Permalink
Find file Copy path
52ad884 Oct 11, 2018
1 contributor

Users who have contributed to this file

203 lines (141 sloc) 7.43 KB

Row Based Access

Parquet, of course, is columnar format, and doesn't store data in rows. However, sometimes accessing data in a row-wise fashion is essential in processing algorithms and to display to a user. We as humans better understand rows rather than columns.

Parquet.Net provides out-of-the-box helpers to represent data in row format, however before using it consider the following:

  • Can you avoid using row based access? If yes, don't use row based access.
  • Row based helpers add a lot of overhead on top of parquet data as it needs to be transformed on the fly from columns to rows internally, and this cannot be done in performant way.
  • If your data access code is slow, this is probably because you are using row based access which is relatively slow.

Table

Table is at the root of row-based hierarchy. A table is simply a collection of Rows, and by itself implements IList<Row> interface. This means that you can perform any operations you normally do with IList<T> in .NET. A row is just a collection of untyped objects:

Rows General

Row

Row is a central structure to hold data during row-based access. Essentially a row is an array of untyped objects. The fact that the row holds untyped objects adds a performance penalty on working with rows and tables throught parquet, because all of the data cells needs to be boxed/unboxed for reading and writing. If you can work with column-based data please don't use row-based access at all. However, if you absolutely need row-based access, these helper classes are still better than writing your own helper classes.

Everything in parquet file can be represented as a set of Rows including plain flat data, arrays, maps, lists and structures.

Flat Data

Representing flat data is the most obvious case, you would simply create a row where each element is a value of a row. For instance let's say you need to store a list of cities with their ids looking similar to this:

id city
1 London
2 New York

The corresponding code to create a table with rows is:

var table = new Table(
   new Schema(
      new DataField<int>("id"),
      new DataField<string>("city")));

table.Add(new Row(1, "London"));
table.Add(new Row(2, "New York"));

Both ParquetReader and ParquetWriter has plenty of extension methods to read and write tables.

Arrays (Repeatable fields)

Parquet has an option to store an array of values in a single cell, which is sometimes called a repeatable field. With row-based access you can simply add an array to each cell. For instance let's say you need to create the following table:

ids
1,2,3
4,5,6

The corresponding code to populate this table is:

var table = new Table(
   new Schema(
      new DataField<IEnumerable<int>>("ids")));

table.Add(new Row(new[] { 1, 2, 3 }));
table.Add(new Row(new[] { 4, 5, 6 }));

Dictionaries (Maps)

var schema = new Schema(
   new DataField<string>("city"),
   new MapField("population",
      new DataField<int>("areaId"),
      new DataField<long>("count")));

and you need to write a row that has London as a city and population is a map of 234=>100, 235=>110.

The table should look like:

Column 0 Column 1
Row 0 London List<Row>

where the last cell is the data for your map. As we're in the row-based world, this needs to be represented as a list of rows as well:

Column 0 Column 1
Row 0 234 100
Row 1 235 110

To express this in code:

table.Add("London",
   new List<Row>
   {
      new Row(234, 100L),
      new Row(235, 110L)
   });

Structures

Structures are represented again as Row objects. When you read or write a structure it is embedded into another row's value as a row. To demonstrate, the following schema

var table = new Table(
   new Schema(
      new DataField<string>("isbn"),
      new StructField("author",
         new DataField<string>("firstName"),
         new DataField<string>("lastName"))));

represents a table with two columns - isbn and author, however author is a structure of two fields - firstName and lastName. To add the following data into the table

isbn author
12345-6 Ivan; Gavryliuk
12345-8 Richard; Conway

you would write:

table.Add(new Row("12345-6", new Row("Ivan", "Gavryliuk")));
table.Add(new Row("12345-7", new Row("Richard", "Conway")));

Lists

Lists are easy to get confused with repeatable fields, because they essentially repeat some data in a cell. This is true for a simple data type like a string, int etc., however lists are special in a way that a list item can be anything else, not just a plain data type. In general, when repeated data can be represented as a plain type, always use repeatable field. Repeatable fields are lighter and faster than lists which have extra overhead on serialisation and performance.

Simple Lists

In simple cases, when a list contains a single data element, it will be mapped to a collection of those elements, for instance in the following schema

var table = new Table(
   new Schema(
      new DataField<int>("id"),
      new ListField("cities",
         new DataField<string>("name"))));

and the following set of data:

id cities
1 London, Derby
2 Paris, New York

can be represented in code as:

table.Add(1, new[] { "London", "Derby" });
table.Add(2, new[] { "Paris", "New York" });

As you can see, it's no different to repeatable fields (in this case a repeatable string) however it will perform much slower due to transformation costs are higher.

Lists of Stuctures

A more complicated use case of lists where they actually make some sense is using structures (although lists can contain any subclass of Field). Let's say you have the the following schema definition:

var t = new Table(
   new DataField<int>("id"),
   new ListField("structs",
      new StructField("mystruct",
         new DataField<int>("id"),
         new DataField<string>("name"))));

and would like to add the following data:

id structs
1 id: 1, name: Joe; id: 2, name: Bloggs
1 id: 3, name Star; id: 4, name: Wars

which essentially creates a list of structures with two fields - id and name in a single table cell. To add the data to the table:

t.Add(1, new[] { new Row(1, "Joe"), new Row(2, "Bloggs") });
t.Add(2, new[] { new Row(3, "Star"), new Row(4, "Wars") });

ToString Overloads

.ToString() overloads on both Table and Row format data in single-line, single-quote JSON for your convenience. For instance, table.ToString() may produce following results:

{'id': 1, 'strings': ['1', '2', '3']}
{'id': 2, 'strings': []}
{'id': 3, 'strings': ['1', '2', '3']}
{'id': 4, 'strings': []}

which means that this table contains 4 rows, each row is a single-line JSON document. All the rows are separated by a line break character. This decision was made based on the fact that multiline JSON can be read directly by Apache Spark, and it's much more easier to parse a large document by splitting it by line separator character to get the next row.

Single quotes are chosen only based on the fact that in C# language it's hard to encode strings for tests with double quotes as you need to escape them. However, you can produce double quotes by using ToString overload and passing "j" as a format string, like .ToString("j").