GDS Sessions (For AuraDB)
This Jupyter notebook is hosted here in the Neo4j Graph Data Science Client Github repository.
The notebook shows how to use the graphdatascience
Python library to
create, manage, and use a GDS Session.
We consider a graph of people and fruits, which we’re using as a simple example to show how to connect your AuraDB instance to a GDS Session, run algorithms, and eventually write back your analytical results to the AuraDB database. We will cover all management operations: creation, listing, and deletion.
If you are using self managed DB, follow this example.
1. Prerequisites
This notebook requires having an AuraDB instance available and have the GDS sessions feature enabled for your tenant. Contact your account manager to get the features enabled.
We also need to have the graphdatascience
Python library installed,
version 1.12a1
or later.
%pip install "graphdatascience>=1.12a1"
2. Aura API credentials
A GDS Session is managed via the Aura API. In order to use the Aura API, we need to have Aura API credentialsn.
Using these credentials, we can create our GdsSessions
object, which
is the main entry point for managing GDS Sessions.
import os
from graphdatascience.session import GdsSessions, AuraAPICredentials
client_id = os.environ["AURA_API_CLIENT_ID"]
client_secret = os.environ["AURA_API_CLIENT_SECRET"]
# If your account is a member of several tenants, you must also specify the tenant ID to use
tenant_id = os.environ.get("AURA_API_TENANT_ID", None)
sessions = GdsSessions(api_credentials=AuraAPICredentials(client_id, client_secret, tenant_id=tenant_id))
3. Creating a new session
A new session is created by calling sessions.get_or_create()
. As the
data source, we assume here that an AuraDB instance has been created and
is available for access. We need to pass the database address, username
and password to the DbmsConnectionInfo
class.
We also need to specify the size of the session. Please refer to the API reference documentation or the manual for a full list.
Finally, we need to give our session a name. We will call ours
people-and-fruits'. It is possible to reconnect to an existing session by calling`get_or_create
with the same session name and configuration.
We will also set a time-to-live (TTL) for the session. This ensures that our session is automatically deleted after being unused for 5 hours. This is a good practice to avoid incurring costs should we forget to delete the session ourselves.
import os
from datetime import timedelta
from graphdatascience.session import DbmsConnectionInfo, AlgorithmCategory
# Identify the AuraDB instance
db_connection = DbmsConnectionInfo(uri=os.environ["AURA_DB_ADDRESS"], username=os.environ["AURA_DB_USER"], password=os.environ["AURA_DB_PW"])
# Create a GDS session!
memory = sessions.estimate(
node_count=20,
relationship_count=50,
algorithm_categories=[AlgorithmCategory.CENTRALITY, AlgorithmCategory.NODE_EMBEDDING],
)
gds = sessions.get_or_create(
# we give it a representative name
session_name="people-and-fruits",
memory=memory,
db_connection=db_connection,
ttl=timedelta(hours=5),
)
4. Listing sessions
Now that we have created a session, let’s list all our sessions to see what that looks like
sessions.list()
5. Adding a dataset
We assume that the configured AuraDB instance is empty. We will add our dataset using standard Cypher.
In a more realistic scenario, this step is already done, and we would just connect to the existing database.
data_query = """
CREATE
(dan:Person {name: 'Dan', age: 18, experience: 63, hipster: 0}),
(annie:Person {name: 'Annie', age: 12, experience: 5, hipster: 0}),
(matt:Person {name: 'Matt', age: 22, experience: 42, hipster: 0}),
(jeff:Person {name: 'Jeff', age: 51, experience: 12, hipster: 0}),
(brie:Person {name: 'Brie', age: 31, experience: 6, hipster: 0}),
(elsa:Person {name: 'Elsa', age: 65, experience: 23, hipster: 1}),
(john:Person {name: 'John', age: 4, experience: 100, hipster: 0}),
(apple:Fruit {name: 'Apple', tropical: 0, sourness: 0.3, sweetness: 0.6}),
(banana:Fruit {name: 'Banana', tropical: 1, sourness: 0.1, sweetness: 0.9}),
(mango:Fruit {name: 'Mango', tropical: 1, sourness: 0.3, sweetness: 1.0}),
(plum:Fruit {name: 'Plum', tropical: 0, sourness: 0.5, sweetness: 0.8})
CREATE
(dan)-[:LIKES]->(apple),
(annie)-[:LIKES]->(banana),
(matt)-[:LIKES]->(mango),
(jeff)-[:LIKES]->(mango),
(brie)-[:LIKES]->(banana),
(elsa)-[:LIKES]->(plum),
(john)-[:LIKES]->(plum),
(dan)-[:KNOWS]->(annie),
(dan)-[:KNOWS]->(matt),
(annie)-[:KNOWS]->(matt),
(annie)-[:KNOWS]->(jeff),
(annie)-[:KNOWS]->(brie),
(matt)-[:KNOWS]->(brie),
(brie)-[:KNOWS]->(elsa),
(brie)-[:KNOWS]->(jeff),
(john)-[:KNOWS]->(jeff);
"""
# making sure the database is actually empty
assert gds.run_cypher("MATCH (n) RETURN count(n)").squeeze() == 0, "Database is not empty!"
# let's now write our graph!
gds.run_cypher(data_query)
gds.run_cypher("MATCH (n) RETURN count(n) AS nodeCount")
6. Projecting Graphs
Now that we have imported a graph to our database, we can project it
into our GDS Session. We do that by using the gds.graph.project()
endpoint.
The remote projection query that we are using selects all Person
nodes and their LIKES
relationships, and all Fruit
nodes and
their LIKES
relationships. Additionally, we project node properties
for illustrative purposes. We can use these node properties as input to
algorithms, although we do not do that in this notebook.
G, result = gds.graph.project(
"people-and-fruits",
"""
CALL {
MATCH (p1:Person)
OPTIONAL MATCH (p1)-[r:KNOWS]->(p2:Person)
RETURN
p1 AS source, r AS rel, p2 AS target,
p1 {.age, .experience, .hipster } AS sourceNodeProperties,
p2 {.age, .experience, .hipster } AS targetNodeProperties
UNION
MATCH (f:Fruit)
OPTIONAL MATCH (f)<-[r:LIKES]-(p:Person)
RETURN
p AS source, r AS rel, f AS target,
p {.age, .experience, .hipster } AS sourceNodeProperties,
f { .tropical, .sourness, .sweetness } AS targetNodeProperties
}
RETURN gds.graph.project.remote(source, target, {
sourceNodeProperties: sourceNodeProperties,
targetNodeProperties: targetNodeProperties,
sourceNodeLabels: labels(source),
targetNodeLabels: labels(target),
relationshipType: type(rel)
})
""",
)
str(G)
7. Running Algorithms
We can now run algorithms on the projected graph. This is done using the standard GDS Python Client API. There are many other tutorials covering some interesting things we can do at this step, so we will keep it rather brief here.
We will simply run PageRank and FastRP on the graph.
print("Running PageRank ...")
pr_result = gds.pageRank.mutate(G, mutateProperty="pagerank")
print(f"Compute millis: {pr_result['computeMillis']}")
print(f"Node properties written: {pr_result['nodePropertiesWritten']}")
print(f"Centrality distribution: {pr_result['centralityDistribution']}")
print("Running FastRP ...")
frp_result = gds.fastRP.mutate(
G,
mutateProperty="fastRP",
embeddingDimension=8,
featureProperties=["pagerank"],
propertyRatio=0.2,
nodeSelfInfluence=0.2,
)
print(f"Compute millis: {frp_result['computeMillis']}")
# stream back the results
gds.graph.nodeProperties.stream(G, ["pagerank", "fastRP"], separate_property_columns=True, db_node_properties=["name"])
8. Writing back to AuraDB
The GDS Session’s in-memory graph was projected from data in our specified AuraDB instance. Write back operations will thus persist the data back to the same AuraDB. Let’s write back the results of the PageRank and FastRP algorithms to the AuraDB instance.
# if this fails once with some error like "unable to retrieve routing table"
# then run it again. this is a transient error with a stale server cache.
gds.graph.nodeProperties.write(G, ["pagerank", "fastRP"])
Of course, we can just use .write
modes as well. Let’s run Louvain
in write mode to show:
gds.louvain.write(G, writeProperty="louvain")
We can now use the gds.run_cypher()
method to query the updated
graph. Note that the run_cypher()
method will run the query on the
AuraDB instance.
gds.run_cypher(
"""
MATCH (p:Person)
RETURN p.name, p.pagerank AS rank, p.louvain
ORDER BY rank DESC
"""
)
9. Deleting the session
Now that we have finished our analysis, we can delete the session. The results that we produced were written back to our AuraDB instance, and will not be lost. If we computed additional things that we did not write back, those will be lost.
Deleting the session will release all resources associated with it, and stop incurring costs.
gds.delete()
# or sessions.delete("people-and-fruits")
# let's also make sure the deleted session is truly gone:
sessions.list()
# Lastly, let's clean up the database
gds.run_cypher("MATCH (n:Person|Fruit) DETACH DELETE n")