the data type, whereas RDBMS cannot live without them.

Some important considerations for how we use this information:

  • The data structure must be retained (i.e., we must be able to identify the fields, columns or other elements of the data).
  • The data format and integrity must be maintained (the data should not be corrupted, shortened or reduced in any way).
  • The data must be able to be efficiently transferred (sending a 1GB file that only contains 15KB of valid information is not efficient).

In this chapter we’ll examine some of the key differences and problems with transferring data that transcend the mechanics of the process, and how to deal with them effectively. Although the three primary aspects, basic formatting, structural comparisons and data type limitations are handled separately here, there are few occasions when you can truly treat these elements individually. We’ll see some examples of this as we go through.

Basic Interchange Formats

When you start to move data efficiently between the different database types that you exist you will find a number of different possible interchange formats, and the primary issue with all of them is exactly how efficiently, and more importantly accurately, they enable the information to be exchanged.

First, the data will need to have been encoded with some kind of encapsulation format. This format is what describes the individual structure, and is entirely dependent on the underlying data that is being exchanged. For example, if the data has a very rigid structure then that will obviously normally enforce the format of the information. Row-based data, for example, can normally be encoded using CSV or even a fixed-width record format.

The second aspect is the encoding and formatting of the information itself. Exchanging data using CSV is fine, providing that you can correctly identify the format of the text itself. Should it be encoded in UTF-8? Or UTF-32? Is plain ASCII better? What if it contains a mix of these characters, should UTF-8 be used as the standard and the actual encoding handled by the database target when the data is imported?

In fact, many of the principles about translating information between different databases also rely on basic best practice for how you design and structure the data in the target system to begin with. Normalisation of the data structure for information can normally be abpplied to any database, even those that might have a loose or undefined structure have conventions. It’s unlikely, for example, that you will call a recipe title field ‘title’ in one record and ‘recipename’ in another record of the same database because your application will be a mess.

Of course, there are times when you may be merging, combining or otherwise consolidating data from a wide variety of different documents, records or blocks of information. There it is up to you to ultimately pick a standardisation for it to be useful to you once it’s been moved into the target system.

As a rough guide for the types of operation and translation that might take place, the following table highlights the kind of structural transformation and changes you might need to make when moving between some of the most common database environments.

Table 2-1: Structural Mappings between database environments

RDBMS Columnar Store Document Database Freetext/unstructured data store
RDBMS Vendor specific only Vendor specific only Field mappings only Application specific
Columnar Store Vendor specific only Vendor specific only Field mappings only Application specific
Document Database Field mappings only Field mappings only Vendor specific only Application specific
Freetext/unstructured data store Application specific Application specific Application specific Application specific


  • Vendor specific only changes are those that are directly related to the capabilities of the source or target database. MySQL for example supports the ENUM and SET compound field types, whereas Oracle, PostgreSQL and SQL Server do not. Moving from one to the other may required changes.
  • Field mappings only refers to how you map the source fields or columns to the target fields/columns. Depending on the target this may include compound and/or JOIN based translation. For example, when moving from a document database to an RDBMS you might convert a compound field into a single field, or a lookup table. When translating from an RDBMS to a document store, the data might be combined using a JOIN into a single target field.
  • Application specific changes are those that will entirely depend on how you to use the information. Translating document data into freetext databases is unlikely to require any changes. But converting freetext info into an RDBMS format is going to require some significant identification and translation.

Let’s dig into some more of the specific challenges.

Row-Based Data

For row-based data, the information can generally represented and formatted as one of the regularly used and displayed formats, such as Character Separated Values (i.e. CSV), or in a fixed width format. Row-based data (which includes the column-based data used in big data stores) is probably one of the easiest formats of data to exchange. In nearly all cases the list of columns is usually pretty well fixed and the format of the data is well known because the columns are fixed.

Character Separated Values (not Comma-separated values) is one of the oldest methods of exchanging fixed format data like this, it was often used as the only available method for exchanging information in a reliable fashion. Historically most tabulated data like this tended to be financial information, and so the content and format of the information was relatively simple. As such, the most common format was to use carriage-returns (or the operating system equivalent, which could be newlines or carriage-return and newline characters) to separate the records, while the individual fields in each row of data were separated by a comma (hence Comma-Separated Values as the CSV).

For example:

1085,Creamy egg and leek special,,4,1,0,0,0.0,0.0,0
87,Chakchouka,A traditional Arabian and North African dish and often accompanied with slices of cooked meat                      ,4,1,0,0,0.0,0.0,0
347,Mozzarella and olive pizza,A very simple pizza base made without yeast topped with traditional Italian ingredients. Look out for a low fat type of Mozzarella or Cheddar cheese if a low fat diet is being followed.,4,1,0,0,0.0,0.0,0
720,Savoury pancakes,Basic pancake recipe. Allow 30 min standing time for batter.,8,1,0,0,0.0,0.0,0
477,Seafood filling for pancakes,,8,1,0,0,0.0,0.0,0

The problem with commas and carriage-return characters is that, as computers got more complex, and the data they stored got equally more complex, how do you determine between a comma in some text, and a comma separating a field? What if you transfer a text string that contains a newline or carriage return. You don’t want that interpreted as the end of the record if it happens to part of the field. The initial solution is to use some kind of further delimiter. For example, using double-quotes:

"1085","Creamy egg and leek special","","4","1","0","0","0.0","0.0","0"
"87","Chakchouka","A traditional Arabian and North African dish and often accompanied with slices of cooked meat                      ","4","1","0","0","0.0","0.0","0"
"347","Mozzarella and olive pizza","A very simple pizza base made without yeast topped with traditional Italian ingredients. Look out for a low fat type of Mozzarella or Cheddar cheese if a low fat diet is being followed.","4","1","0","0","0.0","0.0","0"
"720","Savoury pancakes","Basic pancake recipe. Allow 30 min standing time for batter.","8","1","0","0","0.0","0.0","0"
"477","Seafood filling for pancakes","","8","1","0","0","0.0","0.0","0"

This doesn’t fix the problem, it just diverts your attention for long enough to realize that now what happens if one of the delimiting characters needs to be used in the text? We could escape it, by prefixing it with a backslash:

"700","Garlic mushroom kebabs","The longer you leave these mushrooms to marinate, the better they will taste.\nGood for barbecue.","8","1","0","0","0.0","0.0","0"

But now we’re getting complex, both to read and write the information, the level of complexity is increasing to point of introducing further possible methods of corrupting the data as it gets transferred.

The alternative is to use a different delimiter that is unlikely to be used within the text in any form. Hadoop in fact follows this model, using the hex characters, 0x01 and 0x0A to delimit records and fields. As binary characters these are unlikely to be used in what is normally human-readable text. Of course, once you start transferring binary data, you need to find another method, such as hex-encoding binary data.

"700","Garlic mushroom kebabs",VGhlIGxvbmdlciB5b3UgbGVhdmUgdGhlc2UgbXVzaHJvb21zIHRvIG1hcmluYXRlLCB0aGUgYmV0dGVyIHRoZXkgd2lsbCB0YXN0ZS5cbkdvb2QgZm9yIGJhcmJlY3VlLgo=,"8","1","0","0","0.0","0.0","0"

The other alternative is to use a fixed width format. This has the advantage that providing you know the exact widths of the individual fields, encoding errors are eliminated because the characters are no longer significant in the format of the information.

The downside of the fixed-width format is that the size of the fields, records, and ultimately files, can become prohibitively large if you are exchanging potentially large or unlimited fields. For example, BLOB types in most databases can be MB or GB in size; expressing that in fixed width format is obviously not a practical solution.

Record-Based Data

Record based is information that may not necessarily be identifiable or resolvable by an easy to use row specific format or structure such as that used in CSV exchange. Complex table data, or information that that is made up of a combination of fixed fields and BLOB fields, for example, is unlikely to reliably, or efficiently, transferred. The problems of character and structural formats will ultimately make using that information difficult or computationally expensive when actively sharing the information – for example by making the file sizes too big to be practically exchanged.

A very typical example of record based information is either information from a document-based database, a free texrt database, or where the information that makes up the majority of the content is in fact really an attachment or noit inline field and database data. Think of an email message; the message, the address, from, subject are all examples of easily identifiable and classifiable database information. But what do you do with an attachment that might have been sent along with the recored?

How about documents generally? Metadata about those documents could be represented in a typical RDBMS row, but not the document itself. But the combination of the two – the metadata and the actual document together make up a ‘record’ that you may want to effectively share all or part of with another database.

When transferring record-based data, think first about what constitutes the record and how that can be represented in the different environments. Then move on to understand how the fields and individual data points can be translate into a format suitable for the target database. With record-based data, it may be that you have a massive volume of data and compound records that when move from a document store to a transactional RDBMS require 20, 30 or 200 rows of data to be represented properly; this is not a problem, providing you find a method for identifying all of the row data that refers to the record is handled correctly.

In general with a record based database the easiest approach is to actually translate the data at the source into something that can be imported directly into the target format. For example, from a record-based environment you can generate four different blocks of CSV import data, one for each table and portion of the source information.

The opposite is actually also true; when converting data from a column or table store into a record based format, it normally makes sense to do this on the basis of translating the key data into the new structure before doing the transfer. As a general rule, either use the native target format if you can, or make use of representable formats such as JSON to do the actual encapsulation of the information. Many record or document-based environments already use JSON, or a format similar to this.

    "title": "Fried chilli potatoes",
    "preptime": "5"
    "servings": "4",
    "totaltime": "10",
    "subtitle": "A new way with chips.",
    "cooktime": "5",
    "ingredients": [
            "ingredtext": "chilli powder",
            "ingredient": "chilli powder",
            "meastext": "3-6 tsp"
            "ingredtext": "potatoes, peeled and cut into wedges",
            "ingredient": "potatoes",
            "meastext": "900 g"
            "ingredtext": "vegetable oil for deep frying",
            "ingredient": "vegetable oil for deep frying",
            "meastext": ""

One final consideration is those situations where there is no structure – or the structure is so ephemeral or complex that there is no way to map information. You cannot, indeed should not, either impose a structure just for the sake of one, or inversely rely on sub-standard features in the target database just because it makes your life easier.

Some examples of this include trying to extract key fields or information from unstructured data that are complicated either to identify, or to map back to the original. Extracting a telephone number from a massive text string just because you can does not mean that the telephone number you have extracted is really the one that should be associated with this record in your database. Equally, relying on full-text searching engines within RDBMS environments can be problematic.

Is that a Column or a Field?

Not all fields and columns are created equal, and many of these difficulties come from the complexities or abilities of the database environment being used. Specifically, at which point do you treat a fragment of the data that you are dealing with as a column, or a field, or even just a uniquely identifiable piece of information?

As you move between different environments, the differences become more pronounced or more difficult to identify. True columnar stores, for example, tend to have a significantly reduced number of datatypes and support, and that often limits your ability to store certain values and information. For example, compound values, or specialist types, such as XML, GeoData and spatial points may be storable in one database but not another.

Consider this record, taken from a CouchDB (JSON document) database:

   "title" : "Chicken Curry",
   "Keywords" : [
   "id" : "8023754"

Now is the ‘Keywords’ compound object in the document a field, or is it a column? In MySQL we could translate this into a SET datatype, a special datatype, providing we knew what all the possible values for that column are. In Oracle, a field that has multiple possible values like this would normally either be split into separate columns as a bit or boolean value, or it would converted to a lookup table, as in the diagram below.


Depending on your use case, within a strict columnar environment such as Cassandra or HP Vertica you might actually consider going the other way and repeating the information with the keyword data in individual rows, like this:

dbadmin=> select * from recipes_kwbytext;
   id    |        title         |     kw
 8023754 | Chicken Curry        | Indian
 8023754 | Chicken Curry        | Chicken
 8023754 | Chicken Curry        | Okra
 8023754 | Chicken Curry        | Spicy
(4 rows)

With a column store this can be more efficient if what you are looking for is patterns in the data, because repeated data like this is easy to extract and identify. In this case, what we’ve done is convert something that is a compound field in a document store into multiple rows with the same ID in a column store. This solution can also be used in environments where there are no JOIN operations, or a JOIN is expensive, but where the information is still required at each level. Good examples here are many of the document stores and structureless environments such as Hadoop.

Can you Bypass Datatypes?

Another temptation when translating data between very different database formats is simply to ignore the formatting, and especially the very strict datatypes, that might normally define the data being stored.

This is particularly a problem within those environments where there may be a very limited set of datatypes to work with and can be application specific. For example, the Hive database environment within Hadoop is reqally little more than a thin veneer over a text format used to store the actual data. When you define a table within Hive and then select the rows from the table, Hive parses each row and uses that to display the value in the corresponding format.

This can cause problems for certain data, for example, numbers that are too big, dates that don’t match the very limited set of date formats supported by the Hive parser. In the long term, this causes corruption of the data that you have transferred.

For this reason, some people choose to create tables within Hive that use the Text datatype to display the information rather than the true underlying Integer or Floating Point value as it ensuires the raw value, not the interpreted value will be used.

The same process can be used when moving data; extract the ‘raw’ value rather than hope the source or target database will interpret, store and display the information in the right format.

If you are only sharing or displaying the information in the new target database then there is probably no reason to worry. If you start processing or actively using the data, this is where corruption can occur if you are not storing the information correctly. For example, if an integer is stored and then incremented, you want 10,000 to become 10,001, not 100001.

The bottom line, you can bypass the datatype, but probably shouldn’t if you hope to use the information in the format in which you’ve decided to store it. If you have a datatype, and can identify it from the source material, then use it if the target environment can handle and interpret it correctly. See the notes later in this chapter on limitations in different environments.

Optimization and Performance

Irrespective of the the database environment and the reason for you moving the data, the end goal should always be to move data into a target in a format that will be efficient to use at the target end.

In many document or unstructured systems, or those with very flexible storage mechanism such as Hadoop, the performance will often be predicated not on the structure of the information, but what information is closest to you, or what can be pre-mapped or organized through a map reduce or index generation exercise.

Conversely, RDBMS require highly structured and organized data structures both with and without indexing to provide the best performance. Columnar stores are often much more efficient if you can logically group or sort information together. Some will handle this automatically for you, otherwise are more efficient if you can pre-determine the distribution of the data on which you are most likely to sort and query on. That might mean that when you transfer the data, you sort the generated file by that column or columns before loading. In some extreme examples it may be that you load the data in an unordered format and then move again into a new table with the right column structure.

Don’t be afraid of making the wrong decision, because you can often sort this structure out during a secondary or tertiary stage, but equally don’t ignore it. Having to parse or process large bodies of data a second or third time will be impractical if you are sharing or replicating data compared to single, isolated, dumps.

Ensure Two-way Validity

Without good reason, you should always try and avoid making any kind of translation of format that cannot be either reversed, undone, or translated back into it’s original format, even if that might make the process a little more complicated. Remember that data is often a living organism when being actively used and employed. Therefore doing too much to format, combine, extract or otherwise manipulate the information can then make it difficult to be used again elsewhere.

Note that this is not about normalization. Normalization in typical database parlance means finding the right, fixed, database type for the field data, making it the most efficient choice, and understanding the limits and structure of the data so that you can decide whether a field should be 10 bytes or 12 bytes long. Doing this normally results in identifying the data structure, lookup tables, relations and joins so that you have the right structure. In this context, normalization is really about making the data look like a typical table structure in an RDBMS; normalization for document databases is entirely different. Normalisation for data interoperability is another level still, and we’ve already seen a number of different examples of that.

Instead, think about the quality of the data and how it should be used, while keeping in mind that the structure required for efficiency in an RDBMS may be completely different to the efficient storage of the same information in a document DB, or when resolved down to a text fragment in an unstructured data store.

To return to a well-trodden example, in the section ‘Is that a Column or a Field?’ we looked at the translation of compound values into single or multiple fields. Using a comma to separate the potential values means that we could split the value back out. If the field had been correctly translated either to boolean columns or a linked table is easier to translate back again into a whole variety of formats.

When representing a compound type, think about how you would reverse the structure so that it could be used the other way round. For example, if you decide to dump the information out to another table or column structure, make sure that you add identifiers so that you can track the record it came from (which you’ll probably need anyway), and can reformat it back into that format. If you’ve split it out into multiple rows in a columnar store, make sure you know how to combine it back and deduplicate the information again if you need to send it the other way.

The compound types are the most complex single field type here because there are so many opportunities for you to mess up the translation, but the same is also true for basic structural information, and even more so if you decide that you only want to transfer a smaller number of fields of data from one database to another. Either transfer everything, or transfer what you need and include information (like a unique identifier) so that you can associate it back with the data you extracted it from. Once you’ve lost the context of the information, it can be impossible to get it back and the result is a useless dataset.

Database Metadata

We’ve concentrated very heavily on the actualy data you are storing and want to work with, but what about metadata:

  • Sequences, current auto-increment values?
  • Whether you transfer or record these is going to depend on exactly how you want to use the information when it reaches your destination (or comes back).
  • How about the definition of the structure that you are using? Do you want to be able to share and use that? What happens when the structure changes. Do you want to track and identify those changes?
  • When doing a one-time export of information from one database to another you can be sure about the structure and what you expect to get from it. But what happens when you repeat the same export multiple times? Or when replicating?
  • If you are sharing data between two different systems and integrating them, knowing the sequence number may be irrelevant unless you can synchronize the generation of the number so that it can be used by the multiple databases in an effective manner. Perhaps your databases could use a better unique identification method, rather than relying on a monotonically increasing sequence number, such as UUIDs or using a central identifier registry?

How to address these different problems will be covered in later chapters, but it’s important to think about it here as it has a knock on effect to other areas. For example, when moving unstructured or document based databases into multiple separate tables, you need to identify and tie that information together, where a UUID is important, and it therefore becomes a critical part of the data structure that you swap.