Client Testing

Testing Error Handling with Buggify

FoundationDB clients need to handle errors correctly. Wrong error handling can lead to many bugs - in the worst case it can lead to a corrupted database. Because of this it is important that an application or layer author tests properly their application during failure scenarios. But this is non-trivial. In a development environment cluster failures are very unlikely and it is therefore possible that certain types of exceptions are never tested in a controlled environment.

The simplest way of testing for these kind of errors is a simple mechanism called Buggify. If this option is enabled in the client, the client will randomly throw errors that an application might see in a production environment. Enable this option in testing will greatly improve the probability that error handling is tested properly.

Options to Control Buggify

There are four network options to control the buggify behavior. By default, buggify is disabled (as it will behave in a way that is not desirable in a production environment). The options to control buggify are:

  • buggify_enable This option takes no argument and will enable buggify.

  • buggify_disable This can be used to disable buggify again.

  • client_buggify_section_activated_probability (default 25) A number between 0 and 100.

  • client_buggify_section_fired_probability (default 25) A number between 0 and 100.

The way buggify works is by enabling sections in the code first that get only executed with a certain probability. Generally these code sections will simply introduce a synthetic error.

When a section is passed for the first time, the client library will decide randomly whether that code section will be enabled or not. It will be enabled with a probability of client_buggify_section_activated_probability.

Whenever the client executes a buggify-enabled code-block, it will randomly execute it. This is to make sure that a certain exception doesn’t always fire. The probably for executing such a section is client_buggify_section_fired_probability.

Simulation and Cluster Workloads

FoundationDB comes with its own testing framework. Tests are implemented as workloads. A workload is nothing more than a class that gets called by server processes running the tester role. Additionally, a fdbserver process can run a simulator that simulates a full fdb cluster with several machines and different configurations in one process. This simulator can run the same workloads you can run on a real cluster. It will also inject random failures like network partitions and disk failures.

This tutorial explains how one can implement a workload, how one can orchestrate a workload on a cluster with multiple clients, and how one can run a workload within a simulator. Running in a simulator is also useful as it does not require any setup: you can simply run one command that will provide you with a fully functional FoundationDB cluster.

General Overview

Workloads in FoundationDB are generally compiled into the binary. However, FoundationDB also provides the ability to load workloads dynamically. This is done through dlopen (on Unix like operating systems) or LoadLibrary (on Windows).

Parallelism and Determinism

A workload can run either in a simulation or on a real cluster. In simulation, fdbserver will simulate a whole cluster and will use a deterministic random number generator to simulate random behavior and random failures. This random number generator is initialized with a random seed. In case of a test failure, the user can reuse the given seed and rerun the same test in order to further observe and debug the behavior.

However, this will only work as long as the workload doesn’t introduce any non-deterministic behavior. One example of non-deterministic behavior is the running multiple threads.

The workload is created in the main network thread and it will run in the main network thread. Because of this, using any blocking function (for example blockUntilReady on a future object) will result in a deadlock. Using the callback API is therefore required if one wants to keep the simulator’s deterministic behavior.

For existing applications and layers, however, not using the blocking API might not be an option. For these use-cases, a user can chose to start new threads and use the blocking API from within these threads. This will mean that test failures will be non-deterministic and might be hard to reproduce.

To start a new thread, one has to “bind” operating system threads to their simulated processes. This can be done by setting the ProcessId in the child threads when they get created. In Java this is done by only starting new threads through the provided Executor. In the C++ API one can use the FDBWorkloadContext to do that. For example:

template<class Fun>
std::thread startThread(FDBWorkloadContext* context, Fun fun) {
    auto processId = context->getProcessID();
    return std::thread([context, processID, fun](
        context->setProcessID(processID);
        fun();
    ));
}

Finding the Shared Object

When the test starts, fdbserver needs to find the shared object to load. The name of this shared object has to be provided.

For Java, we provide an implementation in libjava_workloads.so which can be built out of the sources. The tester will look for the key libraryName in the test file which should be the name of the library without extension and without the lib prefix (so java_workloads if you want to write a Java workload).

By default, the process will look for the library in the directory ../shared/foundationdb/ relative to the location of the fdbserver binary. If the library is somewhere else on the system, one can provide the absolute path to the library (only the folder, not the file name) in the test file with the libraryPath option.

Implementing a C++ Workload

In order to implement a workload, one has to build a shared library that links against the fdb client library. This library has to exppse a function (with C linkage) called workloadFactory which needs to return a pointer to an object of type FDBWorkloadFactory. This mechanism allows the other to implement as many workloads within one library as she wants. To do this the pure virtual classes FDBWorkloadFactory and FDBWorkload have to be implemented.

FDBWorkloadFactory *workloadFactory(FDBLogger*)

This function has to be defined within the shared library and will be called by fdbserver for looking up a specific workload. FDBLogger will be passed and is guaranteed to survive for the lifetime of the process. This class can be used to write to the FoundationDB traces. Logging anything with severity FDBSeverity::Error will result in a hard test failure. This function needs to have c-linkage, so define it in a extern "C" block.

std::shared_ptr<FDBWorkload> FDBWorkload::create(const std::string &name)

This is the only method to be implemented in FDBWorkloadFactory. If the test file contains a key-value pair workloadName the value will be passed to this method (empty string otherwise). This way, a library author can implement many workloads in one library and use the test file to chose which one to run (or run multiple workloads either concurrently or serially).

std::string FDBWorkload::description() const

This method has to return the name of the workload. This can be a static name and is primarily used for tracing.

bool FDBWorkload::init(FDBWorkloadContext *context)

Right after initialization

void FDBWorkload::setup(FDBDatabase *db, GenericPromise<bool> done)

This method will be called by the tester during the setup phase. It should be used to populate the database.

void FDBWorkload::start(FDBDatabase *db, GenericPromise<bool> done)

This method should run the actual test.

void FDBWorkload::check(FDBDatabase *db, GenericPromise<bool> done)

When the tester completes, this method will be called. A workload should run any consistency/correctness tests during this phase.

void FDBWorkload::getMetrics(std::vector<FDBPerfMetric> &out) const

If a workload collects metrics (like latencies or throughput numbers), these should be reported back here. The multitester (or test orchestrator) will collect all metrics from all test clients and it will aggregate them.

Implementing a Java Workload

In order to implement your own workload in Java you can simply create an implementation of the abstract class AbstractWorkload. A minimal implementation will look like this:

package my.package;
import com.apple.foundationdb.testing.Promise;
import com.apple.foundationdb.testing.AbstractWorkload;
import com.apple.foundationdb.testing.WorkloadContext;

class MinimalWorkload extends AbstractWorkload {
    public MinimalWorkload(WorkloadContext ctx) {
        super(ctx);
    }

    @Override
    public void setup(Database db, Promise promise) {
        log(20, "WorkloadSetup", null);
        promise.send(true);
    }

    @Override
    public void start(Database db) {
        log(20, "WorkloadStarted", null);
        promise.send(true);
    }

    @Override
    public boolean check(Database db) {
        log(20, "WorkloadFailureCheck", null);
        promise.send(true);
    }
}

The lifecycle of a test will look like this:

  1. All testers will create an instance of the AbstractWorkload implementation.

  2. All testers will (in parallel but not guaranteed exactly at the same time) call setup and they will wait for all of them to finish. This phase can be used to pre-populate data.

  3. All tester will then call start (again, in parallel) and wait for all of them to finish.

  4. All testers will then call check on all testers and use the returned boolean to determine whether the test succeeded.

All these methods take a Database object as an argument. This object can be used to create and execute transactions against the cluster.

When implementing workloads, an author has to follow these rules:

  • To write tracing to the trace-files one should use AbstractWorkload.log. This Method takes three arguments: an integer for severity (5 means debug, 10 means log, 20 means warning, 30 means warn always, and 40 is a severe error). If any tester logs something of severity 40, the test run is considered to have failed.

  • In order to increase throughput on the cluster, an author might want to spawn several threads. However, threads MUST only be spawn through the Executor instance one can get from AbstractWorkload.getExecutor(). Otherwise, a simulation test will probably segfault. The reason for this is that we need to keep track of which simulated machine a thread corresponds to internally.

Within a workload you have access to the WorkloadContext which provides additional information about the current execution environment. The context can be accessed through this.context and provides the following methods:

  • String getOption(String name, String defaultValue). A user can provide parameters to workloads through a configuration file (explained further down). These parameters are provided to all clients through the context and can be accessed with this method.

  • int getClientId() and int getClientCount(). An author can determine how many clients are running in the cluster and each of those will get a globally unique ID (a number between 0 and clientCount - 1). This is useful for example if you want to generate transactions that are guaranteed to not conflict with transactions from other clients.

  • int getSharedRandomNumber(). At startup a random number will be generated. This will allow for generating the same random numbers across several machines if this number is used as a seed.

Running a Workload in the Simulator

We’ll first walk how one can run a workload in a simulator. FoundationDB comes already with a large number of workloads. But some of them can’t be run in simulation while other don’t work on a real cluster. Most will work on both though. To look for examples how these can be ran, you can find configuration files in the tests directory in the FoundationDB source tree.

We will now go through an example how you can write a relatively complex test and run it in the simulator. Writing and running tests in the simulator is a simple two-step process.

  1. Write the test.

  2. Run fdbserver in simulation mode and provide it with the test file.

Write the Test

A workload is not a test. A test is a simple test file that tells the test orchestrator which workloads it should run and in which order. Additionally one can provide parameters to workloads through this file.

A test file might look like this:

testTitle=MyTest
  testName=External
  libraryName=java_workloads
  workloadName=my.package.MinimalWorkload
  classPath=PATH_TO_JAR_OR_DIR_CONTAINING_WORKLOAD,OTHER_DEPENDENCIES

  testName=Attrition
  testDuration=5.0
  reboot=true
  machinesToKill=3

testTitle=AnotherTest
  testName=External
  libraryName=java_workloads
  workloadName=my.package.MinimalWorkload
  classPath=PATH_TO_JAR_OR_DIR_CONTAINING_WORKLOAD,OTHER_DEPENDENCIES
  someOption=foo

  testName=External
  libraryName=java_workloads
  workloadName=my.package.AnotherWorkload
  classPath=PATH_TO_JAR_OR_DIR_CONTAINING_WORKLOAD,OTHER_DEPENDENCIES
  anotherOption=foo

This test will do the following:

  1. First it will run MinimalWorkload without any parameter.

  2. After 5.0 seconds the simulator will reboot 3 random machines (this is what Attrition does and this workload is provided by FoundationDB. This is one of the few workloads that only work in the simulator).

  3. When all workloads are finished, it will run MinimalWorkload again. This time it will have the option someOption set to foo. Additionally it will run AnotherWorkload in parallel.

How to set the Class Path correctly

As you can see from above example, we can set the classpath through two different mechanisms. However, one has to be careful as they can’t be used interchangeably.

  • You can set a class path through the JVM argument -Djava.class.path=.... This is how you have to pass the path to the FoundationDB client library (as the client library is needed during the initialization phase). However, only the first specified section will have any effect as the other Workloads will run in the same VM (and arguments, by nature, can only be passed once).

  • The classPath option. This option will add all paths (directories or JAR-files) to the classPath of the JVM while it is running. Not being able to add the path will result in a test failure. This is useful to add different dependencies to different workloads. A path can appear more than once across sections. However, they must not conflict with each other as we never remove something from the classpath.

Run the simulator

This step is very simple. You can simply run fdbserver with role simulator and pass the test with -f:

fdbserver -r simulation -f testfile.txt

Running a Workload on an actual Cluster

Running a workload on a cluster works basically the smae way. However, one must actually setup a cluster first. This cluster must run between one and many server processes with the class test. So above 2-step process becomes a bit more complex:

  1. Write the test (same as above).

  2. Set up a cluster with as many test clients as you want.

  3. Run the orchestrator to actually execute the test.

Step 1. is explained further up. For step 2., please refer to the general FoundationDB configuration. The main difference to a normal FoundationDB cluster is that some processes must have a test class assigned to them. This can be done in the foundationdb.conf. For example this file would create a server with 8 processes of which 4 would act as test clients.

[fdbmonitor]
user = foundationdb
group = foundationdb

[general]
restart-delay = 60
cluster-file = /etc/foundationdb/fdb.cluster

## Default parameters for individual fdbserver processes
[fdbserver]
command = /usr/sbin/fdbserver
public-address = auto:$ID
listen-address = public
datadir = /var/lib/foundationdb/data/$ID
logdir = /var/log/foundationdb

[fdbserver.4500]
[fdbserver.4501]
[fdbserver.4502]
[fdbserver.4503]
[fdbserver.4510]
class = test
[fdbserver.4511]
class = test
[fdbserver.4512]
class = test
[fdbserver.4513]
class = test

Running the actual test can be done with fdbserver as well. For this you can call the process with the multitest role:

fdbserver -r multitest -f testfile.txt

This command will block until all tests are completed.