Status

Current state: "Under Discussion"Discarded, replaced by KIP-519: Make SSL context/engine configuration extensible

Discussion thread: here

JIRA: here

...

Of course it would be easy to simply change the Kafka source code and ship a customized distribution. The idea is you should be able to replace the implementation through configuration, without rebuilding a custom Kafka distribution or resorting to classpath tricks to shadow Kafka classes.

SslFactory handles two different tasks: creating the SSLContext and reconfiguration. There is no need to make reconfiguration pluggable. This functionality should remain in SslFactory to be shared by all pluggable implementations. SslFactory only needs to delegate the creation of the SSLContext to the pluggable interface.

Since SslFactory will handle reconfiguration, the idea is to make the configuration immutable in the pluggable factory. SslFactory would create a new pluggable factory every time the configuration changes. The pluggable factory creates its SSLContext when it is configured and never changes it. It turns out SslFactory does not really need the SSLContext, so it can use the new pluggable factory as an SSLEngine factory instead.

Public Interfaces

This KIP will introduce a new configuration option ssl.sslfactorysslengine.factory.class to be added to SslConfigs. This will automatically add the new option to AdminClientConfig, ConsumerConfig and ProducerConfig.

A new public interface named SslFactory SslEngineFactory will be created. A default implementation named DefaultSslFactory DefaultSslEngineFactory with the existing behavior will be included with Kafka. The implementation is private, but the class name will leak as the default value of ssl.sslfactorysslengine.factory.class

An application developer can implement SslFactory SslEngineFactory and set ssl.sslengine.sslfactoryfactory.class to the fully qualified class name to instruct Kafka to use this implementation instead of the default implementation.

...

A new configuration option will be added to SslConfigs:

  public static final String SSL_SSLENGINE_SSLFACTORYFACTORY_CLASS_CONFIG = "ssl.sslengine.sslfactoryfactory.class";
  public static final String SSL_SSLFACTORYSSLENGINE_FACTORY_CLASS_DOC = "...";
  public static final String DEFAULT_SSL_SSLENGINE_SSLFACTORYFACTORY_CLASS = "org.apache.kafka.common.security.ssl.DefaultSslFactoryDefaultSslEngineFactory";

  config.define(SslConfigs.SSL_SSLENGINE_SSLFACTORYFACTORY_CLASS_CONFIG, ConfigDef.Type.CLASS, SslConfigs.DEFAULT_SSL_SSLFACTORYSSLENGINE_FACTORY_CLASS, ConfigDef.Importance.LOW, SslConfigs.SSL_SSLFACTORYSSLENGINE_FACTORY_CLASS_DOC)

...

A new public interface named SslFactory SslEngineFactory will be created.
public  public interface SslFactorySslEngineFactory extends Reconfigurable {
      /* valid values are Mode.CLIENT or Mode.SERVER */
   public static final String SSL_MODE_CONFIG = "ssl.mode";
   /* valid values are Boolean.TRUE or Boolean.FALSE */
   public static final String SSL_VERIFY_KEYSTORE_USING_TRUSTSTORE_CONFIG = "ssl.verify.keystore.using.truststore";
void configure(Map<String, ?> configs, Mode mode);
         public SSLEngine createSslEngine(String peerHost, int peerPort);
  }
The source code of SslFactory will contain javadoc comments to describe the interface and each declaration.
An implementation of SslFactory will be instantiated with a default constructor. The non-default constructors in DefaultSslFactory will be removed. The member clientAuthConfigOverride will also be removed. The default constructor of DefaultSslFactory does nothing and therefore can be omitted too.
DefaultSslFactory can no longer rely on constructor arguments. Instead it will receive these values as regular config options in configure((Map<String, ?> configs). This KIP introduces two new config keys ssl.mode and ssl.verify.keystore.using.truststore. The configure() method in DefaultSslFactory will be modified to initialize the mode from the ssl.mode config and the boolean keystoreVerifiableUsingTruststore from the ssl.verify.keystore.using.truststore config. The variable clientAuthConfig will always be initialized from the ssl.client.auth config. It is the caller's responsibility to override ssl.client.auth before calling configure() because it can no longer be overridden by a constructor argument.
In SaslChannelBuilder and SslChannelBuilder, the type of the member sslFactory will still be SslFactory but now it refers to the interface. Instead of calling the SslFactory constructor directly, we will call a new method createSslFactory(). This method will be duplicated in both channel builders.
createSslFactory() will alter (a copy of?) the configs to contain the values for what used to be passed as arguments in the SslFactory constructor. The SSL mode will be stored with the key ssl.mode; if not null clientAuthConfigOverride will be stored with the key ssl.client.auth overwriting the value already there; keystoreVerifiableUsingTruststore will be stored with the key ssl.verify.keystore.using.truststore.
createSslFactory() will call the default constructor of the pluggable SSL Interface by calling Class.newInstance() followed by a call to configure() to pass the configs.

SslFactory finds the class name of the SslEngineFactory implementation in the ssl.sslengine.factory.class config and instantiates it using the default constructor by calling Class.newInstance(). The configs and mode are then passed to configure. This creates the SSLContext which remains private to the SslEngineFactory implementation. SslFactory can now call createSslEngine(peerHost, peerPort) as often as needed.
This assumes SslFactory copies the configs and overwrites ssl.client.auth when there is a clientAuthConfigOverride.
When reconfiguring, only the keystore and truststore configs are allowed to change. This is the existing behavior and this KIP simply preserves it. SslFactory must keep a copy of the previous configs to know the original values for the other configs. This copy also guarantees we don't overwrite anything in the caller's configs.
SslFactory must inspect the keystore and truststore to detect a future change in configuration. In particular, it needs to keep the lastModified times, plus the CertificateEntries of the keystore. The configuration might change during the small interval between the inspection and loading the files in the SslEngineFactory implementation. By always inspecting the files before calling configure, we make sure the SslEngineFactory implementation will always have the same or newer configuration and reconfiguration will never miss an update later. 
It is allowed to keep ssl.keystore.location and/or ssl.truststore.location empty. This means the discovery of the keystore and truststore is fully delegated to the SslEngineFactory implementation. This turns off reconfiguration when both the old and new configuration delegates to the SslEngineFactory implementation.
SslFactory.sslContext() will be removed because the SslContext is now private to the SslEngineFactory implementation.
Compatibility, Deprecation, and Migration Plan
This KIP is considered backwards compatible. SslFactory is not part of the Kafka public API. The SslFactory API remains the same anyway except for the removal of sslContext().
EchoServer is the only caller of SslFactory.sslContext(), This EchoServer calls SslFactory directly and will no longer compile. This is an internal test so backwards compatibility does not apply. We will update EchoServer to call DefaultSslFactory directly instead.
Compatibility, Deprecation, and Migration Plan
Because SslFactory was not a public API, this KIP is considered backwards compatible. Applications that called SslFactory directly will have to be updated but that's the risk taken when calling private APIs.
The default value for the new config ssl.sslengine.sslfactoryfactory.class is org.apache.kafka.common.security.ssl.DefaultSslFactory DefaultSslEngineFactory which selects the existing implementation.
The SslFactory implementation is instantiated by the SslChannelBuilder or the SaslChannelBuilder which will hide all the changes.
EchoServer calls SslFactory directly, but that's an internal test. We will update EchoServer as part of this KIPimplements the existing behavior.
Rejected Alternatives
Kafka could define a new configuration option to hold an instance of SSLSocketFactory. This is similar to many Java libraries that accept an instance of SSLSocketFactory. This was rejected because Kafka tries to be language neutral. It was thought it would make it more difficult to support C and Python.
Kafka's naming convention does not use a special tag for interfaces. Accordingly, these interface names were rejected ISslFactoryISslEngineFactory, SslFactoryIFaceSslEngineFactoryIFace, SslFactoryInterfaceSslEngineFactoryInterface.
It was possible to preserve the name SslFactory for the default implementation and use the name PluggableSslFactory for the interface. Since SslFactory was not part of the public API, it was decided we could rename the class and reuse the nicer name for the interface.
The constants SSL_MODE_CONFIG and SSL_VERIFY_KEYSTORE_USING_TRUSTSTORE_CONFIG are not user configurable and therefore must not appear in SslConfigs as ConfigDefs. Declaring the constants in SslConfigs without a ConfigDef would just cause confusion.
The only call to SslFactory.sslContext() is in EchoServer which does not count because it is an internal test. We decided to leave the sslContext() method out of the interfaceWe could expose the SSLContext with an accessor in the SslEngineFactory interface, but it was only used by an internal test, so we decided to leave it out.
Kafka is not consistent to name configs that hold class names. Compare [partitioner.class, interceptor.classes, principal.builder.class, sasl.server.callback.handler.class, sasl.client.callback.handler.class, sasl.login.class] versus [key.deserializer, value.deserializer, key.serializer, value.serializer]. It appears the serializer/deserializer configs are a special case. Therefore the name ssl.sslfactorysslengine.factory.class was selected instead of ssl.sslengine.sslfactoryfactory
We could require the Pluggable SSL Factory implementation to have the same 3-argument constructor as the old SslFactory. We find the arguments clientAuthConfigOverride and keystoreVerifiableUsingTruststore are somewhat arbitrary for a general pluggable interface. We prefer to use a default constructor for the pluggable SSL Factory implementationadd clientAuthConfigOverride as a third argument for SslEngineFactory.configure(). This looked like an internal feature of Kafka that leaked into the more general interface. We thought overwriting ssl.client.auth in a copy of the configs was cleaner.
We could pass the keystore and truststore as arguments to the SslEngineFactory since SslFactory already loaded them. This felt awkward since the keystore and truststore configs would still be present in the configs passed to configure(). It also felt clumsy when the SslEngineFactory uses its own keystore and truststore, for example, when reusing an SSLContext from the rest of the application.
We need the ability to send custom configuration options to the PluggableSslFactory implementation. This will could be covered in another KIP.