The node now sends a transaction to the verifier if any of its attachments were compiled with Kotlin 1.2 (the net.corda.node.verification.external system property has been removed). It uses kotlinx-metadata to read the Kotlin metadata in the attachment to determine this. For now this scanning is done each time the attachment is loaded from the database.
The existing external verification integration tests were converted into smoke tests so that 4.11 nodes could be involved. This required various improvements to NodeProcess.Factory. A new JAVA_8_HOME environment variable, pointing to JDK 8, is required to run these tests.
There is still some follow-up work that needs to be done:
Sending transactions from a 4.11 node to a 4.12 node works, but not the other way round. A new WireTransaction component group needs to be introduced for storing 4.12 attachments so that they can be safely ignored by 4.11 nodes, and the 4.12 node needs to be able to load both 4.11 and 4.12 versions of the same contracts CorDapp so that they can be both attached to the transaction.
Even though attachments are cached when retrieved from the database, the Kotlin metadata version should be stored in the attachments db table, rather than being scanned each time.
Finally, VerificationService was refactored into NodeVerificationSupport and can be passed into SignedTransaction.verifyInternal, instead of needing the much heavier VerifyingServiceHub. This makes it easier for internal tools to verify transactions and spawn the verifier if necessary.
* ENT-11055: Basic external verification
Introduction of the external transaction verifier, a separate JVM process for verifying `SignedTransaction`s. The end goal is for this verifier to be built with Kotlin 1.2 so that it creates a compatible verification environment for transactions with 4.11 contracts. For now however the verifier is built against Kotlin 1.8, same as the node.
External verification is enabled when the the system property `net.corda.node.verification.external` is set to `true`. When enabled, all verification requests made via `SignedTransaction.verify` are sent to the external verifier, regardless of the transaction content. It will do the vast bulk of the verification and then send the result back, namely if an exception occurred. If it did, then it's re-thrown in the node.
The external verifier is a stateless process, with no connection to the node's database. All transaction resolution information needed to create the relevant ledger transaction object are made to the node, which waits in a loop servicing these requests until it receives the result. The verifier Jar is embedded in the Corda node Jar, and is extracted and run when needed for the first time. The node opens up a local port for the verifier to communicate with, which is specified to the verifier in the process command line. This all means there is no extra configuration or deployment required to support external verification.
The existing code had some initial attempts and abstractions to support a future external verification feature. However,
they were either incorrect or didn't quite fit. One such example was `TransactionVerifierService`. It incorrectly operated on the `LedgerTransaction` level, which doesn't work since the transaction needs to be first serialised. Instead a new abstraction, `VerificationSupport` has been introduced, which represents all the operations needed to resolve and verify a `SignedTransaction`, essentially replacing `ServicesForResolution` (a lot of the changes are due to this). The external verifier implements this with a simple RPC mechanism, whilst the node needed a new (internal) `ServiceHub` abstraction, `VerifyingServiceHub`. `ServicesForResolution` hasn't been deleted since it's public API, however all classes implementing it must also implement `VerifyingServiceHub`. This is possible to do without breaking compatibility since `ServicesForResolution` is annotated with `@DoNotImplement`.
Changes to `api-current.txt` were made due to the removal of `TransactionVerifierService`, which was clearly indicated as an internal class, and returning `TransactionBuilder.toLedgerTransactionWithContext` back to an internal method.
* Address review comments
* One bulk load states method
* Merge fix
Replaced usage of `@Test.expected` annotation parameter with more specific exception assertions. This is also needed to migrate away from the explicit timeouts in every tests.
Major changes due to JDK 17:
1. JDK17 JCE Provider now has built-in support for eddsas, corda uses
the bouncycastle (i2p) implementation. This PR removes the conflicting
algorithms from the built-in JCE provider.
2. JavaScript scripting has been removed from the JDK, the corda log4j config was using
scripting to conditionally output additional diagnostic info if the MDC
was populated. This PR has removed the scripting.
3. The artifactory plug-ins used are now deprecated, this PR has removed them
and uses the same code as Corda 5 for publishing to artifactory.
4. Javadoc generation has been modified to use the latest dokka plug-ins.
5. Gradle 7.6 has implemented an incredibly annoying change where transitive
dependencies are not put on the compile classpath, so that they have to be
explicitly added as dependencies to projects.
6. Mockito has been updated, which sadly meant that quite a few source files
have to changes to use the new (org.mockito.kotlin) package name. This makes
this PR appear much larger than it is.
7. A number of tests have been marked as ignored to get a green, broadly they fall
into 3 classes.
The first is related to crypto keypair tests, it appears some logic
in the JDK prefers to use the SunJCE implementation and we prefer to use
bouncycastle. I believe this issue can be fixed with better test setup.
The second group is related to our use of a method called "uncheckedCast(..)",
the purpose of this method was to get rid of the annoying unchecked cast compiler
warning that would otherwise exist. It looks like the Kotlin 1.9 compiler type
inference differs and at runtime sometimes the type it infers is "Void" which causes
an exception at runtime. The simplest solution is to use an explicit cast instead of
unchecked cast, Corda 5 have removed unchecked cast from their codebase.
The third class are a number of ActiveMQ tests which appear to have a memory leak somewhere.
