This discussion is regarding support for Map Data type in Carbon Data.
Carbon Data supports complex and nested data types such as Arrays and
Struts. However, Carbon Data does not support other complex data types such
as Maps and Union which are generally supported by popular opensource file
Supporting Map data type will require changes/additions to the DDL, Query
Syntax, Data Loading and Storage.
I have hosted the design on google docs for review and discussion.
Below is the same inline.
1. DDL Changes
Maps are key->value data types and where the value can be fetched by
providing the key. Hence we need to restrict keys to primitive data types
whereas values can be of any data type supported in Carbon(primitive and
Map data types can be defined in the create table DDL as :-
create table example_table (id Int, name String, salary Int, salary_breakup
map<String, Int>, city String)
2. Data Loading Changes
Carbon should be able to support loading data into tables with Map type
columns from csv files. It should be possible to represent maps in a single
row of csv. This will need carbon to support specifying the delimiters for
1. Between two Key-Value pairs
2. Between each Key and Value in a pair
As Carbon already supports Strut and Array Complex types, the data loading
process already provides support for defining delimiters for complex data
types. Carbon provides two Optional parameters for data loading
1. COMPLEX_DELIMITER_LEVEL_1: will define the delimiter between two
2. COMPLEX_DELIMITER_LEVEL_2: will define the delimiter between each
Key and Value in a pair
With these delimiter options, the below map can be represented in csv:-
Fixed:100,000$Bonus:30,000$Stock:40,000 in the csv file.
3. Query Capabilities
A complex datatype like Map will require additional operators to be
supported in the query language to fully utilize the strength of the data
Maps are sequence of key-value pairs, hence should support looking up value
for a given key. Users could use the ColumnName[“key”] syntax to lookup
values in a map column. For example: salary_breakup[“Fixed”] could be used
to fetch only the Fixed component in the salary breakup.
In Addition, we also need to define how maps can be used in existing
constructs such as select, where(filter), group by etc..
1. Select:- Map data type can be directly selected or only the value
for a given key can be selected as per the requirement. For example:-“Select
name, salary, salary_breakup” will return the content of map long with each
row.“Select name, salary, salary_breakup[“Fixed”]” will return only one
value from the map whose key is “Fixed”2. Filter:-Map data type cannot
be directly used in a where clause as where clause can operate only on
primitive data types. However the map lookup operator can be used in where
clauses. For example:-“Select name, salary where
salary_breakup[“Bonus”]>10,000”*Note: if the value is not of primitive
type, further assessor operators need to be used depending on the type of
value to arrive at a primitive type for the filter expression to be valid.*
3. Group By:- Just like with filters, maps cannot be directly used in a
group by clause, however the lookup operator can be used.
4. Functions:- A size() function can be provided for map types to
determine the number of key-value pairs in a map.
4. Storage changes
As Carbon is a columnar data store, Map values will be stored using 3
1. One Column for representing the Map Data type. Will store the number
of fields and start index, just the same way as it is done for Struts and
2. One Column for the Key
3. One Column for the value, if the value is of primitive data type,
else the value itself will be multiple physical columns depending on the
data type of the value.