Writing oracle services¶

The two basic approaches¶

The architecture provides two ways of implementing oracles with different tradeoffs:

1. Using commands
2. Using attachments

When a fact is encoded in a command, it is embedded in the transaction itself. The oracle then acts as a co-signer to the entire transaction. The oracle’s signature is valid only for that transaction, and thus even if a fact (like a stock price) does not change, every transaction that incorporates that fact must go back to the oracle for signing.

When a fact is encoded as an attachment, it is a separate object to the transaction and is referred to by hash. Nodes download attachments from peers at the same time as they download transactions, unless of course the node has already seen that attachment, in which case it won’t fetch it again. Contracts have access to the contents of attachments when they run.

As you can see, both approaches share a few things: they both allow arbitrary binary data to be provided to transactions (and thus contracts). The primary difference is whether the data is a freely reusable, standalone object or whether it’s integrated with a transaction.

Here’s a quick way to decide which approach makes more sense for your data source:

• Is your data continuously changing, like a stock price, the current time, etc? If yes, use a command.
• Is your data commercially valuable, like a feed which you are not allowed to resell unless it’s incorporated into a business deal? If yes, use a command, so you can charge money for signing the same fact in each unique business context.
• Is your data very small, like a single number? If yes, use a command.
• Is your data large, static and commercially worthless, for instance, a holiday calendar? If yes, use an attachment.
• Is your data intended for human consumption, like a PDF of legal prose, or an Excel spreadsheet? If yes, use an attachment.

Asserting continuously varying data¶

Let’s look at the interest rates oracle that can be found in the NodeInterestRates file. This is an example of an oracle that uses a command because the current interest rate fix is a constantly changing fact.

The obvious way to implement such a service is like this:

1. The creator of the transaction that depends on the interest rate sends it to the oracle.
2. The oracle inserts a command with the rate and signs the transaction.
3. The oracle sends it back.

But this has a problem - it would mean that the oracle has to be the first entity to sign the transaction, which might impose ordering constraints we don’t want to deal with (being able to get all parties to sign in parallel is a very nice thing). So the way we actually implement it is like this:

1. The creator of the transaction that depends on the interest rate asks for the current rate. They can abort at this point if they want to.
2. They insert a command with that rate and the time it was obtained into the transaction.
3. They then send it to the oracle for signing, along with everyone else, potentially in parallel. The oracle checks that the command has the correct data for the asserted time, and signs if so.

This same technique can be adapted to other types of oracle.

The oracle consists of a core class that implements the query/sign operations (for easy unit testing), and then a separate class that binds it to the network layer.

Here is an extract from the NodeInterestRates.Oracle class and supporting types:

/** A [FixOf] identifies the question side of a fix: what day, tenor and type of fix ("LIBOR", "EURIBOR" etc) */
data class FixOf(val name: String, val forDay: LocalDate, val ofTenor: Duration)

/** A [Fix] represents a named interest rate, on a given day, for a given duration. It can be embedded in a tx. */
data class Fix(val of: FixOf, val value: BigDecimal) : CommandData

class Oracle {
    fun query(queries: List<FixOf>, deadline: Instant): List<Fix>

    fun sign(ftx: FilteredTransaction, merkleRoot: SecureHash): DigitalSignature.LegallyIdentifiable
}

Because the fix contains a timestamp (the forDay field), that identifies the version of the data being requested, there can be an arbitrary delay between a fix being requested via query and the signature being requested via sign as the Oracle can know which, potentially historical, value it is being asked to sign for. This is an important technique for continously varying data.

The query method takes a deadline, which is a point in time the requester is willing to wait until for the necessary data to be available. Not every oracle will need this. This can be useful where data is expected to be available on a particular schedule and we use scheduling functionality to automatically launch the processing associated with it. We can schedule for the expected announcement (or publish) time and give a suitable deadline at which the lack of the information being available and the delay to processing becomes significant and may need to be escalated.

Hiding transaction data from the oracle¶

Because the transaction is sent to the oracle for signing, ordinarily the oracle would be able to see the entire contents of that transaction including the inputs, output contract states and all the commands, not just the one (in this case) relevant command. This is an obvious privacy leak for the other participants. We currently solve this with FilteredTransaction-s and the use of Merkle Trees. These reveal only the necessary parts of the transaction to the oracle but still allow it to sign it by providing the Merkle hashes for the remaining parts. See Transaction tear-offs for more details.

Pay-per-play oracles¶

Because the signature covers the transaction, and transactions may end up being forwarded anywhere, the fact itself is independently checkable. However, this approach can still be useful when the data itself costs money, because the act of issuing the signature in the first place can be charged for (e.g. by requiring the submission of a fresh Cash.State that has been re-assigned to a key owned by the oracle service). Because the signature covers the transaction and not only the fact, this allows for a kind of weak pseudo-DRM over data feeds. Whilst a smart contract could in theory include a transaction parsing and signature checking library, writing a contract in this way would be conclusive evidence of intent to disobey the rules of the service (res ipsa loquitur). In an environment where parties are legally identifiable, usage of such a contract would by itself be sufficient to trigger some sort of punishment.

Implement the core classes¶

The key is to implement your oracle in a similar way to the NodeInterestRates.Oracle outline we gave above with both query and sign methods. Typically you would want one class that encapsulates the parameters to the query method (FixOf above), and a CommandData implementation (Fix above) that encapsulates both an instance of that parameter class and an instance of whatever the result of the query is (BigDecimal above).

The NodeInterestRates.Oracle allows querying for multiple Fix-es but that is not necessary and is provided for the convenience of callers who might need multiple and can do it all in one query request. Likewise the deadline functionality is optional and can be avoided initially.

Let’s see what parameters we pass to the constructor of this oracle.

class Oracle(val identity: Party, private val signingKey: KeyPair, val clock: Clock) = TODO()

Here we see the oracle needs to have its own identity, so it can check which transaction commands it is expected to sign for, and also needs a pair of signing keys with which it signs transactions. The clock is used for the deadline functionality which we will not discuss further here.

Assuming you have a data source and can query it, it should be very easy to implement your query method and the parameter and CommandData classes.

Let’s see how the sign method for NodeInterestRates.Oracle is written:

        fun sign(ftx: FilteredTransaction, merkleRoot: SecureHash): DigitalSignature.LegallyIdentifiable {
            if (!ftx.verify(merkleRoot)) {
                throw MerkleTreeException("Rate Fix Oracle: Couldn't verify partial Merkle tree.")
            }

            // Reject if we have something different than only commands.
            val leaves = ftx.filteredLeaves
            require(leaves.inputs.isEmpty() && leaves.outputs.isEmpty() && leaves.attachments.isEmpty())

            val fixes: List<Fix> = ftx.filteredLeaves.commands.
                    filter { identity.owningKey in it.signers && it.value is Fix }.
                    map { it.value as Fix }

            // Reject signing attempt if we received more commands than we should.
            if (fixes.size != ftx.filteredLeaves.commands.size)
                throw IllegalArgumentException()

            // Reject this signing attempt if there are no commands of the right kind.
            if (fixes.isEmpty())
                throw IllegalArgumentException()

            // For each fix, verify that the data is correct.
            val knownFixes = knownFixes // Snapshot
            for (fix in fixes) {
                val known = knownFixes[fix.of]
                if (known == null || known != fix)
                    throw UnknownFix(fix.of)
            }

            // It all checks out, so we can return a signature.
            //
            // Note that we will happily sign an invalid transaction, as we are only being presented with a filtered
            // version so we can't resolve or check it ourselves. However, that doesn't matter much, as if we sign
            // an invalid transaction the signature is worthless.
            return signingKey.signWithECDSA(merkleRoot.bytes, identity)
        }

Here we can see that there are several steps:

1. Ensure that the transaction we have been sent is indeed valid and passes verification, even though we cannot see all of it.
2. Check that we only received commands as expected, and each of those commands expects us to sign for them and is of the expected type (Fix here).
3. Iterate over each of the commands we identified in the last step and check that the data they represent matches exactly our data source. The final step, assuming we have got this far, is to generate a signature for the transaction and return it.

Binding to the network via a CorDapp plugin¶

The first step is to create a service to host the oracle on the network. Let’s see how that’s implemented:

    class Service(val services: PluginServiceHub) : AcceptsFileUpload, SingletonSerializeAsToken() {
        val oracle: Oracle by lazy {
            val myNodeInfo = services.myInfo
            val myIdentity = myNodeInfo.serviceIdentities(type).first()
            val mySigningKey = services.keyManagementService.toKeyPair(myIdentity.owningKey.keys)
            Oracle(myIdentity, mySigningKey, services.clock)
        }

        init {
            // Note: access to the singleton oracle property is via the registered SingletonSerializeAsToken Service.
            // Otherwise the Kryo serialisation of the call stack in the Quasar Fiber extends to include
            // the framework Oracle and the flow will crash.
            services.registerFlowInitiator(RatesFixFlow.FixSignFlow::class) { FixSignHandler(it, this) }
            services.registerFlowInitiator(RatesFixFlow.FixQueryFlow::class) { FixQueryHandler(it, this) }
        }

        private class FixSignHandler(val otherParty: Party, val service: Service) : FlowLogic<Unit>() {
            @Suspendable
            override fun call() {
                val request = receive<RatesFixFlow.SignRequest>(otherParty).unwrap { it }
                send(otherParty, service.oracle.sign(request.ftx, request.rootHash))
            }
        }

        private class FixQueryHandler(val otherParty: Party, val service: Service) : FlowLogic<Unit>() {
            companion object {
                object RECEIVED : ProgressTracker.Step("Received fix request")
                object SENDING : ProgressTracker.Step("Sending fix response")
            }

            override val progressTracker = ProgressTracker(RECEIVED, SENDING)

            init {
                progressTracker.currentStep = RECEIVED
            }

            @Suspendable
            override fun call(): Unit {
                val request = receive<RatesFixFlow.QueryRequest>(otherParty).unwrap { it }
                val answers = service.oracle.query(request.queries, request.deadline)
                progressTracker.currentStep = SENDING
                send(otherParty, answers)
            }
        }

This may look complicated, but really it’s made up of some relatively simple elements (in the order they appear in the code):

1. Accept a PluginServiceHub in the constructor. This is your interface to the Corda node.
2. Ensure you extend the abstract class SingletonSerializeAsToken (see The Corda plugin framework).
3. Create an instance of your core oracle class that has the query and sign methods as discussed above.
4. Register your client sub-flows (in this case both in RatesFixFlow. See the next section) for querying and signing as initiating your service flows that actually do the querying and signing using your core oracle class instance.
5. Implement your service flows that call your core oracle class instance.

The final step is to register your service with the node via the plugin mechanism. Do this by implementing a plugin. Don’t forget the resources file to register it with the ServiceLoader framework (see The Corda plugin framework).

class Plugin : CordaPluginRegistry() {
     override val servicePlugins: List<Class<*>> = listOf(Service::class.java)
}

Providing client sub-flows for querying and signing¶

We mentioned the client sub-flow briefly above. They are the mechanism that clients, in the form of other flows, will interact with your oracle. Typically there will be one for querying and one for signing. Let’s take a look at those for NodeInterestRates.Oracle.

    class FixQueryFlow(val fixOf: FixOf, val oracle: Party) : FlowLogic<Fix>() {
        @Suspendable
        override fun call(): Fix {
            val deadline = suggestInterestRateAnnouncementTimeWindow(fixOf.name, oracle.name, fixOf.forDay).end
            // TODO: add deadline to receive
            val resp = sendAndReceive<ArrayList<Fix>>(oracle, QueryRequest(listOf(fixOf), deadline))

            return resp.unwrap {
                val fix = it.first()
                // Check the returned fix is for what we asked for.
                check(fix.of == fixOf)
                fix
            }
        }
    }

    class FixSignFlow(val tx: TransactionBuilder, val oracle: Party, val filterFuns: FilterFuns) : FlowLogic<DigitalSignature.LegallyIdentifiable>() {
        @Suspendable
        override fun call(): DigitalSignature.LegallyIdentifiable {
            val wtx = tx.toWireTransaction()
            val partialMerkleTx = FilteredTransaction.buildMerkleTransaction(wtx, filterFuns)
            val rootHash = wtx.id

            val resp = sendAndReceive<DigitalSignature.LegallyIdentifiable>(oracle, SignRequest(rootHash, partialMerkleTx))
            return resp.unwrap { sig ->
                check(sig.signer == oracle)
                tx.checkSignature(sig)
                sig
            }
        }
    }

You’ll note that the FixSignFlow requires a FilterFuns instance with the appropriate filter to include only the Fix commands. You can find a further explanation of this in Transaction tear-offs.