Loading Graphs

GraphScope models graph data as Property Graph, in which the edges/vertices are labeled and each label may have many properties.

Load Built-in Datasets

GraphScope comes with a set of popular datasets, and utility functions to load them into memory, makes it easy for user to get started. Here’s an example:

from graphscope.dataset import load_ogbn_mag
graph = load_ogbn_mag()

In standalone mode, it will automatically download the data to ~/.graphscope/dataset, and it will remain in there for future usage.

However, in kubernetes mode, it’s not trivial to download the data to the pod’s local storage, so we provide an option to mount a OSS bucket, which contains all available datasets. For example, we load the same graph as above, but this time graphscope is running in a Kubernetes cluster:

import graphscope
from graphscope.dataset import load_ogbn_mag
sess = graphscope.session(cluster_type='k8s', mount_dataset='/dataset')
graph = load_ogbn_mag(sess, '/dataset/ogbn_mag_small')

Here, we first created a Session in a Kubernetes cluster, and mount the dataset bucket to /dataset, this path is relative to Pods. Then we pass that session as first parameter, and /dataset/ogbn_mag_small as second parameter. The /dataset is the root path of datasets which we have assigned by mount_dataset, the ogbn_mag_small is the sub folder name of the dataset.

You can view all available datasets in here , and get details description and usage in those source files.

Loading Your Own Datasets

However, it’s more common that user need to load there own data and do some analysis. To load a property graph to GraphScope, we provide a method g() defined in Session.

To build a property graph on GraphScope, we firstly create an empty graph using g().

import graphscope
sess = graphscope.session()
# Use `sess = graphscope.session(cluster_type='hosts')` if you are in standalone mode.
graph = sess.g()

The class Graph has several methods:

def add_vertices(self, vertices, label="_", properties=None, vid_field=0):
    pass

def add_edges(self, edges, label="_e", properties=None, src_label=None, dst_label=None, src_field=0, dst_field=1):
    pass

These methods helps users to construct the schema of the property graph iteratively.

We will use files in ldbc_sample through this tutorial. You can get the files in here. And you can inspect the graph schema by using print(graph.schema).

Adding Vertices

We can add a kind of vertices to graph, the method has the following parameters:

vertices: A location for the vertex data source, which can be a file location, or a numpy, etc. See more details in Loader object.

A simple example:

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv')

It will read data from the location /home/ldbc_sample/person_0_0.csv. Since we didn’t give additional arguments, these vertices will be labeled _ by default, using the first column in the file as their ID, and other columns as their properties. Both the names and data types of properties will be deduced.

Another commonly used parameter is label:

label: The label name of the vertex, default to _.

Since a property graph allows many kinds of vertices, it is suggested for users to give each kind of vertices a meaningful label name. For example:

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')

Then we have a graph with one kind of vertices, its label name is person.

In addition, each kind of labeled vertices have their own properties. Here is the third parameter:

properties: A list of properties, Optional, default to None.

This parameter selects the corresponding columns from the source data file or pandas DataFrames as properties. Please note that the values of this parameter should exist in the file/DataFrame. By default( values None), all columns except the vid_field column will be added as properties. If it equals to a empty list [], then no properties will be added.

For example:

# properties will be firstName,lastName,gender,birthday,creationDate,locationIP,browserUsed
graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person', properties=None)

# properties will be firstName, lastName
graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person', properties=['firstName', 'lastName'])

# no properties
graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person', properties=[])

vid_field determines which column used as vertex ID. (as well as the source ID or destination ID when loading edges.)

It can be a str, the name of columns, or int, representing the index of the columns.

By default, the value is 0, hence the first column will be used as vertex ID.

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', vid_field='firstName')

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', vid_field=0)

Adding Edges

Next, let’s take a look on the parameters for loading edges.

edges: The location indicating where to read the data. e.g.,

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')
# Note we already added a vertex label named 'person'.
graph = graph.add_edges('/home/ldbc_sample/person_knows_person_0_0.csv', src_label='person', dst_label='person')

This will load an edge which label is _e (the default value), its source vertex and destination vertex will be person, using the first column as the source vertex ID, the second column as the destination vertex ID, the others as properties.

Similar to vertices, we can use parameter label to assign label name and properties to select properties.

label: The label name of the edges, default to _e . (It’s recommended to use a meaningful label name.) properties: A list of properties, default to None (add all columns as properties).

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')
graph = graph.add_edges('/home/ldbc_sample/person_knows_person_0_0.csv', label='knows', src_label='person', dst_label='person')

Differ to vertices, edges have some additional parameters.

src_label: The label name of the source vertex. dst_label: The label name of the destination vertex, it can be different to the src_label, src_field and dst_field: The columns used for source(destination) vertex id. Default to 0 and 1, respectively.

e.g.,

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')
graph = graph.add_vertices('/home/ldbc_sample/comment_0_0.csv', label='comment')
# Note we already added a vertex label named 'person'.
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv', label='likes', src_label='person', dst_label='comment')

The value and behavior is similar to vid_field in Vertex, except for it takes two columns as edges is constituted by source vertex id and destination vertex id. Here’s an example:

Examples of src_field and dst_field:

# Steps to init a graph and add vertices are omitted
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv', label='likes', src_label='person', dst_label='comment', src_field='Person.id', dst_field='Comment.id')
# Or use the index.
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv', label='likes', src_label='person', dst_label='comment', src_field=0, dst_field=1)

Advanced Usages

Here are some advanced usages to deal with homogeneous graphs or very complex graphs.

Deduce vertex labels when not ambiguous

If there is only one kind of vertices in a graph, the vertex label can be omitted. GraphScope will infer the source and destination vertex label to that very label.

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')
# GraphScope will assign `src_label` and `dst_label` to `person` automatically.
graph = graph.add_edges('/home/ldbc_sample/person_knows_person_0_0.csv')

Deduce vertex from edges

If user add edges with unseen src_label or dst_label, graphscope will extract an vertex table from the given labels from the edge data.

graph = sess.g()
# Deduce vertex label `person` from the source and destination endpoints of edges.
graph = graph.add_edges('/home/ldbc_sample/person_knows_person_0_0.csv', src_label='person', dst_label='person')

graph = sess.g()
# Deduce the vertex label `person` from the source endpoint,
# and vertex label `comment` from the destination endpoint of edges.
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv', label='likes', src_label='person', dst_label='comment')

Multiple relations

In some cases, an edge label may connect two kinds of vertices. For example, in a graph, two kinds of edges are labeled with likes but represents two relations. i.e., person -> likes <- comment and person -> likes <- post.

In this case, we can simply add the relation again with the same edge label, but with different source and destination labels.

# Steps to init a graph and add vertices are omitted
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv',
        label="likes",
        src_label="person", dst_label="comment",
    )

graph = graph.add_edges('/home/ldbc_sample/person_likes_post_0_0.csv',
        label="likes",
        src_label="person", dst_label="post",
    )

Specify data types of properties manually

GraphScope will deduce data types from input files, and it works as expected in most cases. However, sometimes user may want to determine the data types as well, e.g.

graph = sess.g()
graph = graph.add_vertices('/home/ldbc_sample/post_0_0.csv', label='post', properties=['content', ('length', 'int'), ])

It forces the property to be (casted and) loaded as specified data type. The format of this parameter is tuple(s) with the name and the type. e.g., in this case, the property length will have type int rather than the default int64_t. The options of the types are int, int64, float, double, or str.

Other parameters of graph

The class Graph has three meta options, which are:

  • oid_type, can be int64_t or string. Default to int64_t in consideration of efficiency. But if the ID column can’t be represented by int64_t, then we should use string.

  • directed, boolean value and default to True. Controls to load an directed or undirected graph.

  • generate_eid, bool, default to True, whether to generate an unique id for all edges automatically.

Putting them Together

Let’s make this example complete.

graph = sess.g(oid_type='int64_t', directed=True, generate_eid=True)
graph = graph.add_vertices('/home/ldbc_sample/person_0_0.csv', label='person')
graph = graph.add_vertices('/home/ldbc_sample/comment_0_0.csv', label='comment')
graph = graph.add_vertices('/home/ldbc_sample/post_0_0.csv', label='post')

graph = graph.add_edges('/home/ldbc_sample/person_knows_person_0_0.csv', label='knows', src_label='person', dst_label='person')
graph = graph.add_edges('/home/ldbc_sample/person_likes_comment_0_0.csv', label='likes', src_label='person', dst_label='comment')
graph = graph.add_edges('/home/ldbc_sample/person_likes_post_0_0.csv', label='likes', src_label='person', dst_label='post')

print(graph.schema)

A more complex example to load LDBC snb graph can be find here.

Loading From Pandas or Numpy

The data source aforementioned is an object of Loader. A loader wraps a location or the data itself. GraphScope supports load a graph from pandas dataframes or numpy ndarrays, making it easy to construct a graph right in the python console.

Apart from the loader, the other fields like properties, label, etc. are the same as examples above.

From Pandas

import pandas as pd

df_v = pd.read_csv('/home/ldbc_sample/comment_0_0.csv', sep='|')
df_e = pd.read_csv('/home/ldbc_sample/comment_replyOf_comment_0_0.csv', sep='|')

# use a dataframe as datasource, properties omitted,
# for edges, col_0/col_1 will be used as src/dst by default.
# for vertices, col_0 will be used as vertex_id by default.
graph = sess.g().add_vertices(df_v).add_edges(df_e)

From Numpy

Note that each array is a column, we pass it like as COO matrix format to the loader.

import numpy

array_v = [df_v[col].values for col in df_v.columns.values]
array_e = [df_e[col].values for col in df_e.columns.values]

graph = sess.g().add_vertices(array_v).add_edges(array_e)

Loader Variants

When a loader wraps a location, it may only contains a str. The string follows the standard of URI. When receiving a request for loading graph from a location, graphscope will parse the URI and invoke corresponding loader according to the schema.

Currently, graphscope supports loaders for local, s3, oss, hdfs. Under the hood, data is loaded distributedly by v6d, v6d takes advantage of fsspec to resolve specific scheme and formats. Any additional configurations can be passed in kwargs of Loader, which will be parsed directly by the specific class. e.g., host and port to hdfs, or access-id, secret-access-key to oss or s3.

from graphscope.framework.loader import Loader

ds1 = Loader("file:///var/datafiles/group.e")
ds2 = Loader("oss://graphscope_bucket/datafiles/group.e", key='access-id', secret='secret-access-key', endpoint='oss-cn-hangzhou.aliyuncs.com')
ds3 = Loader("hdfs:///datafiles/group.e", host='localhost', port='9000', extra_conf={'conf1': 'value1'})
d34 = Loader("s3://datafiles/group.e", key='access-id', secret='secret-access-key', client_kwargs={'region_name': 'us-east-1'})

User can implement customized driver to support additional data sources. Take ossfs as an example, User need to subclass AbstractFileSystem, which is used as resolve to specific protocol scheme, and AbstractBufferFile to do read and write. The only methods user need to override is _upload_chunk, _initiate_upload and _fetch_range. In the end user need to use fsspec.register_implementation('protocol_name', 'protocol_file_system') to register corresponding resolver.

Techniques targeting at large graph

Tips for reduce memory consumption of graphs

  • Tune the parameter of graph constructor sess.g()
    • Set directed=False to use an undirected graph if you do not require edge directionality. Undirected graphs require less memory than directed graphs for the same data size, as we do not need to store edge directions.

    • Set generate_eid=False if you do not require edge ids for interactive engine (GIE) operations.

    • Set retain_oid=False if you do not require the ID column as a property for interactive engine (GIE) operations.

    • Set oid_type=’int32_t’ when the ID does not exceed 2^31 - 1.

    • Providing a complete schema that specifies the data type of each property instead of allowing GraphScope to infer it from data could benefit most cases.

  • Filter out super vertices according to the requirements of the business scenario. For certain business scenarios or algorithms, high precision may not be necessary, especially when dealing with very large graph data.