Transactions for Apache HBase™: Cask Tephra provides globally consistent transactions on top of Apache HBase. While HBase provides strong consistency with row- or region-level ACID operations, it sacrifices cross-region and cross-table consistency in favor of scalability. This trade-off requires application developers to handle the complexity of ensuring consistency when their modifications span region boundaries. By providing support for global transactions that span regions, tables, or multiple RPCs, Tephra simplifies application development on top of HBase, without a significant impact on performance or scalability for many workloads.
Tephra leverages HBase's native data versioning to provide multi-versioned concurrency control (MVCC) for transactional reads and writes. With MVCC capability, each transaction sees its own consistent "snapshot" of data, providing snapshot isolation of concurrent transactions.
Tephra consists of three main components:
- Transaction Server - maintains global view of transaction state, assigns new transaction IDs and performs conflict detection;
- Transaction Client - coordinates start, commit, and rollback of transactions; and
- TransactionProcessor Coprocessor - applies filtering to the data read (based on a given transaction's state) and cleans up any data from old (no longer visible) transactions.
A central transaction manager generates a globally unique, time-based transaction ID for each transaction that is started, and maintains the state of all in-progress and recently committed transactions for conflict detection. While multiple transaction server instances can be run concurrently for automatic failover, only one server instance is actively serving requests at a time. This is coordinated by performing leader election amongst the running instances through ZooKeeper. The active transaction server instance will also register itself using a service discovery interface in ZooKeeper, allowing clients to discover the currently active server instance without additional configuration.
A client makes a call to the active transaction server in order to start a new transaction. This
returns a new transaction instance to the client, with a unique transaction ID (used to identify
writes for the transaction), as well as a list of transaction IDs to exclude for reads (from
in-progress or invalidated transactions). When performing writes, the client overrides the
timestamp for all modified HBase cells with the transaction ID. When reading data from HBase, the
client skips cells associated with any of the excluded transaction IDs. The read exclusions are
applied through a server-side filter injected by the TransactionProcessor coprocessor.
The TransactionProcessor coprocessor is loaded on all HBase tables where transactional reads
and writes are performed. When clients read data, it coordinates the server-side filtering
performed based on the client transaction's snapshot. Data cells from any transactions that are
currently in-progress or those that have failed and could not be rolled back ("invalid"
transactions) will be skipped on these reads. In addition, the TransactionProcessor cleans
up any data versions that are no longer visible to any running transactions, either because the
transaction that the cell is associated with failed or a write from a newer transaction was
successfully committed to the same column.
More details on how Tephra transactions work and the interactions between these components can be found in our Transactions over HBase presentation.
Status of continuous integration build at Travis CI:
The latest JDK or JRE version 1.7.xx or 1.8.xx for Linux, Windows, or Mac OS X must be installed in your environment; we recommend the Oracle JDK.
To check the Java version installed, run the command:
$ java -version
Tephra is tested with the Oracle JDKs; it may work with other JDKs such as Open JDK, but it has not been tested with them.
Once you have installed the JDK, you'll need to set the JAVA_HOME environment variable.
Tephra requires a working HBase and HDFS environment in order to operate. Tephra supports these component versions:
| Component | Source | Supported Versions |
|---|---|---|
| HDFS | Apache Hadoop | 2.0.2-alpha through 2.6.0 |
| CDH or HDP | (CDH) 5.0.0 through 5.4.4 or (HDP) 2.0, 2.1, 2.2 or 2.3 | |
| MapR | 4.1 (with MapR-FS) | |
| HBase | Apache | 0.96.x, 0.98.x, 1.0.x, and 1.1.x |
| CDH or HDP | (CDH) 5.0.0 through 5.4.4 or (HDP) 2.0, 2.1, 2.2 or 2.3 | |
| MapR | 4.1 (with Apache HBase) | |
| Zookeeper | Apache | Version 3.4.3 through 3.4.5 |
| CDH or HDP | (CDH) 5.0.0 through 5.4.4 or (HDP) 2.0, 2.1, 2.2 or 2.3 | |
| MapR | 4.1 |
Note: Components versions shown in this table are those that we have tested and are confident of their suitability and compatibility. Later versions of components may work, but have not necessarily have been either tested or confirmed compatible.
You can get started with Tephra by building directly from the latest source code:
git clone https://github.com/caskdata/tephra.git cd tephra mvn clean package
After the build completes, you will have a full binary distribution of Tephra under the
tephra-distribution/target/ directory. Take the tephra-<version>.tar.gz file and install
it on your systems.
For any client applications, add the following dependencies to any Apache Maven POM files (or your build system's equivalent configuration), in order to make use of Tephra classes:
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-api</artifactId> <version>0.6.2</version> </dependency> <dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-core</artifactId> <version>0.6.2</version> </dependency>
Since the HBase APIs have changed between versions, you will need to select the appropriate HBase compatibility library.
For HBase 0.96.x:
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-hbase-compat-0.96</artifactId> <version>0.6.2</version> </dependency>
For HBase 0.98.x:
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-hbase-compat-0.98</artifactId> <version>0.6.2</version> </dependency>
For HBase 1.0.x:
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-hbase-compat-1.0</artifactId> <version>0.6.2</version> </dependency>
If you are running the CDH 5.4 version of HBase 1.0.x (this version contains API incompatibilities with Apache HBase 1.0.x):
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-hbase-compat-1.0-cdh</artifactId> <version>0.6.2</version> </dependency>
For HBase 1.1.x:
<dependency> <groupId>co.cask.tephra</groupId> <artifactId>tephra-hbase-compat-1.1</artifactId> <version>0.6.2</version> </dependency>
Tephra makes use of a central transaction server to assign unique transaction IDs for data modifications and to perform conflict detection. Only a single transaction server can actively handle client requests at a time, however, additional transaction server instances can be run simultaneously, providing automatic failover if the active server becomes unreachable.
The Tephra transaction server can be deployed on the same cluster nodes running the HBase HMaster
process. The transaction server requires that the HBase libraries be available on the server's
Java CLASSPATH.
The transaction server supports the following configuration properties. All configuration
properties can be added to the hbase-site.xml file on the server's CLASSPATH:
| Name | Default | Description |
|---|---|---|
data.tx.bind.port |
15165 | Port to bind to |
data.tx.bind.address |
0.0.0.0 | Server address to listen on |
data.tx.server.io.threads |
2 | Number of threads for socket IO |
data.tx.server.threads |
20 | Number of handler threads |
data.tx.timeout |
30 | Timeout for a transaction to complete (seconds) |
data.tx.long.timeout |
86400 | Timeout for a long running transaction to complete (seconds) |
data.tx.cleanup.interval |
10 | Frequency to check for timed out transactions (seconds) |
data.tx.snapshot.dir |
HDFS directory used to store snapshots of tx state | |
data.tx.snapshot.interval |
300 | Frequency to write new snapshots |
data.tx.snapshot.retain |
10 | Number of old transaction snapshots to retain |
data.tx.metrics.period |
60 | Frequency for metrics reporting (seconds) |
To run the Transaction server, execute the following command in your Tephra installation:
./bin/tephra start
Any environment-specific customizations can be made by editing the bin/tephra-env.sh script.
Since Tephra clients will be communicating with HBase, the HBase client libraries and the HBase cluster
configuration must be available on the client's Java CLASSPATH.
Client API usage is described in the Client APIs section.
The transaction service client supports the following configuration properties. All configuration
properties can be added to the hbase-site.xml file on the client's CLASSPATH:
| Name | Default | Description |
|---|---|---|
data.tx.client.timeout |
30000 | Client socket timeout (milliseconds) |
data.tx.client.provider |
pool | Client provider strategy:
Note that "thread-local" provider can have a resource leak if threads are recycled |
data.tx.client.count |
50 | Max number of clients for "pool" provider |
data.tx.client.obtain.timeout |
3000 | Timeout (milliseconds) to wait when obtaining clients from the "pool" provider |
data.tx.client.retry.strategy |
backoff | Client retry strategy: "backoff" for back off between attempts; "n-times" for fixed number of tries |
data.tx.client.retry.attempts |
2 | Number of times to retry ("n-times" strategy) |
data.tx.client.retry.backoff.initial |
100 | Initial sleep time ("backoff" strategy) |
data.tx.client.retry.backoff.factor |
4 | Multiplication factor for sleep time |
data.tx.client.retry.backoff.limit |
30000 | Exit when sleep time reaches this limit |
In addition to the transaction server, Tephra requires an HBase coprocessor to be installed on all tables where transactional reads and writes will be performed.
To configure the coprocessor on all HBase tables, add the following to hbase-site.xml.
For HBase 0.96:
<property> <name>hbase.coprocessor.region.classes</name> <value>co.cask.tephra.hbase96.coprocessor.TransactionProcessor</value> </property>
For HBase 0.98:
<property> <name>hbase.coprocessor.region.classes</name> <value>co.cask.tephra.hbase98.coprocessor.TransactionProcessor</value> </property>
For HBase 1.0:
<property> <name>hbase.coprocessor.region.classes</name> <value>co.cask.tephra.hbase10.coprocessor.TransactionProcessor</value> </property>
For the CDH 5.4 version of HBase 1.0:
<property> <name>hbase.coprocessor.region.classes</name> <value>co.cask.tephra.hbase10cdh.coprocessor.TransactionProcessor</value> </property>
For HBase 1.1:
<property> <name>hbase.coprocessor.region.classes</name> <value>co.cask.tephra.hbase11.coprocessor.TransactionProcessor</value> </property>
You may configure the TransactionProcessor to be loaded only on HBase tables that you will
be using for transaction reads and writes. However, you must ensure that the coprocessor is
available on all impacted tables in order for Tephra to function correctly.
Tephra overrides HBase cell timestamps with transaction IDs, and uses these transaction
IDs to filter out cells older than the TTL (Time-To-Live). Transaction IDs are at a higher
scale than cell timestamps. When a regular HBase table that has existing data is
converted to a transactional table, existing data may be filtered out during reads. To
allow reading of existing data from a transactional table, you will need to set the
property data.tx.read.pre.existing as true on the table's table descriptor.
Note that even without the property data.tx.read.pre.existing being set to true,
any existing data will not be removed during compactions. Existing data simply won't be
visible during reads.
Tephra ships with built-in support for reporting metrics via JMX and a log file, using the Dropwizard Metrics library.
To enable JMX reporting for metrics, you will need to enable JMX in the Java runtime
arguments. Edit the bin/tephra-env.sh script and uncomment the following lines, making any
desired changes to configuration for port used, SSL, and JMX authentication:
# export JMX_OPTS="-Dcom.sun.management.jmxremote.ssl=false -Dcom.sun.management.jmxremote.authenticate=false -Dcom.sun.management.jmxremote.port=13001" # export OPTS="$OPTS $JMX_OPTS"
To enable file-based reporting for metrics, edit the conf/logback.xml file and uncomment the
following section, replacing the FILE-PATH placeholder with a valid directory on the local
filesystem:
<appender name="METRICS" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>/FILE-PATH/metrics.log</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<fileNamePattern>metrics.log.%d{yyyy-MM-dd}</fileNamePattern>
<maxHistory>30</maxHistory>
</rollingPolicy>
<encoder>
<pattern>%d{ISO8601} %msg%n</pattern>
</encoder>
</appender>
<logger name="tephra-metrics" level="TRACE" additivity="false">
<appender-ref ref="METRICS" />
</logger>
The frequency of metrics reporting may be configured by setting the data.tx.metrics.period
configuration property to the report frequency in seconds.
The TransactionAwareHTable class implements HBase's HTableInterface, thus providing the same APIs
that a standard HBase HTable instance provides. Only certain operations are supported
transactionally. These are:
| Methods Supported In Transactions |
|---|
exists(Get get) |
exists(List<Get> gets) |
get(Get get) |
get(List<Get> gets) |
batch(List<? extends Row> actions, Object[] results) |
batch(List<? extends Row> actions) |
batchCallback(List<? extends Row> actions, Object[] results, Batch.Callback<R> callback) [0.96] |
batchCallback(List<? extends Row> actions, Batch.Callback<R> callback) [0.96] |
getScanner(byte[] family) |
getScanner(byte[] family, byte[] qualifier) |
put(Put put) |
put(List<Put> puts) |
delete(Delete delete) |
delete(List<Delete> deletes) |
Other operations are not supported transactionally and will throw an UnsupportedOperationException if invoked.
To allow use of these non-transactional operations, call setAllowNonTransactional(true). This
allows you to call the following methods non-transactionally:
| Methods Supported Outside of Transactions |
|---|
getRowOrBefore(byte[] row, byte[], family) |
checkAndPut(byte[] row, byte[] family, byte[] qualifier, byte[] value, Put put) |
checkAndDelete(byte[] row, byte[] family, byte[] qualifier, byte[] value, Delete delete) |
mutateRow(RowMutations rm) |
append(Append append) |
increment(Increment increment) |
incrementColumnValue(byte[] row, byte[] family, byte[] qualifier, long amount) |
incrementColumnValue(byte[] row, byte[] family, byte[] qualifier, long amount, Durability durability) |
incrementColumnValue(byte[] row, byte[] family, byte[] qualifier, long amount, boolean writeToWAL) |
Note that for batch operations, only certain supported operations (get, put, and delete)
are applied transactionally.
To use a TransactionalAwareHTable, you need an instance of TransactionContext.
TransactionContext provides the basic contract for client use of transactions. At each point
in the transaction lifecycle, it provides the necessary interactions with the Tephra Transaction
Server in order to start, commit, and rollback transactions. Basic usage of
TransactionContext is handled using the following pattern:
TransactionContext context = new TransactionContext(client, transactionAwareHTable);
try {
context.start();
transactionAwareHTable.put(new Put(Bytes.toBytes("row"));
// ...
context.finish();
} catch (TransactionFailureException e) {
context.abort();
}- First, a new transaction is started using
TransactionContext.start(). - Next, any data operations are performed within the context of the transaction.
- After data operations are complete,
TransactionContext.finish()is called to commit the transaction. - If an exception occurs,
TransactionContext.abort()can be called to rollback the transaction.
TransactionAwareHTable handles the details of performing data operations transactionally, and
implements the necessary hooks in order to commit and rollback the data changes (see
TransactionAware).
To demonstrate how you might use TransactionAwareHTables, below is a basic implementation of a
SecondaryIndexTable. This class encapsulates the usage of a TransactionContext and provides a simple interface
to a user:
/**
* A Transactional SecondaryIndexTable.
*/
public class SecondaryIndexTable {
private byte[] secondaryIndex;
private TransactionAwareHTable transactionAwareHTable;
private TransactionAwareHTable secondaryIndexTable;
private TransactionContext transactionContext;
private final TableName secondaryIndexTableName;
private static final byte[] secondaryIndexFamily =
Bytes.toBytes("secondaryIndexFamily");
private static final byte[] secondaryIndexQualifier = Bytes.toBytes('r');
private static final byte[] DELIMITER = new byte[] {0};
public SecondaryIndexTable(TransactionServiceClient transactionServiceClient,
HTable hTable, byte[] secondaryIndex) {
secondaryIndexTableName =
TableName.valueOf(hTable.getName().getNameAsString() + ".idx");
HTable secondaryIndexHTable = null;
HBaseAdmin hBaseAdmin = null;
try {
hBaseAdmin = new HBaseAdmin(hTable.getConfiguration());
if (!hBaseAdmin.tableExists(secondaryIndexTableName)) {
hBaseAdmin.createTable(new HTableDescriptor(secondaryIndexTableName));
}
secondaryIndexHTable = new HTable(hTable.getConfiguration(),
secondaryIndexTableName);
} catch (Exception e) {
Throwables.propagate(e);
} finally {
try {
hBaseAdmin.close();
} catch (Exception e) {
Throwables.propagate(e);
}
}
this.secondaryIndex = secondaryIndex;
this.transactionAwareHTable = new TransactionAwareHTable(hTable);
this.secondaryIndexTable = new TransactionAwareHTable(secondaryIndexHTable);
this.transactionContext = new TransactionContext(transactionServiceClient,
transactionAwareHTable,
secondaryIndexTable);
}
public Result get(Get get) throws IOException {
return get(Collections.singletonList(get))[0];
}
public Result[] get(List<Get> gets) throws IOException {
try {
transactionContext.start();
Result[] result = transactionAwareHTable.get(gets);
transactionContext.finish();
return result;
} catch (Exception e) {
try {
transactionContext.abort();
} catch (TransactionFailureException e1) {
throw new IOException("Could not rollback transaction", e1);
}
}
return null;
}
public Result[] getByIndex(byte[] value) throws IOException {
try {
transactionContext.start();
Scan scan = new Scan(value, Bytes.add(value, new byte[0]));
scan.addColumn(secondaryIndexFamily, secondaryIndexQualifier);
ResultScanner indexScanner = secondaryIndexTable.getScanner(scan);
ArrayList<Get> gets = new ArrayList<Get>();
for (Result result : indexScanner) {
for (Cell cell : result.listCells()) {
gets.add(new Get(cell.getValue()));
}
}
Result[] results = transactionAwareHTable.get(gets);
transactionContext.finish();
return results;
} catch (Exception e) {
try {
transactionContext.abort();
} catch (TransactionFailureException e1) {
throw new IOException("Could not rollback transaction", e1);
}
}
return null;
}
public void put(Put put) throws IOException {
put(Collections.singletonList(put));
}
public void put(List<Put> puts) throws IOException {
try {
transactionContext.start();
ArrayList<Put> secondaryIndexPuts = new ArrayList<Put>();
for (Put put : puts) {
List<Put> indexPuts = new ArrayList<Put>();
Set<Map.Entry<byte[], List<KeyValue>>> familyMap = put.getFamilyMap().entrySet();
for (Map.Entry<byte [], List<KeyValue>> family : familyMap) {
for (KeyValue value : family.getValue()) {
if (value.getQualifier().equals(secondaryIndex)) {
byte[] secondaryRow = Bytes.add(value.getQualifier(),
DELIMITER,
Bytes.add(value.getValue(),
DELIMITER,
value.getRow()));
Put indexPut = new Put(secondaryRow);
indexPut.add(secondaryIndexFamily, secondaryIndexQualifier, put.getRow());
indexPuts.add(indexPut);
}
}
}
secondaryIndexPuts.addAll(indexPuts);
}
transactionAwareHTable.put(puts);
secondaryIndexTable.put(secondaryIndexPuts);
transactionContext.finish();
} catch (Exception e) {
try {
transactionContext.abort();
} catch (TransactionFailureException e1) {
throw new IOException("Could not rollback transaction", e1);
}
}
}
}- Currently, column family
Deleteoperations are implemented by writing a cell with an empty qualifier (emptybyte[]) and empty value (emptybyte[]). This is done in place of native HBaseDeleteoperations so the delete marker can be rolled back in the event of a transaction failure -- normal HBaseDeleteoperations cannot be undone. However, this means that applications that store data in a column with an empty qualifier will not be able to store empty values, and will not be able to transactionally delete that column. - Column
Deleteoperations are implemented by writing a empty value (emptybyte[]) to the column. This means that applications will not be able to store empty values to columns. - Invalid transactions are not automatically cleared from the exclusion list. When a transaction is
invalidated, either from timing out or being invalidated by the client due to a failure to rollback
changes, its transaction ID is added to a list of excluded transactions. Data from invalidated
transactions will be dropped by the
TransactionProcessorcoprocessor on HBase region flush and compaction operations. Currently, however, transaction IDs can only be manually removed from the list of excluded transaction IDs, using theco.cask.tephra.TransactionAdmintool.
Interested in helping to improve Tephra? We welcome all contributions, whether in filing detailed bug reports, submitting pull requests for code changes and improvements, or by asking questions and assisting others on the mailing list.
Bugs and tasks are tracked in a public JIRA issue tracker.
Tephra User Group: [email protected]
The tephra-user mailing list is primarily for users using the product to develop applications. You can expect questions from users, release announcements, and any other discussions that we think will be helpful to the users.
Tephra Developer Group and Development Discussions: [email protected]
The tephra-dev mailing list is essentially for developers actively working on the product, and should be used for all our design, architecture and technical discussions moving forward. This mailing list will also receive all JIRA and GitHub notifications.
Have questions about how Tephra works, or need help using it? Drop by the #tephra
chat room on irc.freenode.net.
We have a simple pull-based development model with a consensus-building phase, similar to Apache's voting process. If you’d like to help make Tephra better by adding new features, enhancing existing features, or fixing bugs, here's how to do it:
- If you are planning a large change or contribution, discuss your plans on the
tephra-devmailing list first. This will help us understand your needs and best guide your solution in a way that fits the project. - Fork Tephra into your own GitHub repository.
- Create a topic branch with an appropriate name.
- Work on the code to your heart's content.
- Once you’re satisfied, create a pull request from your GitHub repo (it’s helpful if you fill in all of the description fields).
- After we review and accept your request, we’ll commit your code to the caskdata/tephra repository.
Thanks for helping to improve Tephra!
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this product except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
Cask, Cask Tephra and Tephra are trademarks of Cask Data, Inc. All rights reserved.
Apache, Apache HBase, and HBase are trademarks of The Apache Software Foundation. Used with permission. No endorsement by The Apache Software Foundation is implied by the use of these marks.