A newer version of this documentation is available.

View Latest

Distributed Transactions from the Java SDK

    +
    A practical guide to using Couchbase’s distributed ACID transactions, via the Java API.

    This document presents a practical HOWTO on using Couchbase transactions, following on from our transactions documentation.

    You may also want to start with our transactions examples repository, which features useful downloadable examples of using Distributed Transactions.

    Javadocs are available online.

    Requirements

    • Couchbase Server 6.6.1 or above.

    • Couchbase Java client 3.1.5 or above. It is recommended to follow the transitive dependency for the transactions library from Maven.

    • NTP should be configured so nodes of the Couchbase cluster are in sync with time.

    • The application, if it is using extended attributes (XATTRs), must avoid using the XATTR field txn, which is reserved for Couchbase use.

    If using a single node cluster (for example, during development), then note that the default number of replicas for a newly created bucket is 1. If left at this default, then all Key-Value writes performed at with durability will fail with a DurabilityImpossibleException. In turn this will cause all transactions (which perform all Key-Value writes durably) to fail. This setting can be changed via GUI or command line. If the bucket already existed, then the server needs to be rebalanced for the setting to take effect.

    Getting Started

    Couchbase transactions require no additional components or services to be configured. Simply add the transactions library into your project. The latest version, as of May 2022, is 1.2.4.

    With Gradle this can be accomplished by modifying these sections of your build.gradle file like so:

    dependencies {
        compile group: 'com.couchbase.client', name: 'couchbase-transactions', version: '1.2.4'
    }

    Or with Maven:

    <dependency>
        <groupId>com.couchbase.client</groupId>
        <artifactId>couchbase-transactions</artifactId>
        <version>1.2.4</version>
    </dependency>

    This will automatically pull in any transitive dependencies of this library, including a compatible version of the Couchbase Java SDK.

    A complete simple gradle project is available on our transactions examples repository.

    Initializing Transactions

    To make it easy to use any of the following examples, here are all the imports used by them:

    import java.time.Duration;
    import java.util.Arrays;
    import java.util.List;
    import java.util.Optional;
    import java.util.concurrent.atomic.AtomicBoolean;
    import java.util.logging.Logger;
    
    import com.couchbase.client.core.cnc.Event;
    import com.couchbase.client.core.cnc.RequestSpan;
    import com.couchbase.client.java.Bucket;
    import com.couchbase.client.java.Cluster;
    import com.couchbase.client.java.Collection;
    import com.couchbase.client.java.Scope;
    import com.couchbase.client.java.ReactiveCollection;
    import com.couchbase.client.java.ReactiveScope;
    import com.couchbase.client.java.Scope;
    import com.couchbase.client.java.json.JsonArray;
    import com.couchbase.client.java.json.JsonObject;
    import com.couchbase.client.java.kv.GetResult;
    import com.couchbase.client.java.query.QueryOptions;
    import com.couchbase.client.java.query.QueryProfile;
    import com.couchbase.client.java.query.QueryResult;
    import com.couchbase.client.java.query.QueryScanConsistency;
    import com.couchbase.client.tracing.opentelemetry.OpenTelemetryRequestSpan;
    import com.couchbase.transactions.SingleQueryTransactionResult;
    import com.couchbase.transactions.TransactionDurabilityLevel;
    import com.couchbase.transactions.TransactionGetResult;
    import com.couchbase.transactions.TransactionQueryOptions;
    import com.couchbase.transactions.TransactionResult;
    import com.couchbase.transactions.Transactions;
    import com.couchbase.transactions.config.PerTransactionConfigBuilder;
    import com.couchbase.transactions.config.SingleQueryTransactionConfigBuilder;
    import com.couchbase.transactions.config.TransactionConfigBuilder;
    import com.couchbase.transactions.deferred.TransactionSerializedContext;
    import com.couchbase.transactions.error.TransactionCommitAmbiguous;
    import com.couchbase.transactions.error.TransactionFailed;
    import com.couchbase.transactions.log.IllegalDocumentState;
    import com.couchbase.transactions.log.LogDefer;
    import com.couchbase.transactions.log.TransactionCleanupAttempt;
    import com.couchbase.transactions.log.TransactionCleanupEndRunEvent;
    
    import io.opentelemetry.api.trace.Span;
    import reactor.core.publisher.Flux;
    import reactor.core.publisher.Mono;
    import reactor.core.scheduler.Schedulers;

    The starting point is the Transactions object. It is very important that the application ensures that only one of these is created, as it performs automated background processes that should not be duplicated.

    // Initialize the Couchbase cluster
    Cluster cluster = Cluster.connect("localhost", "username", "password");
    Bucket bucket = cluster.bucket("travel-sample");
    Scope scope = bucket.scope("inventory");
    Collection collection = scope.collection("airport");
    
    // Create the single Transactions object
    Transactions transactions = Transactions.create(cluster);

    Multiple Transactions Objects

    Generally an application will need just one Transactions object, and in fact the library will usually warn if more are created. Each Transactions object uses some resources, including a thread-pool.

    There is one rare exception where an application may need to create multiple Transactions objects, which is covered in Custom Metadata Collections.

    Configuration

    Transactions can optionally be globally configured at the point of creating the Transactions object:

    Transactions transactions = Transactions.create(cluster,
            TransactionConfigBuilder.create().durabilityLevel(TransactionDurabilityLevel.PERSIST_TO_MAJORITY)
                    .logOnFailure(true, Event.Severity.WARN)
                    .build());

    The default configuration will perform all writes with the durability setting Majority, ensuring that each write is available in-memory on the majority of replicas before the transaction continues. There are two higher durability settings available that will additionally wait for all mutations to be written to physical storage on either the active or the majority of replicas, before continuing. This further increases safety, at a cost of additional latency.

    A level of None is present but its use is discouraged and unsupported. If durability is set to None, then ACID semantics are not guaranteed.

    Creating a Transaction

    A core idea of Couchbase transactions is that an application supplies the logic for the transaction inside a lambda, including any conditional logic required, and the transaction is then automatically committed. If a transient error occurs, such as a temporary conflict with another transaction, then the transaction will rollback what has been done so far and run the lambda again. The application does have to do these retries and error handling itself.

    Each run of the lambda is called an attempt, inside an overall transaction.

    As with the Couchbase Java Client, you can use the library in either synchronous mode:

    try {
        transactions.run((ctx) -> {
            // 'ctx' is an AttemptContext, which permits getting, inserting,
            // removing and replacing documents, performing N1QL queries, and committing or
            // rolling back the transaction.
    
            // ... Your transaction logic here ...
    
            // This call is optional - if you leave it off, the transaction
            // will be committed anyway.
            ctx.commit();
        });
    } catch (TransactionCommitAmbiguous e) {
        // The application will of course want to use its own logging rather
        // than System.err
        System.err.println("Transaction possibly committed");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    } catch (TransactionFailed e) {
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    or asynchronous modes, using the Project Reactor reactive library:

    Mono<TransactionResult> result = transactions.reactive().run((ctx) -> {
        // 'ctx' is an AttemptContextReactive, providing asynchronous versions of the
        // AttemptContext methods
    
        return
    
        // Your transaction logic here: as an example, get and remove a doc
        ctx.get(collection.reactive(), "document-id").flatMap(doc -> ctx.remove(doc))
    
                // The commit call is optional - if you leave it off,
                // the transaction will be committed anyway.
                .then(ctx.commit());
    }).doOnError(err -> {
        if (err instanceof TransactionCommitAmbiguous) {
            System.err.println("Transaction possibly committed: ");
        } else {
            System.err.println("Transaction failed: ");
        }
    
        for (LogDefer e : ((TransactionFailed) err).result().log().logs()) {
            // System.err is used for example, log failures to your own logging system
            System.err.println(err.toString());
        }
    });
    
    // Normally you will chain this result further and ultimately subscribe. For
    // simplicity, here we just block on the result.
    TransactionResult finalResult = result.block();

    The synchronous mode is the easiest to write and understand. The asynchronous API allows you to build your application in a reactive style, which can help you scale with excellent efficiency. Those new to reactive programming may want to check out the Project Reactor site for more details on this powerful paradigm.

    The lambda gets passed an AttemptContext object, generally referred to as ctx here.

    Since the lambda may be rerun multiple times, it is important that it does not contain any side effects. In particular, you should never perform regular operations on a Collection, such as collection.insert(), inside the lambda. Such operations may be performed multiple times, and will not be performed transactionally. Instead such operations must be done through the ctx object, e.g. ctx.insert().

    Examples

    A code example is worth a thousand words, so here is a quick summary of the main transaction operations. They are described in more detail below.

    With the synchronous API
    Scope inventory = cluster.bucket("travel-sample").scope("inventory");
    
    try {
        TransactionResult result = transactions.run((ctx) -> {
            // Inserting a doc:
            ctx.insert(collection, "doc-a", JsonObject.create());
    
            // Getting documents:
            // Use ctx.getOptional if the document may or may not exist
            Optional<TransactionGetResult> docOpt = ctx.getOptional(collection, "doc-a");
    
            // Use ctx.get if the document should exist, and the transaction
            // will fail if it does not
            TransactionGetResult docA = ctx.get(collection, "doc-a");
    
            // Replacing a doc:
            TransactionGetResult docB = ctx.get(collection, "doc-b");
            JsonObject content = docB.contentAs(JsonObject.class);
            content.put("transactions", "are awesome");
            ctx.replace(docB, content);
    
            // Removing a doc:
            TransactionGetResult docC = ctx.get(collection, "doc-c");
            ctx.remove(docC);
    
            // Performing a SELECT N1QL query against a scope:
            QueryResult qr = ctx.query(inventory, "SELECT * FROM hotel WHERE country = $1",
                    TransactionQueryOptions.queryOptions()
                            .parameters(JsonArray.from("United Kingdom")));
            List<JsonObject> rows = qr.rowsAs(JsonObject.class);
    
            // Performing an UPDATE N1QL query on multiple documents, in the `inventory` scope:
            ctx.query(inventory, "UPDATE route SET airlineid = $1 WHERE airline = $2",
                    TransactionQueryOptions.queryOptions()
                            .parameters(JsonArray.from("airline_137", "AF")));
    
            // Committing (the ctx.commit() call is optional)
            ctx.commit();
        });
    } catch (TransactionCommitAmbiguous e) {
        System.err.println("Transaction possibly committed");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    } catch (TransactionFailed e) {
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }
    With the asynchronous API
    ReactiveScope inventory = cluster.bucket("travel-sample").scope("inventory").reactive();
    
    Mono<TransactionResult> result = transactions.reactive().run((ctx) -> {
        return
        // Inserting a doc:
        ctx.insert(collection.reactive(), "doc-a", JsonObject.create())
    
                // Getting and replacing a doc:
                .then(ctx.get(collection.reactive(), "doc-b")).flatMap(docB -> {
                    JsonObject content = docB.contentAs(JsonObject.class);
                    content.put("transactions", "are awesome");
                    return ctx.replace(docB, content);
                })
    
                // Getting and removing a doc:
                .then(ctx.get(collection.reactive(), "doc-c"))
                    .flatMap(doc -> ctx.remove(doc))
    
                // Performing a SELECT N1QL query, in the `inventory` scope:
                .then(ctx.query(inventory, "SELECT * FROM hotel WHERE country = $1",
                        TransactionQueryOptions.queryOptions()
                                .parameters(JsonArray.from("United Kingdom"))))
    
                .doOnNext(queryResult -> queryResult.rowsAs(JsonObject.class)
                        .doOnNext(row -> {
                            // the application would do something with each row here
                        }))
    
                // Performing an UPDATE N1QL query on multiple documents, in the `inventory` scope:
                .then(ctx.query(inventory, "UPDATE route SET airlineid = $1 WHERE airline = $2",
                        TransactionQueryOptions.queryOptions()
                                .parameters(JsonArray.from("airline_137", "AF"))))
    
                .then(ctx.commit());
    
    }).doOnError(err -> {
        if (err instanceof TransactionCommitAmbiguous) {
            System.err.println("Transaction possibly committed: ");
        } else {
            System.err.println("Transaction failed: ");
        }
    
        for (LogDefer e : ((TransactionFailed) err).result().log().logs()) {
            // System.err is used for example, log failures to your own logging system
            System.err.println(err.toString());
        }
    });
    
    // Normally you will chain this result further and ultimately subscribe.
    // For simplicity, here we just block on the result.
    result.block();

    Transaction Mechanics

    While this document is focussed on presenting how transactions are used at the API level, it is useful to have a high-level understanding of the mechanics. Reading this section is completely optional.

    Recall that the application-provided lambda (containing the transaction logic) may be run multiple times by Couchbase transactions. Each such run is called an attempt inside the overall transaction.

    Active Transaction Record Entries

    The first mechanic is that each of these attempts adds an entry to a metadata document in the Couchbase cluster. These metadata documents:

    • Are named Active Transaction Records, or ATRs.

    • Are created and maintained automatically.

    • Begin with "_txn:atr-".

    • Each contain entries for multiple attempts.

    • Are viewable, and they should not be modified externally.

    Each such ATR entry stores some metadata and, crucially, whether the attempt has committed or not. In this way, the entry acts as the single point of truth for the transaction, which is essential for providing an 'atomic commit' during reads.

    Staged Mutations

    The second mechanic is that mutating a document inside a transaction, does not directly change the body of the document. Instead, the post-transaction version of the document is staged alongside the document (technically in its extended attributes (XATTRs)). In this way, all changes are invisible to all parts of the Couchbase Data Platform until the commit point is reached.

    These staged document changes effectively act as a lock against other transactions trying to modify the document, preventing write-write conflicts.

    Cleanup

    There are safety mechanisms to ensure that leftover staged changes from a failed transaction cannot block live transactions indefinitely. These include an asynchronous cleanup process that is started with the creation of the Transactions object, and scans for expired transactions created by any application, on all buckets.

    Note that if an application is not running, then this cleanup is also not running.

    The cleanup process is detailed below in Asynchronous Cleanup.

    Committing

    Only once the lambda has successfully run to conclusion, will the attempt be committed. This updates the ATR entry, which is used as a signal by transactional actors to use the post-transaction version of a document from its XATTRs. Hence, updating the ATR entry is an 'atomic commit' switch for the transaction.

    After this commit point is reached, the individual documents will be committed (or "unstaged"). This provides an eventually consistent commit for non-transactional actors.

    Key-Value Mutations

    Replacing

    Replacing a document requires a ctx.get() call first. This is necessary so that the transactions library can check that the document is not involved in another transaction. If it is, then the transactions library will handle this at the ctx.replace() point. Generally, this involves rolling back what has been done so far, and retrying the lambda.

    With the synchronous API:
    transactions.run((ctx) -> {
        TransactionGetResult doc = ctx.get(collection, "doc-id");
        JsonObject content = doc.contentAs(JsonObject.class);
        content.put("transactions", "are awesome");
        ctx.replace(doc, content);
    });
    Asynchronous API:
    transactions.reactive().run((ctx) -> {
        return ctx.get(collection.reactive(), "doc-id").flatMap(doc -> {
            JsonObject content = doc.contentAs(JsonObject.class);
            content.put("transactions", "are awesome");
            return ctx.replace(doc, content);
        }).then(ctx.commit());
    });

    Removing

    As with replaces, removing a document requires a ctx.get() call first.

    Synchronous API:
    transactions.run((ctx) -> {
        TransactionGetResult doc = ctx.get(collection, "doc-id");
        ctx.remove(doc);
    });
    With the asynchronous API:
    transactions.reactive().run((ctx) -> {
        return ctx.get(collection.reactive(), "anotherDoc").flatMap(doc -> ctx.remove(doc));
    });
    For those using the asynchronous API - some ctx methods, notably ctx.remove(), return Mono<Void>. There is a common 'gotcha' with Mono<Void> in that it does not trigger a 'next' reactive event - only a 'completion' event. This means that some reactive operators chained afterwards - including the common flatMap - will not trigger. Generally, you will to do ctx.remove(…​).then(…​) rather than ctx.remove(…​).flatMap(…​).

    Inserting

    With the asynchronous API:
    transactions.reactive().run((ctx) -> {
        return ctx.insert(collection.reactive(), "docId", JsonObject.create()).then();
    }).block();
    With the synchronous API:
    transactions.run((ctx) -> {
        String docId = "docId";
    
        ctx.insert(collection, docId, JsonObject.create());
    
    });

    Key-Value Reads

    There are two ways to get a document with Key-Value, get and getOptional:

    transactions.run((ctx) -> {
        String docId = "a-doc";
    
        Optional<TransactionGetResult> docOpt = ctx.getOptional(collection, docId);
        TransactionGetResult doc = ctx.get(collection, docId);
    });

    get will cause the transaction to fail with TransactionFailed (after rolling back any changes, of course). It is provided as a convenience method so the developer does not have to check the Optional if the document must exist for the transaction to succeed.

    Gets will 'read your own writes', e.g. this will succeed:

    transactions.run((ctx) -> {
        String docId = "docId";
    
        ctx.insert(collection, docId, JsonObject.create());
    
        Optional<TransactionGetResult> doc = ctx.getOptional(collection, docId);
    
        assert (doc.isPresent());
    });

    N1QL Queries

    As of Couchbase Server 7.0, N1QL queries may be used inside the transaction lambda, freely mixed with Key-Value operations.

    BEGIN TRANSACTION

    There are two ways to initiate a transaction with Couchbase 7.0: via a transactions library, and via the query service directly using BEGIN TRANSACTION. The latter is intended for those using query via the REST API, or using the query workbench in the UI, and it is strongly recommended that application writers instead use the transactions library. This provides these benefits:

    • It automatically handles errors and retrying.

    • It allows Key-Value operations and N1QL queries to be freely mixed.

    • It takes care of issuing BEGIN TRANSACTION, END TRANSACTION, COMMIT and ROLLBACK automatically. These become an implementation detail and you should not use these statements inside the lambda.

    Supported N1QL

    The majority of N1QL DML statements are permitted within a transaction. Specifically: INSERT, UPSERT, DELETE, UPDATE, MERGE and SELECT are supported.

    DDL statements, such as CREATE INDEX, are not.

    Using N1QL

    If you already use N1QL from the Java SDK, then its use in transactions is very similar. It returns the same QueryResult you are used to, and takes most of the same options.

    You must take care to write ctx.query() inside the lambda however, rather than cluster.query() or scope.query().

    An example of selecting some rows from the travel-sample bucket:

    transactions.run((ctx) -> {
        String st = "SELECT * FROM `travel-sample`.inventory.hotel WHERE country = $1";
        QueryResult qr = ctx.query(st, TransactionQueryOptions.queryOptions()
                .parameters(JsonArray.from("United Kingdom")));
        List<JsonObject> rows = qr.rowsAs(JsonObject.class);
    });

    Rather than specifying the full "`travel-sample`.inventory.hotel" name each time, it is easier to pass a reference to the inventory Scope:

    Bucket travelSample = cluster.bucket("travel-sample");
    Scope inventory = travelSample.scope("inventory");
    
    transactions.run((ctx) -> {
        QueryResult qr = ctx.query(inventory, "SELECT * FROM hotel WHERE country = $1",
                TransactionQueryOptions.queryOptions()
                        .parameters(JsonArray.from("United States")));
        List<JsonObject> rows = qr.rowsAs(JsonObject.class);
    });

    An example using a Scope for an UPDATE:

    String hotelChain = "http://marriot%";
    String country = "United States";
    
    transactions.run((ctx) -> {
        QueryResult qr = ctx.query(inventory, "UPDATE hotel SET price = $1 WHERE url LIKE $2 AND country = $3",
                TransactionQueryOptions.queryOptions()
                        .parameters(JsonArray.from(99.99, hotelChain, country)));
        assert(qr.metaData().metrics().get().mutationCount() == 1);
    });

    And an example combining SELECTs and UPDATEs. It’s possible to call regular Java methods from the lambda, as shown here, permitting complex logic to be performed. Just remember that since the lambda may be called multiple times, so may the method.

    transactions.run((ctx) -> {
        // Find all hotels of the chain
        QueryResult qr = ctx.query(inventory, "SELECT reviews FROM hotel WHERE url LIKE $1 AND country = $2",
                TransactionQueryOptions.queryOptions()
                        .parameters(JsonArray.from(hotelChain, country)));
    
        // This function (not provided here) will use a trained machine learning model to provide a
        // suitable price based on recent customer reviews.
        double updatedPrice = priceFromRecentReviews(qr);
    
        // Set the price of all hotels in the chain
        ctx.query(inventory, "UPDATE hotel SET price = $1 WHERE url LIKE $2 AND country = $3",
                TransactionQueryOptions.queryOptions()
                        .parameters(JsonArray.from(updatedPrice, hotelChain, country)));
    });

    Read Your Own Writes

    As with Key-Value operations, N1QL queries support Read Your Own Writes.

    This example shows inserting a document and then selecting it again.

    transactions.run((ctx) -> {
        ctx.query("INSERT INTO `default` VALUES ('doc', {'hello':'world'})");  (1)
    
        // Performing a 'Read Your Own Write'
        String st = "SELECT `default`.* FROM `default` WHERE META().id = 'doc'"; (2)
        QueryResult qr = ctx.query(st);
        assert(qr.metaData().metrics().get().resultCount() == 1);
    });
    1 The inserted document is only staged at this point. as the transaction has not yet committed. Other transactions, and other non-transactional actors, will not be able to see this staged insert yet.
    2 But the SELECT can, as we are reading a mutation staged inside the same transaction.

    Mixing Key-Value and N1QL

    Key-Value operations and queries can be freely intermixed, and will interact with each other as you would expect.

    In this example we insert a document with Key-Value, and read it with a SELECT.

    transactions.run((ctx) -> {
        ctx.insert(collection, "doc", JsonObject.create().put("hello", "world")); (1)
    
        // Performing a 'Read Your Own Write'
        String st = "SELECT `default`.* FROM `default` WHERE META().id = 'doc'"; (2)
        QueryResult qr = ctx.query(st);
        assert(qr.metaData().metrics().get().resultCount() == 1);
    });
    1 As with the 'Read Your Own Writes' example, here the insert is only staged, and so it is not visible to other transactions or non-transactional actors.
    2 But the SELECT can view it, as the insert was in the same transaction.

    Query Options

    Query options can be provided via TransactionQueryOptions, which provides a subset of the options in the Java SDK’s QueryOptions.

    transactions.run((ctx) -> {
        ctx.query("INSERT INTO `default` VALUES ('doc', {'hello':'world'})",
                TransactionQueryOptions.queryOptions().profile(QueryProfile.TIMINGS));
    });

    The supported options are:

    • parameters

    • scanConsistency

    • flexIndex

    • serializer

    • clientContextId

    • scanWait

    • scanCap

    • pipelineBatch

    • pipelineCap

    • profile

    • readonly

    • adhoc

    • raw

    See the QueryOptions documentation for details on these.

    Query Concurrency

    Only one query statement will be performed by the query service at a time. Non-blocking mechanisms can be used to perform multiple concurrent query statements, but this may result internally in some added network traffic due to retries, and is unlikely to provide any increased performance.

    Query Performance Advice

    This section is optional reading, and only for those looking to maximize transactions performance.

    After the first query statement in a transaction, subsequent Key-Value operations in the lambda are converted into N1QL and executed by the query service rather than the Key-Value data service. The operation will behave identically, and this implementation detail can largely be ignored, except for these two caveats:

    • These converted Key-Value operations are likely to be slightly slower, as the query service is optimized for statements involving multiple documents. Those looking for the maximum possible performance are recommended to put Key-Value operations before the first query in the lambda, if possible.

    • Those using non-blocking mechanisms to achieve concurrency should be aware that the converted Key-Value operations are subject to the same parallelism restrictions mentioned above, e.g. they will not be executed in parallel by the query service.

    Single Query Transactions

    This section is mainly of use for those wanting to do large, bulk-loading transactions.

    The query service maintains where required some in-memory state for each document in a transaction, that is freed on commit or rollback. For most use-cases this presents no issue, but there are some workloads, such as bulk loading many documents, where this could exceed the server resources allocated to the service. Solutions to this include breaking the workload up into smaller batches, and allocating additional memory to the query service. Alternatively, single query transaction, described here, may be used.

    Single query transactions have these characteristics:

    • They have greatly reduced memory usage inside the query service.

    • As the name suggests, they consist of exactly one query, and no Key-Value operations.

    You will see reference elsewhere in Couchbase documentation to the tximplicit query parameter. Single query transactions internally are setting this parameter. In addition, they provide automatic error and retry handling.

    Single query transactions may be initiated like so:

    try {
        SingleQueryTransactionResult result = transactions.query(bulkLoadStatement);
    
        QueryResult queryResult = result.queryResult();
    } catch (TransactionCommitAmbiguous e) {
        System.err.println("Transaction possibly committed");
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    } catch (TransactionFailed e) {
        System.err.println("Transaction did not reach commit point");
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    You can also run a single query transaction against a particular Scope (these examples will exclude the full error handling for brevity):

    Bucket travelSample = cluster.bucket("travel-sample");
    Scope inventory = travelSample.scope("inventory");
    
    transactions.query(inventory, bulkLoadStatement);

    and configure it:

    transactions.query(bulkLoadStatement, SingleQueryTransactionConfigBuilder.create()
        // Single query transactions will often want to increase the default timeout
        .expirationTime(Duration.ofSeconds(360))
        .build());

    Query with KV Roles

    To execute a key-value operation within a transaction, users must have the relevant Administrative or Data RBAC roles, and permissions on the relevant buckets, scopes, and collections.

    Similarly, to run a query statement within a transaction, users must have the relevant Administrative or Query & Index RBAC roles, and permissions on the relevant buckets, scopes and collections.

    Refer to Roles for details.

    Query Mode
    When a transaction executes a query statement, the transaction enters query mode, which means that the query is executed with the user’s query permissions. Any key-value operations which are executed by the transaction after the query statement are also executed with the user’s query permissions. These may or may not be different to the user’s data permissions; if they are different, you may get unexpected results.

    Committing

    Committing is automatic: if there is no explicit call to ctx.commit() at the end of the transaction logic callback, and no exception is thrown, it will be committed.

    With the asynchronous API, if you leave off the explicit call to commit then you may need to call .then() on the result of the chain to convert it to the required Mono<Void> return type:

    Mono<TransactionResult> result = transactions.reactive().run((ctx) -> {
        return ctx.get(collection.reactive(), "anotherDoc").flatMap(doc -> {
            JsonObject content = doc.contentAs(JsonObject.class);
            content.put("transactions", "are awesome");
            return ctx.replace(doc, content);
        }).then();
    });

    As soon as the transaction is committed, all its changes will be atomically visible to reads from other transactions. The changes will also be committed (or "unstaged") so they are visible to non-transactional actors, in an eventually consistent fashion.

    Commit is final: after the transaction is committed, it cannot be rolled back, and no further operations are allowed on it.

    An asynchronous cleanup process ensures that once the transaction reaches the commit point, it will be fully committed - even if the application crashes.

    A Full Transaction Example

    Let’s pull together everything so far into a more real-world example of a transaction.

    This example simulates a simple Massively Multiplayer Online game, and includes documents representing:

    • Players, with experience points and levels;

    • Monsters, with hitpoints, and the number of experience points a player earns from their death.

    In this example, the player is dealing damage to the monster. The player’s client has sent this instruction to a central server, where we’re going to record that action. We’re going to do this in a transaction, as we don’t want a situation where the monster is killed, but we fail to update the player’s document with the earned experience.

    (Though this is just a demo - in reality, the game would likely live with the small risk and limited impact of this, rather than pay the performance cost for using a transaction.)

    A complete version of this example is available on our GitHub transactions examples page.

    public void playerHitsMonster(int damage, String playerId, String monsterId) {
        Transactions transactions = getTransactions();
    
        try {
            transactions.run((ctx) -> {
                TransactionGetResult monsterDoc = ctx.get(collection, monsterId);
                TransactionGetResult playerDoc = ctx.get(collection, playerId);
    
                int monsterHitpoints = monsterDoc.contentAs(JsonObject.class).getInt("hitpoints");
                int monsterNewHitpoints = monsterHitpoints - damage;
    
                if (monsterNewHitpoints <= 0) {
                    // Monster is killed. The remove is just for demoing, and a more realistic
                    // example would set a
                    // "dead" flag or similar.
                    ctx.remove(monsterDoc);
    
                    // The player earns experience for killing the monster
                    int experienceForKillingMonster = monsterDoc.contentAs(JsonObject.class)
                            .getInt("experienceWhenKilled");
                    int playerExperience = playerDoc.contentAs(JsonObject.class).getInt("experience");
                    int playerNewExperience = playerExperience + experienceForKillingMonster;
                    int playerNewLevel = calculateLevelForExperience(playerNewExperience);
    
                    JsonObject playerContent = playerDoc.contentAs(JsonObject.class);
    
                    playerContent.put("experience", playerNewExperience);
                    playerContent.put("level", playerNewLevel);
    
                    ctx.replace(playerDoc, playerContent);
                } else {
                    // Monster is damaged but still alive
                    JsonObject monsterContent = monsterDoc.contentAs(JsonObject.class);
                    monsterContent.put("hitpoints", monsterNewHitpoints);
    
                    ctx.replace(monsterDoc, monsterContent);
                }
            });
        } catch (TransactionFailed e) {
            // The operation failed. Both the monster and the player will be untouched.
    
            // Situations that can cause this would include either the monster
            // or player not existing (as get is used), or a persistent
            // failure to be able to commit the transaction, for example on
            // prolonged node failure.
        }
    }

    Concurrency with Non-Transactional Writes

    This release of transactions for Couchbase requires a degree of co-operation from the application. Specifically, the application should ensure that non-transactional writes are never done concurrently with transactional writes, on the same document.

    This requirement is to ensure that the strong Key-Value performance of Couchbase was not compromised. A key philosophy of our transactions is that you 'pay only for what you use'.

    If two such writes do conflict then the transactional write will 'win', overwriting the non-transactional write.

    Note this only applies to writes. Any non-transactional reads concurrent with transactions are fine, and are at a Read Committed level.

    To help detect that this co-operative requirement is fulfilled, the application can subscribe to the client’s event logger and check for any IllegalDocumentState events, like so:

    cluster.environment().eventBus().subscribe(event -> {
        if (event instanceof IllegalDocumentState) {
            // log this event for review
        }
    });

    These events will be raised in the event of a non-transactional write being detected and overridden. The event contains the key of the document involved, to aid the application with debugging.

    Rollback

    If an exception is thrown, either by the application from the lambda, or by the transactions library, then that attempt is rolled back. The transaction logic may or may not be retried, depending on the exception.

    If the transaction is not retried then it will throw a TransactionFailed exception, and its getCause method can be used for more details on the failure.

    The application can use this to signal why it triggered a rollback, as so:

    class BalanceInsufficient extends RuntimeException {
    }
    
    try {
        transactions.run((ctx) -> {
            TransactionGetResult customer = ctx.get(collection, "customer-name");
    
            if (customer.contentAsObject().getInt("balance") < costOfItem) {
                throw new BalanceInsufficient();
            }
            // else continue transaction
        });
    } catch (TransactionCommitAmbiguous e) {
        // This exception can only be thrown at the commit point, after the
        // BalanceInsufficient logic has been passed, so there is no need to
        // check getCause here.
        System.err.println("Transaction possibly committed");
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    } catch (TransactionFailed e) {
        if (e.getCause() instanceof BalanceInsufficient) {
            // Re-raise the error
            throw (RuntimeException) e.getCause();
        } else {
            System.err.println("Transaction did not reach commit point");
    
            for (LogDefer err : e.result().log().logs()) {
                System.err.println(err.toString());
            }
        }
    }

    The transaction can also be explicitly rolled back:

    transactions.run((ctx) -> {
        TransactionGetResult customer = ctx.get(collection, "customer-name");
    
        if (customer.contentAsObject().getInt("balance") < costOfItem) {
            ctx.rollback();
        }
        // else continue transaction
    });

    In this case, if ctx.rollback() is reached, then the transaction will be regarded as successfully rolled back and no TransactionFailed will be thrown.

    After a transaction is rolled back, it cannot be committed, no further operations are allowed on it, and the library will not try to automatically commit it at the end of the code block.

    Error Handling

    As discussed previously, Couchbase transactions will attempt to resolve many errors for you, through a combination of retrying individual operations and the application’s lambda. This includes some transient server errors, and conflicts with other transactions.

    But there are situations that cannot be resolved, and total failure is indicated to the application via errors. These errors include:

    • Any error thrown by your transaction lambda, either deliberately or through an application logic bug.

    • Attempting to insert a document that already exists.

    • Attempting to remove or replace a document that does not exist.

    • Calling ctx.get() on a document key that does not exist.

    Once one of these errors occurs, the current attempt is irrevocably failed (though the transaction may retry the lambda). It is not possible for the application to catch the failure and continue. Once a failure has occurred, all other operations tried in this attempt (including commit) will instantly fail.

    Transactions, as they are multi-stage and multi-document, also have a concept of partial success or failure. This is signalled to the application through the TransactionResult.unstagingComplete() method, described later.

    There are three exceptions that Couchbase transactions can raise to the application: TransactionFailed, TransactionExpired and TransactionCommitAmbiguous. All exceptions derive from TransactionFailed for backwards-compatibility purposes.

    TransactionFailed and TransactionExpired

    The transaction definitely did not reach the commit point. TransactionFailed indicates a fast-failure whereas TransactionExpired indicates that retries were made until the expiration point was reached, but this distinction is not normally important to the application and generally TransactionExpired does not need to be handled individually.

    Either way, an attempt will have been made to rollback all changes. This attempt may or may not have been successful, but the results of this will have no impact on the protocol or other actors. No changes from the transaction will be visible (presently with the potential and temporary exception of staged inserts being visible to non-transactional actors, as discussed under Inserting).

    Handling: Generally, debugging exactly why a given transaction failed requires review of the logs, so it is suggested that the application log these on failure (see Logging). The application may want to try the transaction again later. Alternatively, if transaction completion time is not a priority, then transaction expiration times (which default to 15 seconds) can be extended across the board through TransactionConfigBuilder.

    Transactions transactions = Transactions.create(cluster,
            TransactionConfigBuilder.create().expirationTime(Duration.ofSeconds(120)).build());

    This will allow the protocol more time to get past any transient failures (for example, those caused by a cluster rebalance). The tradeoff to consider with longer expiration times, is that documents that have been staged by a transaction are effectively locked from modification from other transactions, until the expiration time has exceeded.

    Note that expiration is not guaranteed to be followed precisely. For example, if the application were to do a long blocking operation inside the lambda (which should be avoided), then expiration can only trigger after this finishes. Similarly, if the transaction attempts a key-value operation close to the expiration time, and that key-value operation times out, then the expiration time may be exceeded.

    TransactionCommitAmbiguous

    As discussed previously, each transaction has a 'single point of truth' that is updated atomically to reflect whether it is committed.

    However, it is not always possible for the protocol to become 100% certain that the operation was successful, before the transaction expires. That is, the operation may have successfully completed on the cluster, or may succeed soon, but the protocol is unable to determine this (whether due to transient network failure or other reason). This is important as the transaction may or may not have reached the commit point, e.g. succeeded or failed.

    Couchbase transactions will raise TransactionCommitAmbiguous to indicate this state. It should be rare to receive this error.

    If the transaction had in fact successfully reached the commit point, then the transaction will be fully completed ("unstaged") by the asynchronous cleanup process at some point in the future. With default settings this will usually be within a minute, but whatever underlying fault has caused the TransactionCommitAmbiguous may lead to it taking longer.

    If the transaction had not in fact reached the commit point, then the asynchronous cleanup process will instead attempt to roll it back at some point in the future. If unable to, any staged metadata from the transaction will not be visible, and will not cause problems (e.g. there are safety mechanisms to ensure it will not block writes to these documents for long).

    Handling: This error can be challenging for an application to handle. As with TransactionFailed it is recommended that it at least writes any logs from the transaction, for future debugging. It may wish to retry the transaction at a later point, or globally extend transactional expiration times to give the protocol additional time to resolve the ambiguity.

    TransactionResult.unstagingComplete()

    This boolean flag indicates whether all documents were able to be unstaged (committed).

    For most use-cases it is not an issue if it is false. All transactional actors will still all the changes from this transaction, as though it had committed fully. The cleanup process is asynchronously working to complete the commit, so that it will be fully visible to non-transactional actors.

    The flag is provided for those rare use-cases where the application requires the commit to be fully visible to non-transactional actors, before it may continue. In this situation the application can raise an error here, or poll all documents involved until they reflect the mutations.

    If you regularly see this flag false, consider increasing the transaction expiration time to reduce the possibility that the transaction times out during the commit.

    Similar to TransactionResult, SingleQueryTransactionResult also has an unstagingComplete() method.

    Full Error Handling Example

    Pulling all of the above together, this is the suggested best practice for error handling:

    try {
        TransactionResult result = transactions.run((ctx) -> {
            // ... transactional code here ...
        });
    
        // The transaction definitely reached the commit point. Unstaging
        // the individual documents may or may not have completed
    
        if (!result.unstagingComplete()) {
            // In rare cases, the application may require the commit to have
            // completed.  (Recall that the asynchronous cleanup process is
            // still working to complete the commit.)
            // The next step is application-dependent.
        }
    } catch (TransactionCommitAmbiguous err) {
        // The transaction may or may not have reached commit point
        System.err.println("Transaction returned TransactionCommitAmbiguous and may have succeeded, logs:");
    
        // Of course, the application will want to use its own logging rather
        // than System.err
        err.result().log().logs().forEach(log -> System.err.println(log.toString()));
    } catch (TransactionFailed err) {
        // The transaction definitely did not reach commit point
        System.err.println("Transaction failed with TransactionFailed, logs:");
        err.result().log().logs().forEach(log -> System.err.println(log.toString()));
    }

    Asynchronous Cleanup

    Transactions will try to clean up after themselves in the advent of failures. However, there are situations that inevitably created failed, or 'lost' transactions, such as an application crash.

    This requires an asynchronous cleanup task, described in this section.

    Creating the Transactions object spawns a background cleanup task, whose job it is to periodically scan for expired transactions and clean them up. It does this by scanning a subset of the Active Transaction Record (ATR) transaction metadata documents, on each bucket. As you’ll recall from earlier, an entry for each transaction attempt exists in one of these documents. They are removed during cleanup or at some time after successful completion.

    The default settings are tuned to find expired transactions reasonably quickly, while creating negligible impact from the background reads required by the scanning process. To be exact, with default settings it will generally find expired transactions within 60 seconds, and use less than 20 reads per second. This is unlikely to impact performance on any cluster, but the settings may be tuned as desired.

    All applications connected to the same cluster and running transactions will share in the cleanup, via a low-touch communication protocol on the "_txn:client-record" metadata document that will be created in each bucket in the cluster. This document is visible and should not be modified externally as it is maintained automatically. All ATRs on a bucket will be distributed between all cleanup clients, so increasing the number of applications will not increase the reads required for scanning.

    An application may cleanup transactions created by another application.

    It is important to understand that if an application is not running, then cleanup is not running. This is particularly relevant to developers running unit tests or similar.

    If this is an issue, then the deployment may want to consider running a simple application at all times that just opens a transaction, to guarantee that cleanup is running.

    Configuring Cleanup

    The cleanup settings can be configured as so:

    Setting Default Description

    cleanupWindow

    60 seconds

    This determines how long a cleanup 'run' is; that is, how frequently this client will check its subset of ATR documents. It is perfectly valid for the application to change this setting, which is at a conservative default. Decreasing this will cause expiration transactions to be found more swiftly (generally, within this cleanup window), with the tradeoff of increasing the number of reads per second used for the scanning process.

    cleanupLostAttempts

    true

    This is the thread that takes part in the distributed cleanup process described above, that cleans up expired transactions created by any client. It is strongly recommended that it is left enabled.

    cleanupClientAttempts

    true

    This thread is for cleaning up transactions created just by this client. The client will preferentially aim to send any transactions it creates to this thread, leaving transactions for the distributed cleanup process only when it is forced to (for example, on an application crash). It is strongly recommended that it is left enabled.

    Monitoring Cleanup

    If the application wishes to monitor cleanup it may subscribe to these events:

    cluster.environment().eventBus().subscribe(event -> {
        if (event instanceof TransactionCleanupAttempt || event instanceof TransactionCleanupEndRunEvent) {
            // log this event
        }
    });

    TransactionCleanupEndRunEvent is raised whenever a current 'run' is finished, and contains statistics from the run. (A run is typically around every 60 seconds, with default configuration.)

    A TransactionCleanupAttempt event is raised when an expired transaction was found by this process, and a cleanup attempt was made. It contains whether that attempt was successful, along with any logs relevant to the attempt.

    In addition, if cleanup fails to cleanup a transaction that is more than two hours past expiry, it will raise the TransactionCleanupAttempt event at WARN level (rather than the default DEBUG). With most default configurations of the event-bus (see Logging below), this will cause that event to be logged somewhere visible to the application. If there is not a good reason for the cleanup to be failed (such as a downed node that has not yet been failed-over), then the user is encouraged to report the issue.

    Logging

    To aid troubleshooting, each transaction maintains a list of log entries, which can be logged on failure like this:

    } catch (TransactionCommitAmbiguous e) {
        // The application will of course want to use its own logging rather
        // than System.err
        System.err.println("Transaction possibly committed");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    } catch (TransactionFailed e) {
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    or for the asynchronous API:

    }).doOnError(err -> {
        if (err instanceof TransactionCommitAmbiguous) {
            System.err.println("Transaction possibly committed: ");
        } else {
            System.err.println("Transaction failed: ");
        }
    
        for (LogDefer e : ((TransactionFailed) err).result().log().logs()) {
            // System.err is used for example, log failures to your own logging system
            System.err.println(err.toString());
        }
    });

    A failed transaction can involve dozens, even hundreds, of lines of logging, so the application may prefer to write failed transactions into a separate file.

    For convenience there is also a config option that will automatically write this programmatic log to the standard Couchbase Java logging configuration inherited from the SDK if a transaction fails. This will log all lines of any failed transactions, to WARN level:

    .logOnFailure(true, Event.Severity.WARN)

    By default the Couchbase Java logging event-bus is setup to look for and use SLF4J/logback, log4j1, and log4j2 on the classpath, and to fallback to java.util.Logging.

    Please see the Java SDK logging documentation for details.

    Most applications will have their own preferred Java logging solution in-place already. For those starting from scratch here is a complete example using the basic java.util.Logging:

    final Logger LOGGER = Logger.getLogger("transactions");
    
    try {
        TransactionResult result = transactions.run((ctx) -> {
            // ... transactional code here ...
        });
    } catch (TransactionCommitAmbiguous err) {
        // The transaction may or may not have reached commit point
        LOGGER.info("Transaction returned TransactionCommitAmbiguous and" + " may have succeeded, logs:");
        err.result().log().logs().forEach(log -> LOGGER.info(log.toString()));
    } catch (TransactionFailed err) {
        // The transaction definitely did not reach commit point
        LOGGER.info("Transaction failed with TransactionFailed, logs:");
        err.result().log().logs().forEach(log -> LOGGER.info(log.toString()));
    }

    Tracing

    This telemetry is particularly useful for monitoring performance.

    If the underlying Couchbase Java SDK is configured for tracing, then no further work is required: transaction spans will be output automatically. See the Couchbase Java SDK Request Tracing documentation for how to configure this.

    Parent Spans

    While the above is sufficient to use and output transaction spans, the application may wish to indicate that the transaction is part of a larger span — for instance, a user request. It can do this by passing that as a parent span.

    If you have an existing OpenTelemetry span you can easily convert it to a Couchbase RequestSpan and pass it to the transactions library:

    Span span = Span.current(); // this is a span created by your code earlier
    RequestSpan wrapped = OpenTelemetryRequestSpan.wrap(span);
    
    transactions.run((ctx) -> {
        // your transaction
    }, PerTransactionConfigBuilder.create().parentSpan(wrapped).build());

    Concurrent Operations

    The reactive API allows operations to be performed concurrently inside a transaction, which can assist performance.

    Any users of the reactive API are very strongly advised to use at minimum the 1.2.3 release of the transactions library. Prior to this there were specific rules the application had to follow to get thread-safe results, and these rules can be found in the previous version of this document.

    An example of performing parallel operations using the reactive API:

    List<String> docIds = Arrays.asList("doc1", "doc2", "doc3", "doc4", "doc5");
    ReactiveCollection coll = collection.reactive();
    int concurrency = 100; // This many operations will be in-flight at once
    
    TransactionResult result = transactions.reactive((ctx) -> {
        return Flux.fromIterable(docIds)
                .parallel(concurrency)
                .runOn(Schedulers.boundedElastic())
                .concatMap(docId -> ctx.get(collection.reactive(), docId)
                        .flatMap(doc -> {
                            JsonObject content = doc.contentAsObject();
                            content.put("value", "updated");
                            return ctx.replace(doc, content);
                        }))
                .sequential()
                .then();
    }).block();

    Custom Metadata Collections

    As described earlier, transactions automatically create and use metadata documents. By default, these are created in the default collection of the bucket of the first mutated document in the transaction. Optionally, you can instead use a collection to store the metadata documents. Most users will not need to use this functionality, and can continue to use the default behavior. They are provided for these use-cases:

    • The metadata documents contain, for documents involved in each transaction, the document’s key and the name of the bucket, scope and collection it exists on. In some deployments this may be sensitive data.

    • You wish to remove the default collections. Before doing this, you should ensure that all existing transactions using metadata documents in the default collections have finished.

    Usage

    Custom metadata collections are enabled with:

    Collection metadataCollection = null; // this is a Collection opened by your code earlier
    Transactions transactions = Transactions.create(cluster,
            TransactionConfigBuilder.create().metadataCollection(metadataCollection));

    When specified:

    • Any transactions created from this Transactions object, will create and use metadata in that collection.

    • The asynchronous cleanup started by this Transactions object will be looking for expired transactions only in this collection.

    You need to ensure that this application has RBAC data read and write privileges to it, and should not delete the collection subsequently as it can interfere with existing transactions. You can use an existing collection or create a new one.

    Multiple Transactions Objects

    Generally, an application requires only one Transactions object. But in some deployments, an application may need to access multiple custom metadata collections, and it is reasonable to create multiple Transactions objects as this is the only way to enable this.

    The library will not warn about multiple Transactions objects being created in this scenario.

    Deferred Commits

    The deferred commit feature is currently in alpha, and the API may change.

    Deferred commits allow a transaction to be paused just before the commit point. Optionally, everything required to finish the transaction can then be bundled up into a context that may be serialized into a String or byte array, and deserialized elsewhere (for example, in another process). The transaction can then be committed, or rolled back.

    The intention behind this feature is to allow multiple transactions, potentially spanning multiple databases, to be brought to just before the commit point, and then all committed together.

    Here’s an example of deferring the initial commit and serializing the transaction:

    try {
        TransactionResult result = transactions.run((ctx) -> {
            JsonObject initial = JsonObject.create().put("val", 1);
            ctx.insert(collection, "a-doc-id", initial);
    
            // Defer means don't do a commit right now. `serialized` in the result will be
            // present.
            ctx.defer();
        });
    
        // Available because ctx.defer() was called
        assert (result.serialized().isPresent());
    
        TransactionSerializedContext serialized = result.serialized().get();
    
        // This is going to store a serialized form of the transaction to pass around
        byte[] encoded = serialized.encodeAsBytes();
    
    } catch (TransactionFailed e) {
        // System.err is used for example, log failures to your own logging system
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    And then committing the transaction later:

    TransactionSerializedContext serialized = TransactionSerializedContext.createFrom(encoded);
    
    try {
        TransactionResult result = transactions.commit(serialized);
    
    } catch (TransactionFailed e) {
        // System.err is used for example, log failures to your own logging system
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    Alternatively the transaction can be rolled back:

    TransactionSerializedContext serialized = TransactionSerializedContext.createFrom(encoded);
    
    try {
        TransactionResult result = transactions.rollback(serialized);
    
    } catch (TransactionFailed e) {
        // System.err is used for example, log failures to your own logging system
        System.err.println("Transaction did not reach commit point");
    
        for (LogDefer err : e.result().log().logs()) {
            System.err.println(err.toString());
        }
    }

    The transaction expiry timer (which is configurable) will begin ticking once the transaction starts, and is not paused while the transaction is in a deferred state.

    Further Reading