- What is SQErzo
- Which databases are supported
- Why use SQErzo?
- Project status
- Install
- Usage examples
- ChangeLog
- TODO
- References
- License
SQErzo
is simple and tiny ORM (Object-Relational Mapping) for graph databases for Python developers.
It's compatible with databases that supports Open Cypher language.
Database | Status |
---|---|
Neo4j | Supported |
Redis Graph | Supported |
Arango DB | Looking for contributor |
AWS Neptune | Looking for contributor |
Gremlin | Looking for contributor |
SQErzo
intermediates between the graph database and your application logic in a database agnostic way. As such, SQErzo
abstracts the differences between the different databases. For examples:
- RedisGraph doesn't support Date times or CONSTRAINTS,
SQErzo
does the magic to hide that. - Neo4j need different channels for writing than for read.
SQErzo
does the magic to hide that. SQErzo
integrates a in memory cache to avoid queries to Graph DB and try to improve the performance.- Every database uses their own Node/Edge identification system. You need to manage and understand then to realize when a node already exits in Graph DB.
SQErzo
do this for you. It doesn't matter the Graph DB engine you use. SQErzo
was made to avoid you to write useless code. You can create and manage Nodes and Edges in a few lines of code without know Graph DB internals.SQErzo
supports Graph DB bases on Open cypher language (a Graph databases query language). You don't need to learn them to perform day a day operations.
Project is in a very early stage. If you want to use them, have in count that.
Install is easy. Only run:
> pip install sqerzo
> docker run -d -p7474:7474 -p7687:7687 -e NEO4J_AUTH=neo4j/s3cr3t neo4j
> docker run -p 7000:6379 -d --rm redislabs/redisgraph
Create some nodes and setup database in both databases:
- Neo4j
- RedisGraph
Without the need to change any code:
from dataclasses import dataclass
from sqerzo import GraphEdge, GraphNode, SQErzoGraph
class MeetEdge(GraphEdge):
pass
@dataclass
class UserNode(GraphNode):
name: str = None
def create_graph(connection_string: str):
gh = SQErzoGraph(connection_string)
gh.truncate() # Drop database
u1 = UserNode(name=f"UName-1")
gh.save(u1)
d1 = UserNode(name=f"DName-2")
gh.save(d1)
u1_meet_g1 = MeetEdge(
source=u1,
destination=d1
)
gh.save(u1_meet_g1)
if __name__ == '__main__':
create_graph("redis://127.0.0.1:7000/?graph=email")
create_graph("neo4j://neo4j:s3cr3t@127.0.0.1:7687/?graph=email")
This is the result database in Node4j:
This is the result database in RedisGrap:
SQErzo
try to be simple. So, if you want to do complex queries, you'll write them in the DB Engine language.
This example explains how to perform a query in Open Cypher language and map the results to Python Classes:
from dataclasses import dataclass
from sqerzo import GraphEdge, GraphNode, SQErzoGraph
class MeetEdge(GraphEdge):
pass
class WorksWithEdge(GraphEdge):
pass
@dataclass
class UserNode(GraphNode):
__keys__ = "email"
name: str = None
email: str = None
def create_graph(connection_string: str, nodes_count = 500):
gh = SQErzoGraph(connection_string)
gh.truncate() # Drop database
#
# Add some data and relations: User1 -[meet]-> User 2
#
with gh.transaction() as tx:
for n in range(nodes_count):
u1_name = f"uname{n}"
d1_name = f"dname{n}"
u1 = UserNode(name=u1_name, email=f"{u1_name}@{u1_name}.com")
d1 = UserNode(name=d1_name, email=f"{d1_name}@{d1_name}.com")
tx.add(u1)
tx.add(d1)
u2_meet_u1 = MeetEdge(
source=u1,
destination=d1
)
u1_meet_u2 = MeetEdge(
source=d1,
destination=u1
)
tx.add(u1_meet_u2)
tx.add(u2_meet_u1)
#
# HERE STARTS THE QUERY
#
# Execute will return a list of lists: [
# [UserNode("u1"), UserNode("u2")],
# [UserNode("u1"), UserNode("u2")],
# ...
# ]
q = gh.Query.raw(
"match (u1:User)-[:Meet]->(u2:User) return u1, u2"
).execute(map_to={"u1": UserNode, "u2": UserNode})
print(q)
if __name__ == '__main__':
count = 1000
create_graph("redis://127.0.0.1:7000/?graph=email", nodes_count=count)
create_graph("neo4j://neo4j:s3cr3t@127.0.0.1:7687/?graph=email", nodes_count=count)
Transactions are useful if you need add a lot of data. You add nodes and edges to a transaction. When they finish then perform the insertions to the database in a very efficient way:
from dataclasses import dataclass
from sqerzo import GraphEdge, GraphNode, SQErzoGraph
class MeetEdge(GraphEdge):
pass
@dataclass
class UserNode(GraphNode):
__keys__ = ["name"]
name: str = None
def create_graph(connection_string: str):
gh = SQErzoGraph(connection_string)
gh.truncate() # Drop database
with gh.transaction() as tx: # Transaction starts here
for n in range(500): # Inserts 1000 nodes (500 * 2) and 500 relations
u1 = UserNode(name=f"UName-{n}")
d1 = UserNode(name=f"DName-{n}")
tx.add(u1)
tx.add(d1)
u1_meet_g1 = MeetEdge(
source=u1,
destination=d1
)
tx.add(u1_meet_g1)
if __name__ == '__main__':
print("Redis...")
create_graph("redis://127.0.0.1:7000/?graph=email")
print("Neo4j...")
create_graph("neo4j://neo4j:s3cr3t@127.0.0.1:7687/?graph=email")
If you need a more complex example, you can find in it examples/email_graph.py.
At this example we load a random generated mail inbox (generation script is also available) into a Graph Database following this Neo4j Blog Post suggestions.
- Added queries support for raw queries in DB engine language
- Improved speed at insertion by 100x
- Add support for
UNIQUE
create_constraints_nodes - Add support for
INDEXES
create_constraints_nodes - Add support for raw Cypher query
- Errors, issues, new features and something else
- Complete refactor to easy add new backends
- Complete refactor to easy add new backends
- Add new methods: fetch_many, fetch_one, raw_query, save, update & transaction
- Add new examples
- Improved the way to build the Node to avoid waste memory.
- Improve documentation
- Improve cypher query to avoid query raises when a transaction insert a duplicate node
- Add support for Arango DB
- Add support for AWS Neptune
- Add support for Gremlin
- Add support for dates to RedisGraph using transformation of dates to numbers
- Implementation of Query builder. Add some method to
Query
builder class. Here some possible examples:
from dataclasses import dataclass
from sqerzo import GraphEdge, GraphNode, SQErzoGraph
class MeetEdge(GraphEdge):
pass
class WorksWithEdge(GraphEdge):
pass
@dataclass
class UserNode(GraphNode):
__keys__ = "email"
name: str = None
email: str = None
@dataclass
class OtherUserNode(GraphNode):
__keys__ = "email"
name: str = None
email: str = None
gh = SQErzoGraph("redis://")
gh.Q().from(Node1).to(node2).execute()
gh.Q().from(name="me").to(UserNode).execute()
gh.Q().from(name="me", email="me@me.com").to((UserNode, "User")).execute()
gh.Q().from(name="me").across((WorksWithEdge, "WorksWith")).to((UserNode, "OtherUser")).execute()
gh.Q().to((UserNode, "OtherUser")).execute()
gh.Q().from(OtherUserNode).execute()
gh.Q().from(UserNode).execute()
I tried to use good practices for building SQErzo
. Some references I used:
- https://medium.com/neo4j/cypher-query-optimisations-fe0539ce2e5c
- https://hub.packtpub.com/advanced-cypher-tricks/
- https://gist.github.com/jexp/caeb53acfe8a649fecade4417fb8876a
This project is distributed under BSD license