We follow the Google Java Style Guide. See https://github.com/google/google-java-format for IDE plugins. The rules are not configurable.
The build will fail if the code is not formatted. To format all files from the command line, run:
mvn fmt:format
Some aspects are not covered by the formatter: braces must be used with if, else, for, do
and while statements, even when the body is empty or contains only a single statement.
The build will fail if XML files are not formatted correctly. Run the following command before you commit:
mvn xml-format:xml-formatThe formatter does not enforce a maximum line length, but please try to keep it below 100 characters to keep files readable across all mediums (IDE, terminal, Github...).
Similarly, enforce a right margin of 100 characters in those files. Editors and IDEs generally have a way to configure this (for IDEA, install the "Wrap to column" plugin).
Do not use static imports. They make things harder to understand when you look at the code someplace where you don't have IDE support, like Github's code view.
Avoid abbreviations in class and variable names. A good rule of thumb is that you should only use
them if you would also do so verbally, for example "id" and "config" are probably reasonable.
Single-letter variables are permissible if the variable scope is only a few lines, or for commonly
understood cases (like i for a loop index).
Keep source files short. Short files are easy to understand and test. The average should probably be around 200-300 lines.
All types in "API" packages must be documented. For "internal" packages, documentation is optional, but in no way discouraged: it's generally a good idea to have a class-level comment that explains where the component fits in the architecture, and anything else that you feel is important.
You don't need to document every parameter or return type, or even every method. Don't document something if it is completely obvious, we don't want to end up with this:
/**
* Returns the name.
*
* @return the name
*/
String getName();On the other hand, there is often something useful to say about a method, so most should have at least a one-line comment. Use common sense.
Driver users coding in their IDE should find the right documentation at the right time. Try to think of how they will come into contact with the class. For example, if a type is constructed with a builder, each builder method should probably explain what the default is when you don't call it.
Avoid using too many links, they can make comments harder to read, especially in the IDE. Link to a
type the first time it's mentioned, then use a text description ("this registry"...) or an @code
block. Don't link to a class in its own documentation. Don't link to types that appear right below
in the documented item's signature.
/**
* @return this {@link Builder} <-- completely unnecessary
*/
Builder withLimit(int limit) {We use SLF4J; loggers are declared like this:
private static final Logger LOG = LoggerFactory.getLogger(TheEnclosingClass.class);Logs are intended for two personae:
- Ops who manage the application in production.
- Developers (maybe you) who debug a particular issue.
The first 3 log levels are for ops:
ERROR: something that renders the driver -- or a part of it -- completely unusable. An action is required to fix it: bouncing the client, applying a patch, etc.WARN: something that the driver can recover from automatically, but indicates a configuration or programming error that should be addressed. For example: the driver connected successfully, but one of the contact points in the configuration was malformed; the same prepared statement is being prepared multiple time by the application code.INFO: something that is part of the normal operation of the driver, but might be useful to know for an operator. For example: the driver has initialized successfully and is ready to process queries; an optional dependency was detected in the classpath and activated an enhanced feature.
Do not log errors that are rethrown to the client (such as the error that you're going to complete a request with). This is annoying for ops because they see a lot of stack traces that require no actual action on their part, because they're already handled by application code.
Similarly, do not log stack traces for non-critical errors. If you still want the option to get the
trace for debugging, see the Loggers.warnWithException utility.
The last 2 levels are for developers, to help follow what the driver is doing from a "black box" perspective (think about debugging an issue remotely, and all you have are the logs).
TRACE: anything that happens for every user request. Not only request handling, but all related components (e.g. timestamp generators, policies, etc).DEBUG: everything else. For example, node state changes, control connection activity, etc.
Note that DEBUG and TRACE can coexist within the same component, for example the LBP
initializing is a one-time event, but returning a query plan is a per-request event.
Logs statements start with a prefix that identifies its origin, for example:
- for components that are unique to the cluster instance, just the cluster name:
[c0]. - for sessions, the cluster name + a generated unique identifier:
[c0|s0]. - for channel pools, the session identifier + the address of the node:
[c0|s0|/127.0.0.2:9042]. - for channels, the identifier of the owner (session or control connection) + the Netty identifier,
which indicates the local and remote ports:
[c0|s0|id: 0xf9ef0b15, L:/127.0.0.1:51482 - R:/127.0.0.1:9042]. - for request handlers, the session identifier, a unique identifier, and the index of the
speculative execution:
[c0|s0|1077199500|0].
Tests run with the configuration defined in src/test/resources/logback-test.xml. The default level
for driver classes is WARN, but you can override it with a system property: -DdriverLevel=DEBUG.
A nice setup is to use DEBUG when you run from your IDE, and keep the default for the command
line.
When you add or review new code, take a moment to run the tests in DEBUG mode and check if the
output looks good.
The java.util.stream API is often used (abused?) as a "functional API for collections":
List<Integer> sizes = words.stream().map(String::length).collect(Collectors.toList());The perceived advantages of this approach over traditional for-loops are debatable:
- readability: this is highly subjective. But consider the following:
- everyone can read for-loops, whether they are familiar with the Stream API or not. The opposite is not true.
- the stream API does not spell out all the details: what kind of list does
Collectors.toList()return? Is it pre-sized? Mutable? Thread-safe? - the stream API looks pretty on simple examples, but things can get ugly fast. Try rewriting
NetworkTopologyReplicationStrategywith streams.
- concision: this is irrelevant. When we look at code we care about maintainability, not how many keystrokes the author saved. The for-loop version of the above example is just 5 lines long, and your brain doesn't take longer to parse it.
The bottom line: don't try to "be functional" at all cost. Plain old for-loops are often just as simple.
Only use toString() for debug logs or exception messages, and always assume that its format is
unspecified and can change at any time.
If you need a specific string representation for a class, make it a dedicated method with a
documented format, for example toCqlLiteral. Otherwise it's too easy to lose track of the intended
usage and break things: for example, someone modifies your toString() method to make their logs
prettier, but unintentionally breaks the script export feature that expected it to produce CQL
literals.
toString() can delegate to toCqlLiteral() if that is appropriate for logs.
We use the JCIP annotations to document thread-safety policies.
Add them for all new code, with the exception of:
- enums and interfaces;
- utility classes (only static methods);
- test code.
Make sure you import the types from net.jcip, there are homonyms in the classpath.
We use the Spotbugs annotations to document nullability of parameters, method return types and class members.
Please annotate any new class or interface with the appropriate annotations: @NonNull, @Nullable. Make sure you import
the types from edu.umd.cs.findbugs.annotations, there are homonyms in the classpath.
Static imports are permitted in a couple of places:
- All AssertJ methods, e.g.:
assertThat(node.getDatacenter()).isNotNull(); fail("Expecting IllegalStateException to be thrown");
- All Mockito methods, e.g.:
when(codecRegistry.codecFor(DataTypes.INT)).thenReturn(codec); verify(codec).decodePrimitive(any(ByteBuffer.class), eq(ProtocolVersion.DEFAULT));
- All Awaitility methods, e.g.:
await().until(() -> somethingBecomesTrue());
Test methods names use lower snake case, generally start with should, and clearly indicate the
purpose of the test, for example: should_fail_if_key_already_exists. If you have trouble coming
up with a simple name, it might be a sign that your test does too much, and should be split.
We use AssertJ (assertThat) for assertions. Don't use JUnit assertions (assertEquals,
assertNull, etc).
Don't try to generify at all cost: a bit of duplication is acceptable, if that helps keep the tests simple to understand (a newcomer should be able to understand how to fix a failing test without having to read too much code).
Test classes can be a bit longer, since they often enumerate similar test cases. You can also
factor some common code in a parent abstract class named with "XxxTestBase", and then split
different families of tests into separate child classes. For example, CqlRequestHandlerTestBase,
CqlRequestHandlerRetryTest, CqlRequestHandlerSpeculativeExecutionTest...
They live in the same module as the code they are testing. They should be fast and not start any external process. They usually target one specific component and mock the rest of the driver context.
They live in the integration-tests module, and exercise the whole driver stack against an external
process, which can be either one of:
-
Simulacron: simulates Cassandra nodes on loopback addresses; your test must "prime" data, i.e. tell the nodes what results to return for pre-determined queries.
For an example of a Simulacron-based test, see
NodeTargetingIT. -
CCM: launches actual Cassandra nodes locally. The
ccmexecutable must be in the path.You can pass a
-Dccm.versionsystem property to the build to target a particular Cassandra version (it defaults to 3.11.0).-Dccm.directoryallows you to point to a local installation -- this can be a checkout of the Cassandra codebase, as long as it's built. SeeCcmBridgein the driver codebase for more details.For an example of a CCM-based test, see
PlainTextAuthProviderIT.
Integration tests are divided into three categories:
These tests can be run in parallel, to speed up the build. They either use:
- dedicated Simulacron instances. These are lightweight, and Simulacron will manage the ports to make sure that there are no collisions.
- a shared, one-node CCM cluster. Each test works in its own keyspace.
The build runs them with a configurable degree of parallelism (currently 8). The shared CCM cluster
is initialized the first time it's used, and stopped before moving on to serial tests. Note that we
run with parallel=classes, which means methods within the same class never run concurrent to each
other.
To make an integration test parallelizable, annotate it with @Category(ParallelizableTests.class).
If you use CCM, it must be with CcmRule.
For an example of a Simulacron-based parallelizable test, see NodeTargetingIT. For a CCM-based
test, see DirectCompressionIT.
These tests cannot run in parallel, in general because they require CCM clusters of different sizes, or with a specific configuration (we never run more than one CCM cluster simultaneously: it would be too resource-intensive, and too complicated to manage all the ports).
The build runs them one by one, after the parallelizable tests.
To make an integration test serial, do not annotate it with @Category. The CCM rule must be
CustomCcmRule.
For an example, see DefaultLoadBalancingPolicyIT.
Note: if multiple serial tests have a common "base" class, do not pull up CustomCcmRule, each
child class must have its own instance. Otherwise they share the same CCM instance, and the first
one destroys it on teardown. See TokenITBase for how to organize code in those cases.
Not only can those tests not run in parallel, they also require specific environment tweaks, typically system properties that need to be set before initialization.
The build runs them one by one, each in its own JVM fork, after the serial tests.
To isolate an integration test, annotate it with @Category(IsolatedTests.class). The CCM rule
must be CustomCcmRule.
For an example, see HeapCompressionIT.
Do not mix CcmRule and SimulacronRule in the same test. It makes things harder to follow, and
can be inefficient (if the SimulacronRule is method-level, it will create a Simulacron cluster for
every test method, even those that only need CCM).
Rules annotated with @ClassRule wrap the whole test class, and are reused across methods. Try to
use this as much as possible, as it's more efficient. The fields need to be static; also make them
final and use constant naming conventions, like CCM_RULE.
When you use a server rule (CcmRule or SimulacronRule) and a SessionRule at the same level,
wrap them into a rule chain to ensure proper initialization order:
private static final CcmRule CCM_RULE = CcmRule.getInstance();
private static final SessionRule<CqlSession> SESSION_RULE = SessionRule.builder(CCM_RULE).build();
@ClassRule
public static final TestRule CHAIN = RuleChain.outerRule(CCM_RULE).around(SESSION_RULE);Rules annotated with @Rule wrap each test method. Use lower-camel case for field names:
private CcmRule ccmRule = CcmRule.getInstance();
private SessionRule<CqlSession> sessionRule = SessionRule.builder(ccmRule).build();
@ClassRule
public TestRule chain = RuleChain.outerRule(ccmRule).around(sessionRule);Only use this for:
- CCM tests that use
@CassandraRequirementor@DseRequirementrestrictions at the method level (ex:BatchStatementIT). - tests where you really need to restart from a clean state for every method.
It's also possible to use a @ClassRule for CCM / Simulacron, and a @Rule for the session rule.
In that case, you don't need to use a rule chain.
mvn clean test
This currently takes about 30 seconds. The goal is to keep it within a couple of minutes (it runs for each commit if you enable the pre-commit hook -- see below).
mvn clean verify
This currently takes about 9 minutes. We don't have a hard limit, but ideally it should stay within 30 minutes to 1 hour.
You can skip test categories individually with -DskipParallelizableITs, -DskipSerialITs and
-DskipIsolatedITs (-DskipITs still works to skip them all at once).
Simulacron (used in integration tests) relies on loopback aliases to simulate multiple nodes. On Linux or Windows, you shouldn't have anything to do. On MacOS, run this script:
#!/bin/bash
for sub in {0..4}; do
echo "Opening for 127.0.$sub"
for i in {0..255}; do sudo ifconfig lo0 alias 127.0.$sub.$i up; done
done
Note that this is known to cause temporary increased CPU usage in OS X initially while mDNSResponder acclimates itself to the presence of added IP addresses. This lasts several minutes. Also, this does not survive reboots.
The build will fail if some license headers are missing. To update all files from the command line, run:
mvn license:format
Ensure pre-commit.sh is executable, then run:
ln -s ../../pre-commit.sh .git/hooks/pre-commit
This will only allow commits if the tests pass. It is also a good reminder to keep the test suite short.
Note: the tests run on the current state of the working directory. I tried to add a git stash in
the script to only test what's actually being committed, but I couldn't get it to run reliably
(it's still in there but commented). Keep this in mind when you commit, and don't forget to re-add
the changes if the first attempt failed and you fixed the tests.
If you need to install something in your local repository quickly, you can use the fast profile to
skip all "non-essential" checks (licenses, formatting, tests, etc):
mvn clean install -Pfast
You can speed things up even more by targeting specific modules with the -pl option:
mvn clean install -Pfast -pl core,query-builder,mapper-runtime,mapper-processor,bom
Please run the normal build at least once before you push your changes.
Keep your changes focused. Each commit should have a single, clear purpose expressed in its message.
Resist the urge to "fix" cosmetic issues (add/remove blank lines, move methods, etc.) in existing code. This adds cognitive load for reviewers, who have to figure out which changes are relevant to the actual issue. If you see legitimate issues, like typos, address them in a separate commit (it's fine to group multiple typo fixes in a single commit).
Isolate trivial refactorings into separate commits. For example, a method rename that affects dozens of call sites can be reviewed in a few seconds, but if it's part of a larger diff it gets mixed up with more complex changes (that might affect the same lines), and reviewers have to check every line.
Commit message subjects start with a capital letter, use the imperative form and do not end with a period:
- correct: "Add test for CQL request handler"
- incorrect: "
Added test for CQL request handler" - incorrect: "
New test for CQL request handler"
Avoid catch-all messages like "Minor cleanup", "Various fixes", etc. They don't provide any useful information to reviewers, and might be a sign that your commit contains unrelated changes.
We don't enforce a particular subject line length limit, but try to keep it short.
You can add more details after the subject line, separated by a blank line. The following pattern (inspired by Netty) is not mandatory, but welcome for complex changes:
One line description of your change
Motivation:
Explain here the context, and why you're making that change.
What is the problem you're trying to solve.
Modifications:
Describe the modifications you've done.
Result:
After your change, what will change.
Like commits, pull requests should be focused on a single, clearly stated goal.
Don't base a pull request onto another one, it's too complicated to follow two branches that evolve at the same time. If a ticket depends on another, wait for the first one to be merged.
If you have to address feedback, avoid rewriting the history (e.g. squashing or amending commits): this makes the reviewers' job harder, because they have to re-read the full diff and figure out where your new changes are. Instead, push a new commit on top of the existing history; it will be squashed later when the PR gets merged. If the history is complex, it's a good idea to indicate in the message where the changes should be squashed:
* 20c88f4 - Address feedback (to squash with "Add metadata parsing logic") (36 minutes ago)
* 7044739 - Fix various typos in Javadocs (2 days ago)
* 574dd08 - Add metadata parsing logic (2 days ago)
(Note that the message refers to the other commit's subject line, not the SHA-1. This way it's still relevant if there are intermediary rebases.)
If you need new stuff from the base branch, it's fine to rebase and force-push, as long as you don't rewrite the history. Just give a heads up to the reviewers beforehand. Don't push a merge commit to a pull request.