本节文档描述了如何在CockroachDB数据库中使用Camunda流程引擎。

CockroachDB是一个高度可扩展的SQL数据库，作为一个分布式系统运行。因此，与其他Camunda支持的数据库相比，它有不同的要求和行为。出于这个原因，我们调整了流程引擎的行为，并增加了一些额外的机制，以确保流程引擎能够在CockroachDB上正确运行。

连接到 CockroachDB

CockroachDB implements the PostgreSQL wire protocol for communication between a Java application, like the process engine, and the database. It is recommended to use a PostgreSQL JDBC driver version compatible with PostgreSQL 9.5+, i.e. versions 42.X.X of the JDBC driver (see the CockroachDB docs for more details).

CockroachDB 的并发冲突

Whenever the process engine detects a concurrency conflict between transactions, it reports an OptimisticLockingException (for more details please see this docs section), which is then handled internally, or reported to the user. Since CockroachDB implements the SERIALIZABLE transaction isolation level, the process engine’s optimistic locking mechanism does not apply. When CockroachDB is not able to serialize a concurrent transaction, it raises an error that causes the transaction to abort. The process engine converts this error into a CrdbTransactionRetryException, a sub-class of the OptimisticLockingException class.

As the SERIALIZABLE transaction isolation level is stricter than the READ COMMITTED level that the process engine usually operates on, concurrency conflicts may happen in more cases than with the other Camunda-supported databases.

These cases are handled in the following ways:

1. Concurrency conflicts that happen during invocations of regular API calls should be handled in the same way as handling an OptimisticLockingException. Users should catch the exception and retry the API invocation.
2. Concurrency conflicts that occur during job execution are handled by the job executor. Since the reported exception is an instance of the OptimisticLockingException, the job executor will retry the job execution transparently, without consuming the defined job retries.
3. Concurrency conflicts that are detected in a specific set of API calls, where retries made by users or the job executor are impossible, or hard to perform, use a CRDB-specific mechanism that makes transparent retries of these API calls. The set of API calls, as well as the CRDB-specific mechanism, are described in the section below.

自定义CockroachDB事务重试机制

As mentioned in the section above, when using CockroachDB, a selected set of API calls are transparently retried using a CRDB-specific mechanism:

• When performing a deployment of BPMN/CMMN/DMN models and associated resources with the following calls:
  • DeploymentBuilder#deploy()
  • DeploymentBuilder#deployWithResult()
• By building a process engine with ProcessEngineConfiguration#buildProcessEngine-- :
  • When the history level is determined.
  • When the history cleanup job is created.
  • When the installation id is set.
  • When the telemetry is configured.
• When manually triggering history cleanup with HistoryService#cleanUpHistoryAsync(boolean immediatelyDue) .
• When fetching external tasks:
  • With the External Tasks Rest API endpoint.
  • With the execution of the ExternalTaskService#fetchAndLock API calls.
• When acquiring jobs for execution by the job executor.

It is important to note that the CockroachDB retry mechanism is applied to API calls where an OptimisticLockingException is handled internally in the regular case. The mechanism is also applied to API calls that utilize a pessimistic lock, which is redundant and disabled when CockroachDB is used (please see the section bellow for more details).

The number of retries the mechanism is allowed to perform can be configured with the commandRetries configuration property (more details here). The default value is 0 and can be set to an appropriate value based on the use-case.

When the number of retries is exhausted or is set to 0, a CrdbTransactionRetryException will be thrown to the user. The API call will then need to be manually retried by the user.

流程引擎行为的差异

CockroachDB-specific behaviors with alternatives

This section covers the differences process engine behavior when it is used with CockroachDB versus usage with other Camunda-supported databases. It provides descriptions and recommendations on how to configure and use the process engine with CockroachDB.

Un-ignorable historic OptimisticLockingException

The process engine may generate large amounts of historical data, and provides the history cleanup feature to ensure that “old” data is removed. The History Cleanup Removal-Time-based Strategy allows for historical data associated with still running Process Instances to be cleaned up. Since running Process Instances continue generating historical data, removing the same data in parallel is viewed as an OptimisticLockingException.

In the regular case, this historic OptimisticLockingException is handled internally and ignored since it is expected (and can be controlled by the skipHistoryOptimisticLockingExceptions configuration property). However, due to the SERIALIZABLE transaction isolation, CockroachDB detects this concurrency conflict by itself and requires for a complete transaction rollback and retry.

Consequently, when CockroachDB is used, the process engine will not attempt to ignore the reported CrdbTransactionRetryException, but will report it to the user so that the associated business logic may be retried.

Pessimistic locks replacement “behavior”

As mentioned above, the CockroachDB-specific retry mechanism in the process engine replaces the usage of pessimistic locks in the following API calls:

• When performing a deployment of BPMN/CMMN/DMN models and associated resources with the following calls:
  • DeploymentBuilder#deploy()
  • DeploymentBuilder#deployWithResult()
• By building a process engine with ProcessEngineConfiguration#buildProcessEngine() :
  • When the history level is determined.
  • When the history cleanup job is created.
  • When the installation id is set.
  • When the telemetry is configured.
• When manually triggering history cleanup with HistoryService#cleanUpHistoryAsync(boolean immediatelyDue) .

The role of the pessimistic locks in these API calls is not to prevent concurrent modification on the locked data, but to serve as barriers (or flags) to ensure the serialized, single execution of a given action (e.g. multiple deployments, multiple process engine startups, etc). Since CockroachDB ensures a serialized execution through its implementation of the SERIALIZABLE transaction isolation level, the pessimistic locks are redundant, or even counter-productive, as they introduce an additional waiting time overhead for each concurrent transaction that attempts to acquire the pessimistic lock.

When a process engine is started, it will invoke the commands to determine the history level, configure the history cleanup job, as well as set an installation id and telemetry status. All of these invocations had their pessimistic locks replaced with the CockroachDB-specific retry mechanism.

As a result of this replacement, multiple process engine startups will take longer to finish, as each concurrent startup may experience a command failure and perform a retry instead of waiting to acquire a pessimistic lock. The number of retries is defined by the commandRetries configuration property described in the CockroachDB retry mechanism section. Once the retries are exhausted, the startup will have to be retried manually by the caller (user). The same scenario applies when performing multiple concurrent deployments.

Using External Transaction management with the Spring/Java EE integrations

The Spring/Java EE/Quarkus Transaction integrations enable developers to manage transactions through the respective framework. This means that the process engine doesn’t control when transactions are started, committed, or rolled back.

When using CockroachDB, the retry mechanism works as expected as long as engine-managed transactions are used. However, in the case of externally managed transactions, this mechanism will not work correctly.

In this case, we suggest to disable the retry mechanism by setting the commandRetries configuration property (documented here) to 0. This will allow the CrdbTransactionRetryException to bubble up to the Spring/Java EE layer, where the transaction rollback and retry can be correctly implemented by the application developers.

CockroachDB-specific behaviors without alternatives

This section covers the differences in behavior between CockroachDB and other Camunda-supported databases that impact how the process engine operates. As opposed to the differences described in the previous section, this section describes more general CockroachDB characteristics, for which an alternative is not available, or more difficult to implement.

Unsupported Exclusive Message correlation

Exclusive Message correlation (see Javadocs ) prevents multiple concurrent message correlations to be performed on a single Process execution. When using the process engine with CockroachDB, it is not possible to use Exclusive Message correlation.

Unsupported Engine-side Query timeout

The jdbcStatementTimeout configuration property (more details here) enables users to set the amount of time in seconds that the JDBC driver will wait for a response from the database. CockroachDB currently doesn’t support query cancellation (CockroachDB issue), so the process engine jdbcStatementTimeout property will not have any effect when used on CockroachDB.

Server-side transaction timeout of 5 minutes

A common feature of databases is to introduce a server-side timeout on long-running transactions. In CockroachDB, this timeout is tied to the access time interval (lease) that CockroachDB nodes have on the data included in the transaction. Transactions running for more than 5 minutes may experience a server-side timeout with the error RETRY_COMMIT_DEADLINE_EXCEEDED.

In cases like these, it is advised to shorten the execution time of the business logic associated with the transaction since setting a custom server-side transaction timeout in CockroachDB is currently not possible (CockroachDB issue).

Blocking Transactions on Database READ

Due to the SERIALIZABLE transaction isolation level, when a transaction attempts to READ data which is currently modified by a concurrent transaction, the former will block until the latter is finished (committed). In the context of the process engine, certain operations may take longer than expected. For example, attempting to view (READ) the runtime state of a Process Instance in Cockpit, while the JobExecutor is executing a Job associated with that Process Instance, will block the Cockpit transaction until the Job is completed.

Concurrency conflicts on unrelated data from the same table

The section on concurrency conflicts in CockroachDB describes how this database handles concurrent conflicts on the same data. In addition, the currently supported CockroachDB version (v20.1.3), may view concurrent modifications of unrelated data of the same table as a conflict (CockroachDB issue).

In the process engine, this may happen when a Process Instance is deleted by one transaction, while a second transaction performs a modification on another, unrelated Process Instance. This results in a CrdbTransactionRetryException , and the modification of the latter Process Instance will have to be retired.