Skip to content

Latest commit

 

History

History
659 lines (503 loc) · 26.6 KB

DEVELOPING.md

File metadata and controls

659 lines (503 loc) · 26.6 KB
Raw

Developing

Setup

We're super excited to have you interested in working on Vector! Before you start you should pick how you want to develop.

For small or first-time contributions, we recommend the Docker method. Prefer to do it yourself? That's fine too!

Using a Docker or Podman environment

Targets: You can use this method to produce AARCH64, Arm6/7, as well as x86/64 Linux builds.

Since not everyone has a full working native environment, we took our environment and stuffed it into a Docker (or Podman) container!

This is ideal for users who want it to "Just work" and just want to start contributing. It's also what we use for our CI, so you know if it breaks we can't do anything else until we fix it. 😉

Before you go further, install Docker or Podman through your official package manager, or from the Docker or Podman sites.

# Optional: Only if you use `podman`
export CONTAINER_TOOL="podman"

If your Linux environment runs SELinux in Enforcing mode, you will need to relabel the vector source code checkout with container_home_t context. Otherwise, the container environment cannot read/write the code:

cd your/checkout/of/vector/
sudo semanage fcontext -a "${PWD}(/.*)?" -t container_file_t
sudo restorecon . -R

By default, make environment style tasks will do a docker pull from GitHub's container repository, you can optionally build your own environment while you make your morning coffee ☕:

# Optional: Only if you want to go make a coffee
make environment-prepare

Now that you have your coffee, you can enter the shell!

# Enter a shell with optimized mounts for interactive processes.
# Inside here, you can use Vector like you have full toolchain (See below!)
make environment
# Try out a specific container tool. (Docker/Podman)
make environment CONTAINER_TOOL="podman"
# Add extra cli opts
make environment CLI_OPTS="--publish 3000:2000"

Now you can use the jobs detailed in "Bring your own toolbox" below.

Want to run from outside of the environment? Clever. Good thinking. You can run any of the following:

# Validate your code can compile
make check ENVIRONMENT=true
# Validate your code actually does compile (in dev mode)
make build-dev ENVIRONMENT=true
# Validate your test pass
make test SCOPE="sources::example" ENVIRONMENT=true
# Validate tests (that do not require other services) pass
make test ENVIRONMENT=true
# Validate your tests pass (starting required services in Docker)
make test-integration SCOPE="sources::example" ENVIRONMENT=true
# Validate your tests pass against a live service.
make test-integration SCOPE="sources::example" AUTOSPAWN=false ENVIRONMENT=true
# Validate all tests pass (starting required services in Docker)
make test-integration ENVIRONMENT=true
# Run your benchmarks
make bench SCOPE="transforms::example" ENVIRONMENT=true
# Format your code before pushing!
make fmt ENVIRONMENT=true

We use explicit environment opt-in as many contributors choose to keep their Rust toolchain local.

Bring your own toolbox

Targets: This option is required for MSVC/Mac/FreeBSD toolchains. It can be used to build for any environment or OS.

To build Vector on your own host will require a fairly complete development environment!

Loosely, you'll need the following:

  • To build Vector: Have working Rustup, Protobuf tools, C++/C build tools (LLVM, GCC, or MSVC), Python, and Perl, make (the GNU one preferably), bash, cmake, GNU coreutils, and autotools.
  • To run integration tests: Have docker available, or a real live version of that service. (Use AUTOSPAWN=false)
  • To run make check-component-features: Have remarshal installed.

If you find yourself needing to run something inside the Docker environment described above, that's totally fine, they won't collide or hurt each other. In this case, you'd just run make environment-generate.

We're interested in reducing our dependencies if simple options exist. Got an idea? Try it out, we'd to hear of your successes and failures!

In order to do your development on Vector, you'll primarily use a few commands, such as cargo and make tasks you can use ordered from most to least frequently run:

# Validate your code can compile
cargo check
make check
# Validate your code actually does compile (in dev mode)
cargo build
make build-dev
# Validate your test pass
cargo test sources::example
make test scope="sources::example"
# Validate tests (that do not require other services) pass
cargo test
make test
# Validate your tests pass (starting required services in Docker)
make test-integration scope="sources::example"
# Validate your tests pass against a live service.
make test-integration scope="sources::example" autospawn=false
cargo test --features docker sources::example
# Validate all tests pass (starting required services in Docker)
make test-integration
# Run your benchmarks
make bench scope="transforms::example"
cargo bench transforms::example
# Format your code before pushing!
make fmt
cargo fmt

If you run make you'll see a full list of all our tasks. Some of these will start Docker containers, sign commits, or even make releases. These are not common development commands and your mileage may vary.

The basics

Directory structure

  • /.github - GitHub & CI related configuration.
  • /benches - Internal benchmarks.
  • /config - Public facing Vector config, included in releases.
  • /distribution - Distribution artifacts for various targets.
  • /docs - Internal documentation for Vector contributors.
  • /lib - External libraries that do not depend on vector but are used within the project.
  • /proto - Protobuf definitions.
  • /rfcs - Previous Vector proposals, a great place to build context on previous decisions.
  • /scripts - Scripts used to generate docs and maintain the repo.
  • /src - Vector source.
  • /tests - Various high-level test cases.
  • /website - Vector's website and external documentation for Vector users.

Makefile

Vector includes a Makefile in the root of the repo. This serves as a high-level interface for common commands. Running make will produce a list of make targets with descriptions. These targets will be referenced throughout this document.

Code style

We use rustfmt on stable to format our code and CI will verify that your code follows this format style. To run the following command make sure rustfmt has been installed on the stable toolchain locally.

# To install rustfmt
rustup component add rustfmt

# To format the code
make fmt

Logging style

  • Always use the Tracing crate's key/value style for log events.
  • Events should be capitalized and end with a period, ..
  • Never use e or err - always spell out error to enrich logs and make it clear what the output is.
  • Prefer Display over Debug, %error and not ?error.

Nope!

warn!("Failed to merge value: {}.", err);

Yep!

warn!(message = "Failed to merge value.", %error);

Feature flags

When a new component (a source, transform, or sink) is added, it has to be put behind a feature flag with the corresponding name. This ensures that it is possible to customize Vector builds. See the features section in Cargo.toml for examples.

In addition, during development of a particular component it is useful to disable all other components to speed up compilation. For example, it is possible to build and run tests only for console sink using

cargo test --lib --no-default-features --features sinks-console sinks::console

In case if the tests are already built and only the component file changed, it is around 4 times faster than rebuilding tests with all features.

Dependencies

Dependencies should be carefully selected and avoided if possible. You can see how dependencies are reviewed in the Reviewing guide.

If a dependency is required only by one or multiple components, but not by Vector's core, make it optional and add it to the list of dependencies of the features corresponding to these components in Cargo.toml.

Guidelines

Sink healthchecks

Sinks may implement a health check as a means for validating their configuration against the environment and external systems. Ideally, this allows the system to inform users of problems such as insufficient credentials, unreachable endpoints, non-existent tables, etc. They're not perfect, however, since it's impossible to exhaustively check for issues that may happen at runtime.

When implementing health checks, we prefer false positives to false negatives. This means we would prefer that a health check pass and the sink then fail than to have the health check fail when the sink would have been able to run successfully.

A common cause of false negatives in health checks is performing an operation that the sink itself does not need. For example, listing all the available S3 buckets and checking that the configured bucket is on that list. The S3 sink doesn't need the ability to list all buckets, and a user that knows that may not have permitted it to do so. In that case, the health check will fail due to bad credentials even through its credentials are sufficient for normal operation.

This leads to a general strategy of mimicking what the sink itself does. Unfortunately, the fact that health checks don't have real events available to them leads to some limitations here. The most obvious example of this is with sinks where the exact target of a write depends on the value of some field in the event (e.g. an interpolated Kinesis stream name). It also pops up for sinks where incoming events are expected to conform to a specific schema. In both cases, random test data is reasonably likely to trigger a potential false-negative result. Even in simpler cases, we need to think about the effects of writing test data and whether the user would find that surprising or invasive. The answer usually depends on the system we're interfacing with.

In some cases, like the Kinesis example above, the right thing to do might be nothing at all. If we require dynamic information to figure out what entity (i.e. Kinesis stream in this case) that we're even dealing with, odds are very low that we'll be able to come up with a way to meaningfully validate that it's in working order. It's perfectly valid to have a health check that falls back to doing nothing when there is a data dependency like this.

With all that in mind, here is a simple checklist to go over when writing a new health check:

  • Does this check perform different fallible operations from the sink itself?
  • Does this check have side effects the user would consider undesirable (e.g. data pollution)?
  • Are there situations where this check would fail but the sink would operate normally?

Not all the answers need to be a hard "no", but we should think about the likelihood that any "yes" would lead to false negatives and balance that against the usefulness of the check as a whole for finding problems. Because we have the option to disable individual health checks, there's an escape hatch for users that fall into a false negative circumstance. Our goal should be to minimize the likelihood of users needing to pull that lever while still making a good effort to detect common problems.

Testing

Testing is very important since Vector's primary design principle is reliability. You can read more about how Vector tests in our testing blog post.

Unit tests

Unit tests refer to the majority of inline tests throughout Vector's code. A defining characteristic of unit tests is that they do not require external services to run, therefore they should be much quicker. You can run them with:

cargo test

Integration tests

Integration tests verify that Vector actually works with the services it integrates with. Unlike unit tests, integration tests require external services to run. A few rules when setting up integration tests:

  • To ensure all contributors can run integration tests, the service must run in a Docker container.
  • The service must be configured on a unique port that is configured through an environment variable.
  • Add a test-integration-<name> to Vector's Makefile and ensure that it starts the service before running the integration test.
  • Add the name of your integration to the include matrix of the test-integration job to Vector's .github/workflows/integration-test.yml workflow.

Once complete, you can run your integration tests with:

make test-integration-<name>

Blackbox tests

Vector also offers blackbox testing via Vector's test harness. This is a complex testing suite that tests Vector's performance in real-world environments. It is typically used for benchmarking, but also correctness testing.

You can run these tests within a PR as described in the CI section.

Tips and tricks

Faster builds With sccache

Vector is a large project with a plethora of dependencies. Changing to a different branch, or running cargo clean, can sometimes necessitate rebuilding many of those dependencies, which has an impact on productivity. One way to reduce some of this cycle time is to use sccache, which caches compilation assets to avoid recompiling them over and over.

sccache works by being configured to sit in front of rustc, taking compilation requests from Cargo and checking the cache to see if it already has the cached compilation unit. It handles making sure that different compiler flags, versions of Rust, etc., are taken into consideration before using a cached asset.

In order to use sccache, you must first install it. There are pre-built binaries for all major platforms to get you going quickly. The usage documentation also explains how to set up your environment to actually use it. We recommend using the $HOME/.cargo/config approach as this can help speed up all of your Rust development work, and not just developing on Vector.

While sccache was originally designed to cache compilation assets in cloud storage, maximizing reusability amongst CI workers, sccache actually supports storing assets locally by default. Local mode works well for local development as it is much easier to delete the cache directory if you ever encounter issues with the cached assets. It also involves no extra infrastructure or spending.

Testing specific components

If you are developing a particular component and want to quickly iterate on unit tests related only to this component, the following approach can reduce waiting times:

  1. Install cargo-watch.

  2. (Only for GNU/Linux) Install LLVM 9 (for example, package llvm-9 on Debian) and set RUSTFLAGS environment variable to use lld as the linker:

    export RUSTFLAGS='-Clinker=clang-9 -Clink-arg=-fuse-ld=lld'

  3. Run in the root directory of Vector's source

    cargo watch -s clear -s \
  'cargo test --lib --no-default-features --features=<component type>-<component id> <component type>::<component id>'

    For example, if the component is reduce transform, the command above turns into

    cargo watch -s clear -s \
  'cargo test --lib --no-default-features --features=transforms-reduce transforms::reduce'

Generating sample logs

We use flog to build a sample set of log files to test sending logs from a file. This can be done with the following commands on Mac with homebrew. Installation instruction for flog can be found here.

flog --bytes $((100 * 1024 * 1024)) > sample.log

This will create a 100MiB sample log file in the sample.log file.

Benchmarking

All benchmarks are placed in the /benches folder. You can run benchmarks via the make bench command. In addition, Vector maintains a full test harness for complex end-to-end integration and performance testing.

Profiling

If you're trying to improve Vector's performance (or understand why your change made it worse), profiling is a useful tool for seeing where time is being spent.

While there are a bunch of useful profiling tools, a simple place to get started is with Linux's perf. Before getting started, you'll likely need to give yourself access to collect stats:

echo -1 | sudo tee /proc/sys/kernel/perf_event_paranoid

You'll also want to edit Cargo.toml and make sure that Vector is being built with debug symbols in release mode. This ensures that you'll get human-readable info in the eventual output:

[profile.release]
debug = true

Then you can start up a release build of Vector with whatever config you're interested in profiling.

cargo run --release -- --config my_test_config.toml

Once it's started, use the ps tool (or equivalent) to make a note of its PID. We'll use this to tell perf which process we would like it to collect data about.

The next step is somewhat dependent on the config you're testing. For this example, let's assume you're using a simple TCP-mode socket source listening on port 9000. Let's also assume that you have a large file of example input in access.log (you can use a tool like flog to generate this).

With all that prepared, we can send our test input to Vector and collect data while it is under load:

perf record -F99 --call-graph dwarf -p $VECTOR_PID socat -dd OPEN:access.log TCP:localhost:9000

This instructs perf to collect data from our already-running Vector process for the duration of the socat command. The -F argument is the frequency at which perf should sample the Vector call stack. Higher frequencies will collect more data and produce more detailed output, but can produce enormous amounts of data that take a very long time to process. Using -F99 works well when your input data is large enough to take a minute or more to process, but feel free to adjust both input size and sampling frequency for your setup.

It's worth noting that this is not the normal way to profile programs with perf. Usually you would simply run something like perf record my_program and not have to worry about PIDs and such. We differ from this because we're only interested in data about what Vector is doing while under load. Running it directly under perf would collect data for the entire lifetime of the process, including startup, shutdown, and idle time. By telling perf to collect data only while the load generation command is running we get a more focused dataset and don't have to worry about timing different commands in quick succession.

You'll now find a perf.data file in your current directory with all the information that was collected. There are different ways to process this, but one of the most useful is to create a flamegraph. For this we can use the inferno tool (available via cargo install):

perf script | inferno-collapse-perf > stacks.folded
cat stacks.folded | inferno-flamegraph > flamegraph.svg

And that's it! You now have a flamegraph SVG file that can be opened and navigated in your favorite web browser.

Domains

This section contains domain specific development knowledge for various areas of Vector. You should scan this section for any relevant domains for your development area.

Kubernetes

Architecture

The Kubernetes integration architecture is largely inspired by the RFC 2221, so this is a concise outline of the effective design, rather than a deep dive into the concepts.

The operation logic

With kubernetes_logs source, Vector connects to the Kubernetes API doing a streaming watch request over the Pods executing on the same Node that Vector itself runs at. Once Vector gets the list of all the Pods that are running on the Node, it starts collecting logs for the logs files corresponding to each of the Pod. Only plaintext (as in non-gzipped) files are taken into consideration. The log files are then parsed into events, and the said events are annotated with the metadata from the corresponding Pods, correlated via the file path of the originating log file. The events are then passed to the topology.

Where to find things

We use custom Kubernetes API client and machinery, that lives at src/kubernetes. The kubernetes_logs source lives at src/sources/kubernetes_logs. There is also an end-to-end (E2E) test framework that resides at lib/k8s-test-framework, and the actual end-to-end tests using that framework are at lib/k8s-e2e-tests.

The Kubernetes-related distribution bit that are at distribution/docker, distribution/kubernetes and our Helm chart can be found at vectordotdev/helm-charts.

The development assistance resources are located at Tiltfile and in the tilt dir.

Development

There is a special flow for when you develop portions of Vector that are designed to work with Kubernetes, like kubernetes_logs source or the deployment/kubernetes/*.yaml configs.

This flow facilitates building Vector and deploying it into a cluster.

Requirements

There are some extra requirements besides what you'd normally need to work on Vector:

Automatic

You can use tilt to detect changes, rebuild your image, and update your Kubernetes resource. Simply start your local Kubernetes cluster and run tilt up from Vector's root dir.

Testing

Integration tests

The Kubernetes integration tests have a lot of parts that can go wrong.

To cope with the complexity and ensure we maintain high quality, we use E2E (end-to-end) tests.

E2E tests normally run at CI, so there's typically no need to run them manually.

Requirements
  • kubernetes cluster (minikube has special support, but any cluster should work)
  • docker
  • kubectl
  • bash
  • cross - cargo install cross
  • helm

Vector release artifacts are prepared for E2E tests, so the ability to do that is required too, see Vector docs for more details.

Notes:

  • minikube had a bug in the versions 1.12.x that affected our test process - see kubernetes/minikube#8799. Use version 1.13.0+ that has this bug fixed.
  • minikube has troubles running on ZFS systems. If you're using ZFS, we suggest using a cloud cluster or minik8s with local registry.
  • E2E tests expect to have enough resources to perform a full Vector build, usually 8GB of RAM with 2CPUs are sufficient to successfully complete E2E tests locally.
Tutorial

To run the E2E tests, use the following command:

CONTAINER_IMAGE_REPO=<your name>/vector-test make test-e2e-kubernetes

Where CONTAINER_IMAGE_REPO is the docker image repo name to use, without part after the :. Replace <your name> with your Docker Hub username.

You can also pass additional parameters to adjust the behavior of the test:

  • QUICK_BUILD=true - use development build and an image from the dev flow instead of a production docker image. Significantly speeds up the preparation process, but doesn't guarantee the correctness in the release build. Useful for development of the tests or Vector code to speed up the iteration cycles.

  • USE_MINIKUBE_CACHE=true - instead of pushing the built docker image to the registry under the specified name, directly load the image into a minikube-controlled cluster node. Requires you to test against a minikube cluster. Eliminates the need to have a registry to run tests. When USE_MINIKUBE_CACHE=true is set, we provide a default value for the CONTAINER_IMAGE_REPO so it can be omitted. Can be set to auto (default) to automatically detect whether to use minikube cache or not, based on the current kubectl context. To opt-out, set USE_MINIKUBE_CACHE=false.

  • CONTAINER_IMAGE=<your name>/vector-test:tag - completely skip the step of building the Vector docker image, and use the specified image instead. Useful to speed up the iterations speed when you already have a Vector docker image you want to test against.

  • SKIP_CONTAINER_IMAGE_PUBLISHING=true - completely skip the image publishing step. Useful when you want to speed up the iteration speed and when you know the Vector image you want to test is already available to the cluster you're testing against.

  • SCOPE - pass a filter to the cargo test command to filter out the tests, effectively equivalent to cargo test -- $SCOPE.

Passing additional commands is done like so:

QUICK_BUILD=true USE_MINIKUBE_CACHE=true make test-e2e-kubernetes

or

QUICK_BUILD=true CONTAINER_IMAGE_REPO=<your name>/vector-test make test-e2e-kubernetes