# Integrating Other Data Sources

Data source plays an important role in **GL**, through which **GL** loads the raw data to build graph and sync states among distributed servers.
The data source is implemented by a kind of **file system**.
Currently, GL has implemented the file system interface to operate local files.
When running distributedly, a distributed file system is needed.
**GL** supports to be launched by KubeFlow and all servers access the file system by mounting **NFS** to local.

If it is difficult to configure your environment with KubeFlow and NFS, you `NEED` to implement a kind of file system to satisfy the requirements of GL.

## File System Interface

The [interface](../../../../graphlearn/src/platform/file_system.h) of file system defines the basic operation for file and directory, such as **Exist**, **Create** and **Delete**.
Besides, we abstract **three** kinds of files that a file system should implement.


### ByteStreamAccessFile

A **ByteStreamAccessFile** enables to read a file in byte stream.
That is, you can create such a file reader, seek to an `offset`, and then read the next `n bytes` until reach the end.

```c++
virtual Status Read(size_t n, LiteString* result, char* buffer) = 0;
```

The **GetFileSize()** interface in file system should return the size in bytes for a **ByteStreamAccessFile**.


### StructuredAccessFile

Different from **ByteStreamAccessFile**, **StructuredAccessFile** will consume the file in records instead of in bytes.
That is, you can create such a file reader, seek to an `offset`, and then read the `next Record` until reach the end.

```c++
virtual Status Read(io::Record* result) = 0;
```

A `Record` contains several columns described by **Schema**.
Each column value may be in type of `Int32`, `Int64`, `Float`, `Double` or `String`.

The raw graph data will be placed in **StructuredAccessFile**s, **GL** will load it and build graph index.
Refer [HERE](../graph/data_loader.md) for more supported raw data formats.

The **GetRecordCount()** interface in file system should return the size in records for a **StructuredAccessFile**.


### WritableFile

A **WritableFile** enables to append pieces of data into a file, and does not care about what the format is.

```c++
virtual Status Append(const LiteString& data) = 0;
```

## Register with Scheme

At last, register the new file system to the **Environment**.
Each file system should has a unique scheme, such as ```hdfs://```.

```
REGISTER_FILE_SYSTEM("scheme", MyFileSystem);
```

After registration, you can use such a file path when building [Graph](../graph/graph_object.md).