Attribute-based access control (ABAC) SDK for Python

ACL, ABAC, access-control, policy, security, authorization, permission, attribute-based-access-control, authz, iam-policy, permissions, policy-management, python
pip install vakt==1.5.0


Vakt logo

Attribute-based access control (ABAC) SDK for Python.

CI Status PyPI version Apache 2.0 licensed



Vakt is an attribute-based and policy-based access control (ABAC) toolkit that is based on policies. ABAC stands aside of RBAC and ACL models, giving you a fine-grained control on definition of the rules that restrict an access to resources and is generally considered a "next generation" authorization model. In its form Vakt resembles IAM Policies, but has a way nicer attribute managing.

See concepts section for more details.

Back to top


Given you have some set of resources, you can define a number of policies that will describe access to them answering the following questions:

  1. What resources (resource) are being requested?
  2. Who is requesting the resource?
  3. What actions (action) are requested to be done on the asked resources?
  4. What are the rules that should be satisfied in the context of the request itself?
  5. What is resulting effect of the answer on the above questions?

The overall diagram of vakt workflow is:

Vakt diagram

Vakt allows you to gain:

  • Policy Based Access Control (vakt is based on Policies that describe access rules, strategies to your resources)
  • Fine-Grained Authorization (vakt Policies give you fine-grained control over resource's, subject's, action's and context's attributes)
  • Dynamic Authorization Management (you can add Policies and change their attributes)
  • Externalized Authorization Management (you can build own external AuthZ server with vakt, see examples)

Back to top


Vakt runs on Python >= 3.6. PyPy implementation is supported as well.

Bare-bones installation with in-memory storage:

pip install vakt

For MongoDB storage:

pip install vakt[mongo]

For SQL storage:

pip install vakt[sql]

For Redis storage:

pip install vakt[redis]

Also see redis-py docs. For example if hiredis is found in the system, it will be used as a faster parser. However vakt doesn't enforce this dependency.

Back to top


A quick dive-in:

import vakt
from vakt.rules import Eq, Any, StartsWith, And, Greater, Less

policy = vakt.Policy(
    actions=[Eq('fork'), Eq('clone')],
    resources=[StartsWith('repos/Google', ci=True)],
    subjects=[{'name': Any(), 'stars': And(Greater(50), Less(999))}],
    context={'referer': Eq('')},
    Allow to fork or clone any Google repository for
    users that have > 50 and < 999 stars and came from Github
storage = vakt.MemoryStorage()
guard = vakt.Guard(storage, vakt.RulesChecker())

inq = vakt.Inquiry(action='fork',
                   subject={'name': 'larry', 'stars': 80},
                   context={'referer': ''})

assert guard.is_allowed(inq)

Or if you prefer Amazon IAM Policies style:

import vakt
from vakt.rules import CIDR

policy = vakt.Policy(
    subjects=[r'<[a-zA-Z]+ M[a-z]+>'],
    resources=['library:books:<.+>', 'office:magazines:<.+>'],
    actions=['read', 'get'],
        'ip': CIDR(''),
    Allow all readers of the book library whose surnames start with M get and read any book or magazine,
    but only when they connect from local library's computer
storage = vakt.MemoryStorage()
guard = vakt.Guard(storage, vakt.RegexChecker())

inq = vakt.Inquiry(action='read',
                   subject='Jim Morrison',
                   context={'ip': ''})

assert guard.is_allowed(inq)

For more examples see here.

Back to top



Policy is a main object for defining rules for accessing resources. The main parts reflect questions described in Concepts section:

  • resources - a list of resources. Answers: what is asked?
  • subjects - a list of subjects. Answers: who asks access to resources?
  • actions - a list of actions. Answers: what actions are asked to be performed on resources?
  • context - rules that should be satisfied by the given inquiry's context.
  • effect - If policy matches all the above conditions, what effect does it imply? Can be either vakt.ALLOW_ACCESS or vakt.DENY_ACCESS

All resources, subjects and actions are described with a list containing strings, regexes, Rules or dictionaries of strings (attributes) to Rules. Each element in list acts as logical OR. Each key in a dictionary of Rules acts as logical AND. context can be described only with a dictionary of Rules.

Depending on a way resources, subjects, actions are described, Policy can have either String-based or Rule-based type. Can be inspected by policy.type. This enforces the use of a concrete Checker implementation. See Checker for more.

from vakt import Policy, ALLOW_ACCESS
from vakt.rules import CIDR, Any, Eq, NotEq, In

# Rule-based policy (defined with Rules and dictionaries of Rules)
    Allow access to administration interface subcategories: 'panel', 'switch' if user is not
    a developer and came from local IP address.
    resources=[{'category': Eq('administration'), 'sub': In(['panel', 'switch'])}],
    subjects=[{'name': Any(), 'role': NotEq('developer')}],
    context={'ip': CIDR('')}

# String-based policy (defined with regular expressions)
    Allow all readers of the book library whose surnames start with M get and read any book or magazine,
    but only when they connect from local library's computer
    subjects=['<[\w]+ M[\w]+>'],
    resources=('library:books:<.+>', 'office:magazines:<.+>'),
    context={'ip': CIDR('')}

Basically you want to create some set of Policies that encompass access rules for your domain and store them for making future decisions by the Guard component.

st = MemoryStorage()
for p in policies:

Additionally you can create Policies with predefined effect classes:

from vakt import PolicyAllow, PolicyDeny, ALLOW_ACCESS, DENY_ACCESS

p = PolicyAllow(1, actions=['<read|get>'], resources=['library:books:<.+>'], subjects=['<[\w]+ M[\w]+>'])
assert ALLOW_ACCESS == p.effect

p = PolicyDeny(2, actions=['<read|get>'], resources=['library:books:<.+>'], subjects=['<[\w]+ M[\w]+>'])
assert DENY_ACCESS == p.effect

Back to top


Inquiry is an object that serves as a mediator between Vakt and outer world request for resource access. All you need to do is take any kind of incoming request (REST request, SOAP, etc.) and build an Inquiry out of it in order to feed it to Vakt. There are no concrete builders for Inquiry from various request types, since it's a very meticulous process and you have hands on control for doing it by yourself. Let's see an example:

from vakt import Inquiry
from flask import request, session


# if policies are defined on some subject's and resource's attributes with dictionaries of Rules:
inquiry2 = Inquiry(subject={'login': request.form['username'], 'role': request.form['user_role']},
                   resource={'book': session.get('book'), 'chapter': request.form['chapter']},
                   context={'ip': request.remote_addr})

# if policies are defined with strings or regular expressions:
inquiry = Inquiry(subject=request.form['username'],
                  context={'ip': request.remote_addr})

Here we are taking form params from Flask request and additional request information. Then we transform them to Inquiry. That's it.

Inquiry has several constructor arguments:

  • resource - any | dictionary of str -> any. What resource is being asked to be accessed?
  • action - any | dictionary str -> any. What is being asked to be done on the resource?
  • subject - any | dictionary str -> any. Who asks for it?
  • context - dictionary str -> any. What is the context of the request?

If you were observant enough you might have noticed that Inquiry resembles Policy, where Policy describes multiple variants of resource access from the owner side and Inquiry describes an concrete access scenario from consumer side.

Back to top


Rules allow you to describe conditions directly on action, subject, resource and context or on their attributes. If at least one Rule in the Rule-set is not satisfied Inquiry is rejected by given Policy.

Attaching a Rule-set to a Policy is simple. Here are some examples:

from vakt import Policy, rules

    subjects=[{'name': rules.Eq('Tommy')}],

    actions=[rules.Eq('get'), rules.Eq('list'), rules.Eq('read')],

        'secret': rules.string.Equal('.KIMZihH0gsrc'),

There are a number of different Rule types, see below.

If the existing Rules are not enough for you, feel free to define your own.

Rule Example in Policy Example in Inquiry Notes
Eq 'age': Eq(40) 'age': 40
NotEq 'age': NotEq(40) 'age': 40
Greater 'height': Greater(6.2) 'height': 5.8
Less 'height': Less(6.2) 'height': 5.8
GreaterOrEqual 'stars': GreaterOrEqual(300) 'stars': 77
LessOrEqual 'stars': LessOrEqual(300) 'stars': 300
Rule Example in Policy Example in Inquiry Notes
Truthy 'admin': Truthy() 'admin': user.is_admin() Evaluates on Inquiry creation
Falsy 'admin': Falsy() 'admin': lambda x: x.is_admin() Evaluates on Inquiry creation
Not 'age': Not(Greater(90)) 'age': 40
And 'stars': And(Greater(50), Less(89)) 'stars': 78 Also, attributes in dictionary of Rules act as AND logic
Or 'stars': Or(Greater(50), Less(120), Eq(8888)) 'stars': 78 Also, rules in a list of, say, actions act as OR logic
Any actions=[Any()] action='get', action='foo' Placeholder that fits any value
Neither subjects=[Neither()] subject='Max', subject='Joe' Not very useful, left only as a counterpart of Any
Rule Example in Policy Example in Inquiry Notes
In 'method': In('get', 'post') 'method': 'get'
NotIn 'method': NotIn('get', 'post') 'method': 'get'
AllIn 'name': AllIn('Max', 'Joe') 'name': ['Max', 'Joe']
AllNotIn 'name': AllNotIn('Max', 'Joe') 'name': ['Max', 'Joe']
AnyIn 'height': AnyIn(5.9, 7.5, 4.9) 'height': [7.55]
AnyNotIn 'height': AnyNotIn(5.9, 7.5, 4.9) 'height': [7.55]
Rule Example in Policy Example in Inquiry Notes
CIDR 'ip': CIDR('') 'ip':
Rule Example in Policy Example in Inquiry Notes
Equal 'name': Equal('max', ci=True) 'name': 'Max' Aliased as StrEqual. Use instead of Eq it you want string-type check and case-insensitivity
PairsEqual 'names': PairsEqual() 'names': ['Bob', 'Bob'] Aliased as StrPairsEqual
RegexMatch 'file': RegexMatch(r'\.rb$') 'file': 'test.rb'
StartsWith 'file': StartsWith('logs-') 'file': 'logs-data-101967.log' Supports case-insensitivity
EndsWith 'file': EndsWith('.log') 'file': 'logs-data-101967.log' Supports case-insensitivity
Contains 'file': Contains('sun') 'file': 'observations-sunny-days.csv' Supports case-insensitivity

Inquiry-related rules are useful if you want to express equality relation between inquiry elements or their attributes.

Rule Example in Policy Example in Inquiry Notes
SubjectMatch resources=[{'id': SubjectMatch()}] Inquiry(subject='Max', resource={'id': 'Max'}) Works for the whole subject value or one of its attributes
ActionMatch subjects=[ActionMatch('id')] Inquiry(subject='Max', action={'method': 'get', id': 'Max'}) Works for the whole action value or one of its attributes
ResourceMatch subjects=[ResourceMatch('id')] Inquiry(subject='Max', resource={'res': 'book', id': 'Max'}) Works for the whole resource value or one of its attributes
SubjectEqual 'data': SubjectEqual() Inquiry(subject='Max') Works only for strings. Favor SubjectMatch
ActionEqual 'data': ActionEqual() Inquiry(action='get') Works only for strings. Favor ActionMatch
ResourceIn 'data': ResourceIn() Inquiry(resource='/books/') Works only for strings. Favor ResourceMatch

Back to top


Checker allows you to check whether Policy matches Inquiry by concrete field (subject, action, etc.). It's used internally by Guard, but you should be aware of Checker types:

  • RulesChecker - universal type that is used to check match of Policies defined with Rules or dictionaries of Rules (Rule-based Policy type). It gives you the highest flexibility. Most of the time you will use this type of Polices and thus this type of a Checker. Besides, it's much more performant than RegexChecker. See benchmark for more details.
from vakt import RulesChecker

ch = RulesChecker()
# etc.
  • RegexChecker - checks match by regex test for policies defined with strings and regexps (String-based Policy type). This means that all you Policies can be defined in regex syntax (but if no regex defined in Policy falls back to simple string equality test) - it gives you better flexibility compared to simple strings, but carries a burden of relatively slow performance. You can configure a LRU cache size to adjust performance to your needs:
from vakt import RegexChecker

ch = RegexChecker(2048)
ch2 = RegexChecker(512)
# etc.

See benchmark for more details.

Syntax for description of Policy fields is:


Where <> are delimiters of a regular expression boundaries part. Custom Policy can redefine them by overriding start_tag and end_tag properties. Generally you always want to use the first variant: <foo.*>.

Due to relatively slow performance of regular expressions execution we recommend to define your policies in regex syntax only when you really need it, in other cases use simple strings: both will work perfectly (and now swiftly!) with RegexChecker.

NOTE. All regex checks are performed in a case-sensitive way by default. Even thought some storages (e.g. MemoryStorage) allow you to specify regex modifiers within the regex string, we do not translate regex modifiers to all storages (e.g. SQLStorage). Also see warning below

WARNING. Please note, that storages have varying level of regexp support. For example, most SQL databases allow to use POSIX metacharacters whereas python re module and thus MemoryStorage does not. So, while defining policies you're safe and sound as long as you understand how storage of your choice handles the regexps you specified.

  • StringExactChecker - the most quick checker:
Checker that uses exact string equality. Case-sensitive.
E.g. 'sun' in 'sunny' - False
     'sun' in 'sun' - True
  • StringFuzzyChecker - quick checker with some extent of flexibility:
Checker that uses fuzzy substring equality. Case-sensitive.
E.g. 'sun' in 'sunny' - True
     'sun' in 'sun' - True

Note, that some Storage handlers can already check if Policy fits Inquiry in find_for_inquiry() method by performing specific to that storage queries - Storage can (and generally should) decide on the type of actions based on the checker class passed to Guard constructor (or to find_for_inquiry() directly).

Regardless of the results returned by a Storage the Checker is always the last row of control before Vakt makes a decision.

Back to top


Guard component is a main entry point for Vakt to make a decision. It has one method is_allowed that passed an Inquiry gives you a boolean answer: is that Inquiry allowed or not?

Guard is constructed with Storage and Checker.

Policies that have String-based type won't match if RulesChecker is used and vise-versa.

st = MemoryStorage()
# And persist all our Policies so that to start serving our library.
for p in policies:

guard = Guard(st, RulesChecker())

if guard.is_allowed(inquiry):
    return "You've been logged-in", 200
    return "Go away, you violator!", 401

To gain best performance read Caching section.

Back to top


Storage is a component that gives an interface for manipulating Policies persistence in various places.

It provides the following methods:

add(policy)                 # Store a Policy
get(uid)                    # Retrieve a Policy by its ID
get_all(limit, offset)      # Retrieve all stored Policies (with pagination)
retrieve_all(batch)         # Retrieve all existing stored Policies (without pagination)
update(policy)              # Store an updated Policy
delete(uid)                 # Delete Policy from storage by its ID
find_for_inquiry(inquiry)   # Retrieve Policies that match the given Inquiry

Storage may have various backend implementations (RDBMS, NoSQL databases, etc.), they also may vary in performance characteristics, so see Caching and Benchmark sections.

Vakt ships some Storage implementations out of the box. See below:


Implementation that stores Policies in memory. It's not backed by any file or something, so every restart of your application will swipe out everything that was stored. Useful for testing.

from vakt import MemoryStorage

storage = MemoryStorage()

MongoDB is chosen as the most popular and widespread NO-SQL database.

from pymongo import MongoClient
from import MongoStorage

client = MongoClient('localhost', 27017)
storage = MongoStorage(client, 'database-name', collection='optional-collection-name')

Default collection name is 'vakt_policies'.

Actions are the same as for any Storage that conforms interface of base class.

Beware that currently MongoStorage supports indexed and filtered-out find_for_inquiry() only for StringExact, StringFuzzy and Regex (since MongoDB version 4.2 and onwards) checkers. When used with the RulesChecker it simply returns all the Policies from the database.


SQL storage is backed by SQLAlchemy, thus it should support any RDBMS available for it: MySQL, Postgres, Oracle, MSSQL, Sqlite, etc.

Given that we support various SQL databases via SQLAlchemy, we don't specify any DB-specific drivers in the vakt dependencies. It's up to the user to provide a desired one. For example: psycopg2 or PyMySQL.

Example for MySQL.

from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker, scoped_session
from import SQLStorage

engine = create_engine('mysql://root:root@localhost/vakt_db')
storage = SQLStorage(scoped_session=scoped_session(sessionmaker(bind=engine)))

# Don't forget to run migrations here (especially for the first time)

Beware that currently SQLStorage supports indexed and filtered-out find_for_inquiry() only for StringExact, StringFuzzy and Regex checkers. When used with the RulesChecker it simply returns all the Policies from the database.

Note that vakt focuses on testing SQLStorage functionality only for two most popular open-source databases: MySQL and Postgres. Other databases support may have worse performance characteristics and/or bugs. Feel free to report any issues.


Redis storage.

RedisStorate stores all Policies in a hash whose key is the collection name and the hash'es key value pairs are Policy UID -> serialized Policy representation.

Default collection name is "vakt_policies".

You can use different Serializers. Any custom or one of the vakt's native. Just pass it to the RedisStorage constructor.

Vakt is shipped with:

  • JSONSerializer
  • PickleSerializer - the fastest. Used as the default one.

Due to serialization/deserialization Redis is not as fast as simple MemoryStorage. You can run the benchmark and check performance for your use-case.

from redis import Redis
from import RedisStorage

client = Redis('', 6379)
yield RedisStorage(client, collection='optional-policies-collection-name')

Back to top


vakt.migration is a set of components that are useful from the perspective of the Storage. It's recommended to favor it over manual actions on DB schema/data since it's aware of Vakt requirements to Policies data. But it's not mandatory, anyway. However it's up to a particular Storage to decide whether it needs migrations or not. It consists of 3 components:

  • Migration
  • MigrationSet
  • Migrator

Migration allows you to describe data modifications between versions. Each storage can have a number of Migration classes to address different releases with the order of the migration specified in order property. Should be located inside particular storage module and implement Migration has 2 main methods (as you might guess) and 1 property:

  • up - runs db "schema" upwards
  • down - runs db "schema" downwards (rolls back the actions of up)
  • order - tells the number of the current migration in a row

MigrationSet is a component that represents a collection of Migrations for a Storage. You should define your own migration-set. It should be located inside particular storage module and implement It has 3 methods that lest unimplemented:

  • migrations - should return all initialized Migration objects
  • save_applied_number - saves a number of a lst applied up migration in the Storage for later reference
  • last_applied - returns a number of a lst applied up migration from the Storage

Migrator is an executor of a migrations. It can execute all migrations up or down, or execute a particular migration if number argument is provided.

Example usage:

from pymongo import MongoClient
from import MongoStorage, MongoMigrationSet
from import Migrator

client = MongoClient('localhost', 27017)
storage = MongoStorage(client, 'database-name', collection='optional-collection-name')

migrator = Migrator(MongoMigrationSet(storage))

Back to top


Vakt has several layers of caching, that serve a single purpose: speed up policy enforcement decisions. In most situations and use-cases you might want to use them all, thus they are designed not to interact with each other, but rather work in tandem (nonetheless you are free to use any single layer alone or any combination of them). That said let's look at all those layers.

Caching RegexChecker

It's relevant only for RegexChecker and allows to cache parsing and execution of regex-defined Policies, which can be very expensive due to inherently slow computational performance of regular expressions and vakt's parsing. When creating a RegexChecker you can specify a cache size for an in-memory LRU (least recently used) cache. Currently only python's native LRU cache is supported.

# preferably size is a power of 2
chk = RegexChecker(cache_size=2048)

# or simply
chk = RegexChecker(2048)

# or 512 by default
chk = RegexChecker()
Caching the entire Storage backend

Some vakt's Storages may be not very clever at filtering Policies at find_for_inquiry especially when dealing with Rule-based policies. In this case they return the whole set of the existing policies stored in the external storage. Needless to say that it makes your application very heavy IO-bound and decreases performance for large policy sets drastically. See benchmark for more details and exact numbers.

In such a case you can use EnfoldCache that wraps your main storage (e.g. MongoStorage) into another one (it's meant to be some in-memory Storage). It returns you a Storage that behind the scene routes all the read-calls (get, get_all, find_for_inquiry, ...) to an in-memory one and all modify-calls (add, update, delete) to your main Storage ( don't worry, in-memory Storage is kept up-to date with the main Storage). In case a requested policy is not found in in-memory Storage it's considered a cache miss and a request is routed to a main Storage.

Also, in order to keep Storages in sync, when you initialize EnfoldCache the in-memory Storage will fetch all the existing Policies from a main one - therefore be forewarned that it might take some amount of time depending on the size of a policy-set. Optionally you can call populate method after initialization, but in this case do not ever call any modify-related methods of EnfoldCache'd storage before populate(), otherwise Storages will be in an unsynchronized state and it'll result in broken Guard functionality.

from vakt import EnfoldCache, MemoryStorage, Policy, Guard, RegexChecker
from import MongoStorage

storage = EnfoldCache(MongoStorage(...), cache=MemoryStorage())
storage.add(Policy(1, actions=['get']))


guard = Guard(storage, RegexChecker())
Caching the Guard

Guard.is_allowed it the the centerpiece of vakt. Therefore it makes ultimate sense to cache it. And create_cached_guard() function allows you to do exactly that. You need to pass it a Storage, a Checker and a maximum size of a cache. It will return you a tuple of: Guard, Storage and AllowanceCache instance:

  • You must do all policies operations with the returned storage (which is a slightly enhanced version of a Storage you provided to the function).
  • The returned Guard is a normal vakt's Guard, but its is_allowed is cached with AllowaceCache.
  • The returned cache is an instance of AllowaceCache and has a handy method info that provides current state of the cache.

How it works?

Only the first Inquiry will be passed to is_allowed, all the subsequent answers for similar Inquiries will be taken from cache. AllowanceCache is rather coarse-grained and if you call Storage's add, update or delete the whole cache will be invalided because the policy-set has changed. However for stable policy-sets it is a good performance boost.

By default AllowanceCache uses in-memory LRU cache and maxsize param is it's size. If for some reason it does not satisfy your needs, you can pass your own implementation of a cache backend that is a subclass of vakt.cache.AllowanceCacheBackend to create_cached_guard as a cache keyword argument.

guard, storage, cache = create_cached_guard(MongoStorage(...), RulesChecker(), maxsize=256)

p1 = Policy(1, actions=[Eq('get')], resources=[Eq('book')], subjects=[Eq('Max')], effect=ALLOW_ACCESS)

# Given we have some inquiries that tend to repeat
inq1 = Inquiry(action='get', resource='book', subject='Max')
inq2 = Inquiry(action='get', resource='book', subject='Jamey')

assert guard.is_allowed(inq1)
assert guard.is_allowed(inq1)
assert guard.is_allowed(inq1)
assert not guard.is_allowed(inq2)
assert guard.is_allowed(inq1)
assert guard.is_allowed(inq1)

# You can check cache state
assert 4 ==
assert 2 ==
assert 2 ==

Back to top


All Policies, Inquiries and Rules can be JSON-serialized and deserialized.

For example, for a Policy all you need is just run:

from vakt.policy import Policy

policy = Policy('1')

json_policy = policy.to_json()
# {"actions": [], "description": null, "effect": "deny", "uid": "1",
# "resources": [], "context": {}, "subjects": []}

policy = Policy.from_json(json_policy)
# <vakt.policy.Policy object at 0x1023ca198>

The same goes for Rules, Inquiries. All custom classes derived from them support this functionality as well. If you do not derive from Vakt's classes, but want this option, you can mix-in vakt.util.JsonSerializer class.

from vakt.util import JsonSerializer

class CustomInquiry(JsonSerializer):

Back to top


Vakt follows a common logging pattern for libraries:

Its corresponding modules log all the events that happen but the log messages by default are handled by NullHandler. It's up to the outer code/application to provide desired log handlers, filters, levels, etc.

For example:

import logging

root = logging.getLogger()

... # here go all the Vakt calls.

Vakt logs can be comprehended in 2 basic levels:

  1. Error/Exception - informs about exceptions and errors during Vakt work.
  2. Info - informs about incoming inquiries, their resolution and policies responsible for this decisions ('vakt.guard' and 'vakt.audit' streams).

Back to top


Vakt allows you to not only watch the incoming inquiries and their resolution, but also keep track of the policies that were responsible for allowing or rejecting the inquiry. It's done via audit logging.

Audit logging is implemented within a standard Python logging framework. You can enable it by subscribing to an audit ('vakt.audit') logging "stream".

Example of configuration in the code:

import logging

logger = logging.getLogger('vakt.audit')

fmt = 'msg: %(message)s | effect: %(effect)s | deciders: %(deciders)s | candidates: %(candidates)s | inquiry: %(inquiry)s'
fileHandler = logging.FileHandler('test.log')

... # here go all the Vakt calls.

Vakt logs all audit records at the INFO level.

The formatter supports the following fields:

  • message - the message that tells what and why happened in the audit.
  • effect - effect that this decision has: 'allow' or 'deny'.
  • candidates - potential policies that were filtered by storage and checkers and may be responsible for the decision.
  • deciders - policies that are responsible for the final decision.
  • inquiry - the inquiry in question.
  • all the standard Python logging fields like time, level, module name, etc.

The deciders and candidates field can be logged in various ways depending on the the audit_policies_cls. It can be passed to the Guard constructor.

Vakt has the following Audit Policies messages classes out of the box:

  • PoliciesNopMsg
  • PoliciesUidMsg (is the default one)
  • PoliciesDescriptionMsg
  • PoliciesCountMsg

Refer to their documentation on how they represent the policies.

WARNING. Please note, that if you have Guard caching enabled, then audit records for the same subsequent inquiries won't be logged because the calls are cached. However the log records from 'vakt.guard' stream will be always logged - they will tell only was the inquiry allowed or not.

Back to top


Most valuable features to be implemented in the order of importance:

  • SQL Storage
  • Rules that reference Inquiry data for Rule-based policies
  • Caching mechanisms (for Storage and Guard)
  • YAML-based language for declarative policy definitions
  • Enhanced audit logging
  • Redis Storage

Back to top


You can see how much time it takes for a single Inquiry to be processed given we have a number of unique Policies in a Storage. For MemoryStorage it measures the runtime of a decision-making process for all the existing Policies when Guard's code iterates the whole list of Policies to decide if Inquiry is allowed or not. In case of other Storages the mileage may vary since they may return a smaller subset of Policies that fit the given Inquiry. Don't forget that most external Storages add some time penalty to perform I/O operations. The runtime also depends on a Policy-type used (and thus checker): RulesChecker performs much better than RegexChecker.


python3 --checker regex --storage memory -n 1000

Output is:

Populating MemoryStorage with Policies
Number of unique Policies in DB: 1,000
Among them Policies with the same regexp pattern: 0
Checker used: RegexChecker
Storage used: MemoryStorage
Number of concurrent threads: 1
Decision for Inquiry took (mean: 0.2062 seconds. stdev: 0.0000)
Inquiry passed the guard? False

Script usage:

usage: [-h] [-n [POLICIES_NUMBER]] [-s {mongo,memory,sql,redis}] [-d [SQL_DSN]] [-c {regex,rules,exact,fuzzy}]
                    [-t [THREADS]] [--regexp] [--same SAME] [--cache CACHE] [--serializer {json,pickle}]

Run vakt benchmark.

optional arguments:
  -h, --help            show this help message and exit
                        number of policies to create in DB (default: 100000)
  -s {mongo,memory,sql,redis}, --storage {mongo,memory,sql,redis}
                        type of storage (default: memory)
  -d [SQL_DSN], --dsn [SQL_DSN]
                        DSN connection string for sql storage (default: sqlite:///:memory:)
  -c {regex,rules,exact,fuzzy}, --checker {regex,rules,exact,fuzzy}
                        type of checker (default: regex)
  -t [THREADS], --threads [THREADS]
                        number of concurrent requests (default: 1)

regex policy related:
  --regexp              should Policies be defined without Regex syntax? (default: True)
  --same SAME           number of similar regexps in Policy
  --cache CACHE         number of LRU-cache for RegexChecker (default: RegexChecker's default cache-size)

Redis Storage related:
  --serializer {json,pickle}
                        type of serializer for policies stored in Redis (default: json)

Back to top


Initial code ideas of Vakt are based on Amazon IAM Policies and Ladon Policies SDK as its reference implementation.

Back to top


To hack Vakt locally run:

$ ...                              # activate virtual environment w/ preferred method (optional)
$ pip install -e .[dev,mongo,sql,redis]  # to install all dependencies
$ pytest -m "not integration"      # to run non-integration tests with coverage report
$ pytest --cov=vakt tests/         # to get coverage report
$ pylint vakt                      # to check code quality with PyLint

To run only integration tests (for Storage adapters other than MemoryStorage):

$ docker run --rm -d -p 27017:27017 mongo
$ # run sql and Redis database here as well...
$ pytest -m integration
$ pytest -m sql_integration

Optionally you can use make to perform development tasks.

Back to top


The source code is licensed under Apache License Version 2.0

Back to top