KeycloakRBACAuthorizer has been superseded by KeycloakAuthorizer and works in both Zookeeper and KRaft mode
While KeycloakRBACAuthorizer can still be used in Zookeeper mode, for the future you should migrate your configuration to use KeycloakAuthorizer:
In your server.properties use:
authorizer.class.name=io.strimzi.kafka.oauth.server.authorizer.KeycloakAuthorizer
As part of supporting KRaft mode the grants mapping logic has changed slightly. Rather than using the access token as a unit of grant, the user id is now used. This results in better sharing of the grants between sessions of the same user, and should also reduce the number of grants held in cache, and the number of refresh requests to the Keycloak server.
Due to these changes additional configuration options have been added:
strimzi.authorization.grants.max.idle.time.secondsspecifies the time after which an idle grant in the cache can be garbage collectedstrimzi.authorization.grants.gc.period.secondsspecifies an interval in which cleaning of stale grants from grants cache is performed
Also, as a result the option strimzi.authorization.reuse.grants now defaults to true, and no longer to false.
See PR 188
Due to integration difficulties of OAuth metrics with Kafka metrics system the OAuth has to instantiate its own copy of metric reporters.
It turns out that some metric reporters don't work correctly when instantiated multiple times. To address that, we no longer use Kafka's metric.reporters configuration.
If strimzi.oauth.metric.reporters is not set OAuth metrics will still instantiate a default org.apache.kafka.common.metrics.JmxReporter if any OAuth metrics are enabled.
In order to install some other metric reporter in addition to JmxReporter both have to be listed.
Also, the suggested way to configure it on the Kafka broker is to set it as an env variable, rather than a property in server.properties file:
export OAUTH_ENABLE_METRICS=true
export STRIMZI_OAUTH_METRIC_REPORTERS=org.apache.kafka.common.metrics.JmxReporter,org.some.package.SomeReporter
bin/kafka-server-start.sh config/server.properties
See PR 193
It is now possible to use JsonPath query to target nested attributes when extracting a principal. For example:
oauth.username.claim="['user.info'].['user.id']"
oauth.fallback.username.claim="['user.info'].['client.id']"
See PR 194
This change introduces a backwards incompatible change in how queries using equals or not equals comparison to null are handled.
Previously the query "@.missing == null" where JWT token claim called missing was not present in the token would evaluate to false.
Similarly the query "@.missing != null" would evaluate to true.
Such behavior is clearly non-intuitive, and was recognised as a bug and fixed in the json-path library.
By bumping the version of json-path to 2.8.0 the behaviour is now fixed. The query "@.missing == null" evaluates to true, and
"@.missing != null" evaluates to false.
The documentation in README.md has always contained a note that one should not use null comparison in the queries.
Those who followed that rule will not be affected.
Project can now be compiled and tests performed by Java 8, Java 11, and Java 17.
The option was ignored due to a bug.
Some optimization have been done to reduce the number of grants requests to the Keycloak.
A retry mechanism for unexpected failures was added. A configuration option strimzi.authorization.http.retries was introduced, that if set to anvalue greater than zero,
results in the initial grants request for the session be immediately repeated upon failure for up to the specified number of times.
Introduced new configuration options oauth.http.retries and oauth.http.retry.pause.millis that can be used to enable
automatically retrying failed requests to the authorization server during authentication (to the token endpoint), and
during token validation (to the introspection and userinfo endpoints).
Added support for OAuth related metrics. It is disabled by default. To enable it set oauth.enable.metrics OAuth configuration option to true. Use metrics.reporters, and other Kafka configuration metrics.* options to configure the behaviour of metrics capture and how they are exported.
See README.md for details.
The Resource Owner Password Credentials support was added for interoperability in existing corporate environments where established security policies prevent using client credentials to authenticate the client applications. The reason can also be purely technical in that the existing Identity and Access Management solution (IAM) only supports user accounts, even where the 'user' is actually an application service.
See README.md for details.
Set this option to true in order to use all the keys in the JWKS response for token signature validation, regardless of their use attribute.
This makes it possible to use authorization servers that don't specify use attribute in JWKS keys.
Truststores with empty password are now supported
Before, when Kafka client or broker connected to the authorization server during authentication or token validation, there was no connect timeout and no read timeout applied. As a result, if a reverse proxy was in front of the authorization server or a network component glitch prevented normal connectivity, it could happen that the authentication request would stall for a long time.
In order to address this, the default connect timeout and read timeout are now both set to 60 seconds and they are configurable via oauth.connect.timeout.seconds and oauth.read.timeout.seconds.
Added an authentication time mechanism on the broker where a JsonPath query can be configured to extract a set of groups from a JWT token during authentication. A custom authorizer can then retrieve this information through OAuthKafkaPrincipal object available during the authorize() call.
When writing a custom authorizer you may need access to the already parsed JWT token or a map of claims returned by the introspection endpoint. A getJSON() method has been added to BearerTokenWithPayload.
The authorizers have been migrated to Authorizer API that has been introduced in Kafka 2.4.0. As a result the authorizer no longer works in Kafka 2.3.x and earlier versions.
The logging to .grant and .deny logs now takes into account hints from Kafka about whether the authorization decision for specific action should be logged or not.
Fixed a bug when parsing a kafka-cluster section of a Keycloak authorization services resource pattern.
If Keycloak authorization services grants were targeted using a pattern ending with * as in kafka-cluster:*,Topic:my_topic or kafka-cluster:prod_*,Topic:my_topic, the parsing was invalid and resulted in matching of the grant rule to always fail (authorization was denied).
Using just Topic:my_topic would correctly match any cluster, and kafka-cluster:my-cluster,Topic:my_topic would match only if my-cluster was set as a cluster name.
If multiple producers or consumers were used concurrently with the same credentials there was a high likelihood of principal presenting as KafkaPrincipal rather than OAuthKafkaPrincipal after successful authentication. As a result, custom authorizer would not recognise and properly match such a session during authorization check. Depending on the custom authorizer it could result in the delegation of authorization decisions to ACL Authorizer, or it might result in denial of permissions.
When preparing an https connection to authorization server the reported error would say that the URL was malformed, and the actual cause was not logged.
By default enabled option oauth.check.access.token.type triggeres a token type check which checks that the value of typ claim in JWT token is set to Bearer. If typ claim is not present it now falls back to checking if token_type claim with value Bearer is present in the access token.
PEM certificates can now be used directly without being converted to Java Keystore or PKCS12 formats.
To use PEM certificates, set the oauth.ssl.truststore.type option to PEM and either specify location of the PEM file in oauth.ssl.truststore.location or set the certificates directly in oauth.ssl.truststore.certificates.
Now JWT token validation uses a different third-party library. As a result ECDSA support no longer requires the BouncyCastle library. Also, some JWT tokens that would fail previously, can now be handled, widening the support of different authorization servers.
Sometimes authorization server may require audience option to be passed when authenticating to the token endpoint.
Pass the configured oauth.scope option on the Kafka broker as scope when performing clientId + secret authentication on the broker
While the option has existed, it was only used for inter-broker authentication, but not for OAuth over PLAIN.
It is now possible to not specify the oauth.token.endpoint.uri parameter when configuring OAuth over PLAIN on the listener, which
results in username to always be treated as an account id, and password to always be treated as a raw access token without any prefix. In this mode the client can't authenticate using client credentials (Client ID + secret) - even if the client sends them, the server will always interpret it as account id, and an access token, resulting in authentication failing due to invalid token.
The initial implementation of OAuth over PLAIN was based on an invalid assumptions about how threads are assigned to handle messages from different connections in Kafka. An internal assertion often got triggered during concurrent usage, causing authentication failures.
A breaking change with how $accessToken is used in 0.7.0 had to be introduced.
Before, you could authenticate with an access token by setting username to $accessToken and setting the password to the access token string.
Now, you have to set username to be the same as the principal name that the broker will extract from the given access token (for example, the value of the claim configured by oauth.username.claim) and set the password to $accessToken: followed by the access token string.
So, compared to before, you now prefix the access token string with $accessToken: in the password parameter, and you have to be careful to properly set the username parameter to the principal name matching that in the access token.
SASL/PLAIN can now be used to perform the authentication using a service account clientId and secret or a long-lived access token.
When configuring OAuth authentication you should configure the custom principal builder factory:
principal.builder.class=io.strimzi.kafka.oauth.server.OAuthKafkaPrinipalBuilder
That is needed by OAuth over PLAIN to function correctly, and is also required by KeycloakRBACAuthorizer to function correctly, so it is best to just always configure it.
See [README.md] for instructions on how to set up the brokers and the clients.
Additional server-side configuration option was added to enable / disable the audience checking during authentication:
oauth.check.audience(e.g. "true")
See [README.md] for more information.
Another token validation mechanism was added which allows using the JSONPath filter queries to express the additional conditions that the token has to match during authentication in order to pass validation.
To enable it set the following option to a valid JSONPath filter query:
oauth.custom.claim.check(e.g. "'kafka-user' in @.roles")
See [README.md] for more information.
Redundant multiple instances of validators were being instantiated resulting in service components, supposed to be singletons, to be instantiated multiple times. That was fixed through internal refactoring.
Improvements have been made to logging for easier tracking of the requests, and better error reporting.
Many improvements have been made to address problems with access tokens expiring, or becoming invalid.
To prevent active sessions operating beyond the access token lifetime, Kafka 2.2 and later has re-authentication support, which has to be explicitly enabled.
Use the following server.properties configuration to enable re-authentication, and force clients to re-authenticate within one hour:
connections.max.reauth.ms=3600000
If the access token expires before that, the re-authentication will be enforced within the access token lifetime. Any non-authentication operation after token expiry will cause the connection to be terminated by the broker.
Re-authentication should be enabled if you want to prevent authenticated sessions from continuing beyond access token expiry.
Also, without re-authentication enabled, the KeycloakRBACAuthorizer will now deny further access as soon as the access token expires.
It would in any case fail fairly quickly as it relies on a valid access token when refreshing the list of grants for the current session.
In previous versions the re-authentication support was broken due to the bug #60. That should now be fixed.
Re-authentication forces active sessions whose access tokens have expired to immediately become invalid.
However, if for some reason you don't want to, or can't use re-authentication, and you don't use KeycloakRBACAuthorizer, but still want to enforce session expiry, or if you simply want to better log if token expiry occurs, the new authorizer denies all actions after the access token of the session has expired.
The client will receive the org.apache.kafka.common.errors.AuthorizationException.
Usage of OAuthSessionAuthorizer is optional.
It 'wraps' itself around another authorizer, and delegates all calls after determining that the current session still contains a valid token.
This authorizer should not be used together with KeycloakRBACAuthorizer, since the latter already performs all the same checks.
Two configuration options have been added for use by this authorizer:
-
strimzi.authorizer.delegate.class.nameSpecifies the delegate authorizer class name to be used.
-
strimzi.authorizer.grant.when.no.delegateEnables this authorizer to work without the delegate.
In order to support the newly added OAuthSessionAuthorizer the JwtKafkaPrincipalBuilder had to be moved to oauth-server module, which called for a different package.
We took the opportunity to also give the class a better name. The old class still exists, but simply extends the new class.
To use the new class your server.properties configuration should contain:
principal.builder.class=io.strimzi.kafka.oauth.server.OAuthKafkaPrincipalBuilder
-
Added session expiry enforcement to KeycloakRBACAuthorizer. If the access token used during authentication expires all the authorizations will automatically be denied. Note, that if re-authentication is enabled, and properly functioning, the access token should be refreshed on the server before ever expiring.
-
The KeycloakRBACAuthorizer will now regularly refresh the list of grants for every active session, and thus allows any permissions changes made at authorization server (Keycloak / RH-SSO) to take effect on Kafka brokers.
For example, if the access token expires in 30 minutes, and grants are refreshed every minute, the revocation of grants will be detected within one minute, and immediately enforced.
Additional server.properties configuration options have been introduced:
-
strimzi.authorization.grants.refresh.period.secondsThe time between two grants refresh job runs. The default value is 60 seconds. If this value is set to 0 or less, refreshing of grants is turned off.
-
strimzi.authorization.grants.refresh.pool.sizeThe number of threads that can fetch grants in parallel. The default value is 5.
When using fast local signature validation using JWKS endpoint keys, if signing keys are suddenly revoked at the authorization server, it takes a while for the Kafka broker to be aware of this change. Until then, the Kafka broker keeps successfully authorizing new connections with access tokens signed using old keys. At the same time it rejects newly issued access tokens that are properly signed with the new signing keys which Kafka broker doesn't yet know about. In order to shorten this mismatch period as much as possible, the broker will now trigger JWKS keys refresh as soon as it detects a new signing key. And it will keep trying if not successful using a so called exponential back-off, where after the unsuccessful attempt it waits a second, then two, then 4, 8, 16, 32, ... It will not flood the server with requests, always pausing for a minimum time between two consecutive keys refresh attempts.
While there would still be a few invalid token errors, the clients, if coded correctly, to reinitialise the KafkaProducer / KafkaConsumer in order to force access token refresh, should quickly recover.
The following new configuration option has been introduced:
-
oauth.jwks.refresh.min.pause.secondsThe minimum pause between two consecutive reload attempts. This prevents flooding the authorization server. The default value is 1 second.
In some cases the client would not receive org.apache.kafka.common.errors.AuthenticationException when it should.
Error messages were also improved to give a more precise reason for the failure.
Some claims are no longer required in token or Introspection Endpoint response (iat).
Others can be configured to not be required:
issclaim is not required ifoauth.check.issueris set tofalsesubclaim is no longer required ifoauth.username.claimis configured since then it is no longer used to extract principal.
Additional options were added to improve interoperability with authorization servers.
The following options were added:
-
oauth.scopeScope can now be specified for the Token endpoint on the Kafka clients and on the Kafka broker for inter-broker communication.
-
oauth.check.issuerIssuer check can now be disabled when configuring token validation on the Kafka broker - some authorization servers don't provide
issclaim. -
oauth.fallback.username.claimPrincipal can now be extracted from JWT token or Introspection endpoint response by using multiple claims. First
oauth.username.claimis attempted (if configured). If the value is not present, the fallback claim is attempted. If neitheroauth.username.claimnoroauth.fallback.username.claimis specified or its value present,subclaim is used. -
oauth.fallback.username.prefixIf principal is set by
oauth.fallback.username.claimthen its value will be prefixed by the value ofoauth.fallback.username.prefix, if specified. -
oauth.userinfo.endpoint.uriSometimes the introspection endpoint doesn't provide any claim that could be used for the principal. In such a case User Info Endpoint can be used, and configuration of
oauth.username.claim,oauth.fallback.username.claim, andoauth.fallback.username.prefixis taken into account. -
oauth.valid.token.typeWhen using the Introspection Endpoint, some servers use custom values for
token_type. If this configuration parameter is set then thetoken_typeattribute has to be present in Introspection Token response, and has to have the specified value.
The job that refreshes the keys would be cancelled if fetching of keys failed due to network error or authorization server glitch.
If token_type was present it was expected to be equal to access_token which is not an OAuth 2.0 spec compliant value.
Token type check is now disabled unless the newly introduced oauth.valid.token.type configuration option is set.
-
Fixed an issue with
keycloakandhydracontainers not visible when starting services in separate shells.The instructions for running
keycloak/hydraseparately omitted the required-f compose.ymlas a first compose file, resulting in a separate bridge network being used. -
Added Spring Security Authorization Server
There is now some TRACE logging support which should only ever be used in development / testing environment because it outputs secrets into the log.
When integrating with your authorization server, enabling TRACE logging on io.strimzi.kafka.oauth logger will output the authorization server responses which can point you to how to correctly configure oauth.* parameters to make the integration work.
The helper library used for JWT / JWKS handling was bumped to version 10.0.0
The following configuration options have been deprecated:
oauth.tokens.not.jwtis now calledoauth.access.token.is.jwtand has a reverse meaning.oauth.validation.skip.type.checkis now calledoauth.check.access.token.typeand has a reverse meaning.
See: Align configuration with Kafka Operator PR (#36).
Scope claim is no longer required in an access token. (#30) That improves compatibility with different authorization servers, since the attribute is not required by OAuth 2.0 specification neither is it used by validation logic.
jackson-core, and jackson-databind libraries have been updated to latest versions. (#33)
Instructions for preparing the environment, building and deploying the latest version of Strimzi Kafka OAuth library with Strimzi Kafka Operator have been added.
See: Hacking on OAuth and deploying with Strimzi Kafka Operator PR (#34)
Fixed enabled remote debugging mode in example compose-authz.yml (#39)
It is now possible to use Keycloak Authorization Services to centrally manage access control to resources on Kafka Brokers (#24) See the tutorial which explains many concepts. For configuration details also see KeycloakRBACAuthorizer JavaDoc.
The JWTSignatureValidator now supports ECDSA signatures, but requires explicit enablement of BouncyCastle security provider (#25)
To enable BouncyCastle set oauth.crypto.provider.bouncycastle to true.
Optionally you may control the order where the provider is installed by using oauth.crypto.provider.bouncycastle.position - by default it is installed at the end of the list of existing providers.
A testsuite based on Arquillian Cube, and using docker containers was added.
Added Ory Hydra authorization server to examples.
Support for token-based authentication that plugs into Kafka's SASL/OAUTHBEARER mechanism to provide:
- Different ways of access token retrieval for Kafka clients (clientId + secret, refresh token, or direct access token)
- Fast signature-checking token validation mechanism (using authorization server's JWKS endpoint)
- Introspection based token validation mechanism (using authorization server's introspection endpoint)
See the tutorial.