Introduction to TigerBeetle Transactions Database

1. Introduction

In this tutorial, we’ll explore the TigerBeetle database engine and learn how we can use it to build a fault-tolerant and high-performance application.

2. Financial Transactions in a Nutshell

Every time we use a debit or credit to buy something online or in a store, there will be a transaction where some currency amount will be transferred from your account to the merchant’s.

Behind the scenes, fees must be deducted from the value transferred and then split among all parties involved (acquirer, card processing company, banks, etc). All those entities must also keep detailed transaction logs, also known as ledgers, after the books where accountants used to keep them.

Nowadays, most financial transaction systems rely on database engines, such as Oracle, SQL Server, and DB2, to store transactions. A typical system will have an accounts table that holds balances and a transactions table that logs every debt or credit made to the accounts.

While this works well, the general purpose of these databases leads to several inefficiencies and, consequently, requires much more resources to deploy and operate at large scales.

3. TigerBeetle’s Approach

A newcomer in the crowded database market, TigerBeetle is a specialized database engine that focuses primarily on financial transactions. By doing so, it can get rid of most of the complexity associated with a general-purpose database engine and, in exchange, claims to be able to deliver up to 1000x throughput improvement.

These are the main simplifications that make this improvement possible:

• Fixed schema
• Fixed-point arithmetic
• Batched transactions
• No general query capabilities

Perhaps the most surprising of these is the fixed schema. TigerBeetle has just two entities: Accounts and Transfers.

An Account stores the balance of some asset (current, stocks, bitcoins, etc.), which can be anything that we can acquire or transfer to/from another Account belonging to the same ledger. An Account also has a few fields that are intended to hold external identifiers, allowing us to link it to traditional system-of-records databases.

To add some credit to an Account we create a Transfer instance that contains the amount, the source Account from which we’ll deduce this amount, and the destination Account.

Those are some important features of Accounts and Transfers:

•  Once created, an Account cannot be deleted
•  The Account‘s initial balance is always zero
•  Transfers are immutable. Once committed they cannot be modified or deleted
•  Transfers require both Accounts to be on the same ledger
• At all times, the sum of debits and credits over all Accounts is zero

4. Deploying TigerBeetle

TigerBeetle is distributed as a statically linked executable, available at the official site.

Before using TigerBeetle, we need to create a data file where it will store its data. This is done using the format command:

$ tigerbeetle format --cluster=0 --replica=0 --replica-count=1 0_0.tigerbeetle

We can now start a standalone instance using the start command:

$ tigerbeetle start --addresses=3000 0_0.tigerbeetle

5. Using TigerBeetle from Java Applications

This is the TigerBeetle’s official client Maven dependency:

<dependency>
    <groupId>com.tigerbeetle</groupId>
    <artifactId>tigerbeetle-java</artifactId>
    <version>0.15.3</version>
</dependency>

The latest version of this library is available from Maven Central.

Important notice: This dependency contains platform-specific native code. Make sure to run your client only on supported architectures

5.1. Connecting to TigerBeetle

The entry point to access TigerBeetle’s functionalities is the Client class. Client instances are thread-safe, so we just need to create a single instance of it in our applications. For Spring-based applications, the simplest approach is to define a @Bean in a @Configuration class, so we can inject it when needed:

@Configuration
public class TigerBeetleConfig {
    @Value("${tigerbeetle.clusterID:0}")
    private BigInteger clusterID;

    @Value("${tb_address:3000}")
    private String[] replicaAddress;

    @Bean
    Client tigerBeetleClient() {
        return new Client(UInt128.asBytes(clusterID), replicaAddress);
    }
}

5.2. Creating Accounts

TigerBeetle’s API doesn’t come with any domain object, so let’s create a simple Account record to store the data we need to create one:

@Builder
public record Account(
    UUID id,
    BigInteger accountHolderId,
    int code,
    int ledger,
    int userData32,
    long userData64,
    BigInteger creditsPosted,
    BigInteger creditsPending,
    BigInteger debtsPosted,
    BigInteger debtsPending,
    int flags,
    long timestamp) {
}

Here, we’ve used a UUID account identifier as a convenience. Internally, TigerBeetle uses 128-bit integers as account identifiers, which the Java API maps to 16-byte arrays. Our domain class has an accountHolderId, which maps to the userData128 field.

The API uses 128-bit integers in many places, but since Java has no equivalent native datatype, it provides the UInt128 utility class that helps to convert from arrays to other formats. Besides UUIDs, we can also use BigIntegers or a pair of regular long integers.

The userData128, userData32, and userData64 fields’ main use is to store secondary identifiers associated with this account. For example, we can use them to store the identifier of this account in an external database.

Now, let’s create an AccountRepository and implement a createAccount() method:

@RequiredArgsConstructor
public class AccountRepository {
    private final Client client;

    public Account createAccount(BigInteger accountHolderId, int code, int ledger,
      int userData32, long userData64, int flags ) {
        AccountBatch batch = new AccountBatch(1);
        byte[] id = UInt128.id(); 
        batch.add();
        batch.setId(id);
        batch.setUserData128(UInt128.asBytes(accountHolderId));
        batch.setLedger(ledger);
        batch.setCode(code);
        batch.setFlags(AccountFlags.HISTORY | flags);

        CreateAccountResultBatch result = client.createAccounts(batch);
        if(result.getLength() > 0) {
            result.next();
            throw new AccountException(result.getResult());
        }

        return findAccountById(UInt128.asUUID(id)).orElseThrow();
    }

   // ... other repository methods omitted
}

The implementation starts by creating an AccountBatch object to hold the batch data. In this example, the batch consists of a single account creation command, but we could easily extend this model to accept multiple requests.

Notice the AccountFlags.HISTORY flag. When set, we’ll be able to query historical balances, as we’ll see later. Also important is the use of UInt128.id() to generate the Account identifier. Values returned from this method are unique and time-based, meaning that if we compare them, we can determine which one was created first.

Once we’ve populated the batch request, we send it to TigerBeetle using the createAccounts() method. This method returns a CreateAccountResultBatch, which will be empty in case of a successful request. Otherwise, there will be an entry for each failed creation request containing the reason for the failure.

As a convenience to the caller, the method returns the Account domain object populated from data recovered from the database, which includes the actual creation timestamp as set by TigerBeetle.

5.3. Looking Up Accounts

To implement findAccountById() we follow a similar pattern as in the previous case. Firstly, we create a batch to hold the identifiers we want to find. To make things simpler, we’ll limit ourselves to just a single account per call.

Next, we submit this batch to TigerBeetle and process the results:

public Optional<Account> findAccountById(UUID id) throws ConcurrencyExceededException {
    IdBatch idBatch = new IdBatch(UInt128.asBytes(id));
    var batch = client.lookupAccounts(idBatch);

    if (!batch.next()) {
        return Optional.empty();
    }

    return Optional.of(mapFromCurrentAccountBatch(batch));
 }

Notice the use of next() to determine whether the given identifier exists or not. This works because of the single account limitation mentioned above.

A variant of this method that supports multiple identifiers is available in the source code. There, we populate the resulting Map using the returned values, leaving null entries for any identifier not found.

5.4. Creating Simple Transfers

Let’s start with the simplest case: a Transfer between two accounts belonging to the same ledger. Besides the source and destination accounts and ledger, we’ll also allow users of our repository to add some metadata: code, userData128, userData64, and userData32. Although optional, these metadata fields are useful to link this Transfer to external systems.

public UUID createSimpleTransfer(UUID sourceAccount, UUID targetAccount, BigInteger amount,
  int ledger, int code, UUID userData128, long userData64, int userData32)  {
    var id = UInt128.id();
    var batch = new TransferBatch(1);

    batch.add();
    batch.setId(id);
    batch.setAmount(amount);
    batch.setCode(code);
    batch.setCreditAccountId(UInt128.asBytes(targetAccount));
    batch.setDebitAccountId(UInt128.asBytes(sourceAccount));
    batch.setUserData32(userData32);
    batch.setUserData64(userData64);
    batch.setUserData128(UInt128.asBytes(userData128));
    batch.setLedger(ledger);

    var batchResults = client.createTransfers(batch);

    if (batchResults.getLength() > 0) {
        batchResults.next();
        throw new TransferException(batchResults.getResult());
    }
    return UInt128.asUUID(id);
}

If the operation succeeds, the amount will be added to the source Account‘s debitsPosted field and the destination Account‘s creditsPosted.

5.5. Balance Queries

When an Account is created with the HISTORY flag set, we can query its balances as they change as a result of a transfer. The API expects an AccountFilter filled with the Account identifier and a time range. The filter also supports parameters to limit the amount and order of returned entries.

This is how we’ve used the getAccountBalances() method to implement the listAccountBalances() repository’s method:

List<Balance> listAccountBalances(UUID accountId, Instant start, Instant end, int limit, boolean lastFirst) {
    var filter = new AccountFilter();
    filter.setAccountId(UInt128.asBytes(accountId));
    filter.setCredits(true);
    filter.setDebits(true);
    filter.setLimit(limit);
    filter.setReversed(lastFirst);
    filter.setTimestampMin(start.toEpochMilli());
    filter.setTimestampMax(end.toEpochMilli());

    var batch = client.getAccountBalances(filter);
    var result = new ArrayList<Balance>();
    while(batch.next()) {
        result.add(
          Balance.builder()
            .accountId(accountId)
            .debitsPending(batch.getDebitsPending())
            .debitsPosted(batch.getDebitsPosted())
            .creditsPending(batch.getCreditsPending())
            .creditsPosted(batch.getCreditsPosted())
            .timestamp(Instant.ofEpochMilli(batch.getTimestamp()))
            .build()
        );
    }

    return result;
 }

Notice that the API’s result doesn’t include information about the associated transactions, thus limiting its use in practice. However, as mentioned in the official documentation, this API is likely to change in future versions.

5.6. Transfer Queries

Currently, the getAccountTransfers() is the most useful of the available query APIs – not a big achievement given there are only two ;^). It works similarly to getAccountBalances(), including the use of AccountFilter to specify the query criteria:

public List<Transfer> listAccountTransfers(UUID accountId, Instant start, Instant end, int limit, boolean lastFirst) {
    var filter = new AccountFilter();
    filter.setAccountId(UInt128.asBytes(accountId));
    filter.setCredits(true);
    filter.setDebits(true);
    filter.setReversed(lastFirst);
    filter.setTimestampMin(start.toEpochMilli());
    filter.setTimestampMax(end.toEpochMilli());
    filter.setLimit(limit);

    var batch = client.getAccountTransfers(filter);
    var result = new ArrayList<Transfer>();
    while(batch.next()) {
        result.add(Transfer.builder()
          .id(UInt128.asUUID(batch.getId()))
          .code(batch.getCode())
          .amount(batch.getAmount())
          .flags(batch.getFlags())
          .ledger(batch.getLedger())
          .creditAccountId(UInt128.asUUID(batch.getCreditAccountId()))
          .debitAccountId(UInt128.asUUID(batch.getDebitAccountId()))
          .userData128(UInt128.asUUID(batch.getUserData128()))
          .userData64(batch.getUserData64())
          .userData32(batch.getUserData32())
          .timestamp(Instant.ofEpochMilli(batch.getTimestamp()))
          .pendingId(UInt128.asUUID(batch.getPendingId()))
          .build());
    }

    return result;
}

5.7. Two-Phase Transfers

TigerBeetle makes a clear distinction between pending and posted transfers. This distinction is made evident by the fact that an Account has four balance fields: two for posted and two for pending values.

In our earlier Transfer example, we didn’t inform its type. In this case, the API defaults to a posted Transfer, meaning the amount will be added directly to the debits_posted or credits_posted field.

To create a pending Transfer, we have to set the PENDING flag:

public UUID createPendingTransfer(UUID sourceAccount, UUID targetAccount, BigInteger amount,
  int ledger, int code, UUID userData128, long userData64, int userData32) throws ConcurrencyExceededException {

    var id = UInt128.id();
    var batch = new TransferBatch(1);
    // ... fill batch data (same as regular Transfer)
    batch.setFlags(TransferFlags.PENDING);

    // ... send transfer and handle results (same as regular Transfer) 
}

A pending Transfer should always be confirmed (POST_PENDING) or canceled (VOID_PENDING) by a later Transfer request. In both cases, we must include the original Transfer identifier in the pendingId field:

public UUID completePendingTransfer(UUID pendingId, boolean success) throws ConcurrencyExceededException {
    var id = UInt128.id();
    var batch = new TransferBatch(1);

    batch.add();
    batch.setId(id)
    batch.setPendingId(UInt128.asBytes(pendingId));
    batch.setFlags(success? TransferFlags.POST_PENDING_TRANSFER : TransferFlags.VOID_PENDING_TRANSFER);

    var batchResults = client.createTransfers(batch);

    if (batchResults.getLength() > 0) {
        batchResults.next();
        throw new TransferException(batchResults.getResult());
    }
    return UInt128.asUUID(id);
}

A typical scenario where this feature can be used is an authorization server that handles requests from an ATM. Firstly, the client informs his account and the requested amount to withdraw. The authorization server then creates a PENDING transaction and returns the generated Transfer identifier.

Next, the ATM proceeds to dispense the money. There are two possible outcomes: if everything goes right, the ATM sends another message to the authorization server and confirms the Transfer.

However, if something goes wrong (e.g. there are no bills available or a stuck dispenser), the ATM cancels the Transfer.

5.8. Two-Phase Transfer Timeouts

To account for a communication failure happening between the initial authorization request and its confirmation or cancellation, we can pass an optional timeout in the first step:

public UUID createExpirablePendingTransfer(UUID sourceAccount, UUID targetAccount, BigInteger amount,
  int ledger, int code, UUID userData128, long userData64, int userData32, int timeout) throws ConcurrencyExceededException {

    var id = UInt128.id();
    var batch = new TransferBatch(1);
    // ... prepare batch (same as regular pending Transfer)
    batch.setTimeout(timeout);

    // ... send batch and handle results (same as regular pending Transfer)
}

If the timeout expires before receiving a request to confirm or cancel it, TigerBeetle will automatically the pending transaction.

5.9. Linked Operations

Often, it’s important to ensure that a group of operations sent to TigerBeetle must either complete or fail as a whole. We can think of it as an analog to regular database transactions, where we can issue multiple inserts and commit them all at once at the end.

To support this scenario TigerBeetle has the concept of linked events. In a nutshell, to create a group of Account or Transfer records as a single transaction, all items except for the last must have the linked flag set:

public List<Map.Entry<UUID,CreateTransferResult>> createLinkedTransfers(List<Transfer> transfers) 
  throws ConcurrencyExceededException {
    var results = new ArrayList<Map.Entry<UUID,CreateTransferResult>>(transfers.size());
    var batch = new TransferBatch(transfers.size());
    for ( Transfer t : transfers) {
        byte[] id = UInt128.id();
        batch.add();
        batch.setId(id);

        // Is this the last transfer to add ?
        if ( batch.getPosition() != transfers.size() -1 ) {
            batch.setFlags(TransferFlags.LINKED);
        }

        batch.setLedger(t.ledger());
        batch.setAmount(t.amount());
        batch.setDebitAccountId(UInt128.asBytes(t.debitAccountId()));
        batch.setCreditAccountId(UInt128.asBytes(t.creditAccountId()));
        if ( t.userData128() != null) {
            batch.setUserData128(UInt128.asBytes(t.userData128()));
        }
        batch.setCode(t.code());
        results.add(new AbstractMap.SimpleImmutableEntry<>(UInt128.asUUID(id), CreateTransferResult.Ok));
    }

    var batchResult = client.createTransfers(batch);
    while(batchResult.next()) {
        var original = results.get(batchResult.getIndex());
        results.set(batchResult.getIndex(), new AbstractMap.SimpleImmutableEntry<>(original.getKey(), batchResult.getResult()));
    }

    return results;
}

TigerBeetle ensures that linked operations will be executed in order and either committed or rolled back fully. Also important is the fact that the side effects of one operation will be visible to the next in the chain.

For instance, consider an Account created with the DEBITS_MUST_NOT_EXCEED_CREDITS flag. If we create two linked transfer commands such that the second one results in an overdraft, both transfers will be rejected:

@Test
void whenSimpleTransfer_thenSuccess() throws Exception {
    var MY_LEDGER = 1000;
    var CHECKING_ACCOUNT = 1000;
    var P2P_TRANSFER = 500;

    var liabilitiesAcc = repo.createAccount(
      BigInteger.valueOf(1000L),
      CHECKING_ACCOUNT,
      MY_LEDGER, 0,0, 0);

    var sourceAcc = repo.createAccount(
      BigInteger.valueOf(1001L),
      CHECKING_ACCOUNT,
      MY_LEDGER, 0,0, AccountFlags.DEBITS_MUST_NOT_EXCEED_CREDITS);

    var targetAcc = repo.createAccount(
      BigInteger.valueOf(1002L),
      CHECKING_ACCOUNT,
      MY_LEDGER, 0, 0, 0);

    List<Transfer> transfers = List.of(
      Transfer.builder()
        .debitAccountId(liabilitiesAcc.id())
        .ledger(MY_LEDGER)
        .code(P2P_TRANSFER)
        .creditAccountId(sourceAcc.id())
        .amount(BigInteger.valueOf(1_000L))
        .build(),
      Transfer.builder()
        .debitAccountId(sourceAcc.id())
        .ledger(MY_LEDGER)
        .code(P2P_TRANSFER)
        .creditAccountId(targetAcc.id())
        .amount(BigInteger.valueOf(2_000L))
        .build()
      );

    var results = repo.createLinkedTransfers(transfers);
    assertEquals(2, results.size());
    assertEquals(CreateTransferResult.LinkedEventFailed, results.get(0).getValue());
    assertEquals(CreateTransferResult.ExceedsCredits, results.get(1).getValue());
}

In this case, we see that the first Transfer, which would succeed in a non-linked scenario, fails because the second one would result in an overdraft.

6. Conclusion

In this article, we’ve explored the TigerBeetle database and its features. Despite its limited query capabilities, it has great performance and runtime guarantees, making it a good candidate for every application where its double-entry ledger model is applicable.