Octopoes
Octopoes is KAT’s knowledge-graph. It stores the knowledge KAT has gathered about its domain. As Octopoes uses XTDB for bi-temporal data-storage, Octopoes keeps the current state of the knowledge-graph, as well as a complete, queryable history of the knowledge-graph.
Instructions
Install dependencies
python3 -m pip install -r requirements.txt
Run Octopoes API
python3 -m uvicorn octopoes.api.api:app [--port 8000]
Run the event processor
python3 -m celery -A octopoes.tasks.tasks worker -B -s /tmp/celerybeat-schedule --loglevel=WARNING
Note: The -B flag instructs celery start the Celery Beat scheduler in the same process
Note: The -s flag is used to specify the beat schedule location and should be writeable by the user the process runs in
Healthcheck
# Return XTDB connection info
curl http://localhost:8000/_dev/health
# Return some XTDB objects (or empty list [])
curl http://localhost:8000/_dev/objects
# To request data for a different KAT-client:
curl http://localhost:8000/clientx/healthcheck
OOI
The domain of discourse, on which Octopoes operates, is described by the OOI datamodel. The OOI (Object of Interest) model is described by entities and relations between them. The OOI model is currently defined in Octopoes itself (module octopoes.model.ooi). However, it is planned to be defined in the Openkat schema registries, decoupling the applicable domain from the logic.
Origin
Each OOI must have an origin to exist in the knowledge-graph. Origins can be supplied to Octopoes in 3 ways:
origin through declaration
origin through observation
origin through inference
Each origin consists of:
the identifier of the origin-method
a source OOI
a set of result OOIs
additional metadata. E.g. the task-ID that made the observation
Origin through declaration
An OOI is declared to exist by a user of KAT.
In this case, OOI B is both source and result
Origin through observation
An observation is reported by a normalizer
An observation has a key that identifies the normalizer
An observation always has a source OOI
An observation always has a (possibly empty) set of result OOIs
Origin through inference
An object is inferred from other objects in the knowledge-graph. This is achieved by rules, declared in bits. A bit is a rule that is applied to a pattern in the knowledge-graph.
Graph mutations
Mutations can only be made by supplying an origin to Octopoes. This can be an origin through declaration, or origin through observation. When, after an origin-update, an OOI is no longer referenced by any origin. The OOI will be deleted from the knowledge-graph.
Example: observation O has result B and C
After a mutation, observation O has result B. C is no longer referenced, and is deleted from the knowledge-graph.
If C had been referenced by another origin, it would not have been deleted.
OOI C is not deleted, since it’s still referenced by Observation P
Code Architecture
In high level, the code architecture is as follows:
Origin gets reported to the API
API calls the service layer
Service layer calls the data layer
Data layer sends out a mutation event
Listener catches the mutation event
Listener calls service layer to process mutation
Sequence: save_origin
Sequence: process update ooi
Crux / XTDB
Crux is the central database of OOIs within KAT. Crux is a graph-database that can store objects (schemalessly), while providing object history and audit-trail functionality out-of-the-box. The term bitemporal means it tracks every object on 2 time axis: valid-time and transaction-time.
Valid-time means the state of an object at a certain time X (mutable).
Transaction-time means the state of an object with all transactions-processed until time Y (immutable)
This is especially useful for forensics-type queries like: What was the state of an object at time X (valid-time), with the information we had at time Y (transaction-time).
Good to know: Crux tracks the history of each object by its primary key.
OOI
OOI objects are instances of relatively simple classes, which inherit from OOIBase.
Because all OOIs are stored in Crux and Crux tracks object history by primary key, KAT defines a way to reliably determine the primary key of an object by its attributes. This is called the natural key of an object.
The main advantage of this method, is that when enough attributes of an OOI are discovered, the primary key of this object is known. This allows reasoning about the exact same objects in several subsystems, without having to query a database.
Consider this (oversimplified) Person class
from octopoes.models import OOI
class Person(OOI):
name: str
last_name: str
age: int
_natural_key_attributes: ['name', 'last_name']
# 2 completely separate systems can instantiate the the following Person OOI:
john = Person(name='John', last_name='Doe', age=42)
# And without having to search this person in a central database, the primary key is known:
john.natural_key # 'John/Doe'
john.primary_key # 'Person/John/Doe'
Note that the primary key consists of the natural key prefixed by the OOI-type, to avoid PK collisions
Relationships
OOIs can be related to each other. At time of writing the OOI data structure looks like this:
Directional arrows indicate a foreign key pointing to referred object
In a one-to-many relationship (A 1-* B), the relationship is stored in B (B points to A). For example, an IP-address belongs to a Network. So the Network primary key is stored as a foreign key in the IP-address object.
from octopoes.models import OOI, Reference
class Network(OOI):
name: str
_natural_key_attrs = ['name']
class IpAddressV6(OOI):
Network: Reference[Network]
address: str
_natural_key_attrs = ['Network', 'address']
A few example records
OOI Reference
Even though foreign keys are actually simple strings, for ease of use these strings are represented in Octopoes by a special Reference object.
OOIRefs can be obtained in several ways.
from octopoes.models import Reference
from octopoes.models.ooi.network import Network, IPAddressV6
# Through the .ref() method of an OOI instance
internet = Network(name='internet')
internet_ref = internet.reference
# Or from string
internet_ref = Reference.from_str('Network|internet')
# Create a related object with a ref
ip = IPAddressV6(network=internet_ref, address='2001:db8::1')
Since an OOIRef is a compound key, individual parts of the foreign key can be retrieved by the parsed property.
from octopoes.models import Reference
ref = Reference.from_str('IpPort|internet|2001:db8::1|tcp|5050')
ref.tokenized.protocol # 'tcp'
ref.tokenized.port # '5050'
ref.tokenized.address.address # '2001:db8::1'
Octopoes API
OctopoesAPIConnector
The OctopoesAPIConnector class provides a python interface for connecting with Octopoes API.
Abstract classes / subclassing
Relationships from an OOI class to another OOI class are inferred through its property types. It is possible to define a relationship to an abstract class.
For querying purposes and grouping purposes these abstract classes can also be used.
from octopoes.models import OOI, Reference
from octopoes.connector.octopoes import OctopoesAPIConnector
# Define abstract class and subclasses
class IPAddress(OOI):
...
class IPAddressV4(IPAddress):
...
class IPAddressV6(IPAddress):
...
# Relationships to abstract class
class IPPort(OOI):
address: Reference[IPAddress] # Any subclass of IPAddress (IPAddressV4, IPAddressV6)
protocol: str
port: int
class TagExample(OOI):
ooi: Reference[OOI] # Any subclass of OOI..
tag: str
# Query abstract class
OctopoesAPIConnector('http://octopoes', '_dev').list({IPAddress})
Querying
Octopoes API uses the OOI model to construct Crux queries. For complex graph-querying, Crux’s pull-syntax is used to build a query tree. Crux can join objects to properties which hold (lists of) foreign keys.
Imagine a query “Give me IpAddressV4 with primary key X and all related objects 2 levels deep”.
What happens under the hood:
A relation map is created with all OOI classes and their relations
A query plan is created by traversing the relation map 2 levels deep. The queryplan is a tree of QueryNode objects
The query plan is transformed into a Crux Datalog query, utilizing its pull syntax to join related objects
Rules: A few rules come into play when planning the query.
Relations are not traversed back through the previous relation. E.g.:
IpAddressV4 -> IpPort -> IpAddressV4Leaf nodes are OOI classes that have too many relations to effectively traverse if they are not the starting node. Currently these are
Network,FindingandJob
Query Plan Visualization: The OOI class tree is traversed 2 levels deep. Bear in mind that both Finding and Job can be related to any OOI, so the following paths are valid:
- IpAddressV4 -> Finding
- IpAddressV4 -> Job
- IpAddressV4 -> DnsARecord -> Finding
- IpAddressV4 -> DnsARecord -> Job
- IpAddressV4 -> IpPort -> Finding
- IpAddressV4 -> IpPort -> Job
- IpAddressV4 -> IpPort -> IpService
- IpAddressV4 -> DnsARecord -> Hostname
....etc
Hence the 1 and 2 levels markers on Finding and Job in the image below.
Tests
The unit tests octopoes/tests are run using
python -m unittest discover octopoes/tests