Skip to content

Latest commit

 

History

History
478 lines (353 loc) · 20.2 KB

DEVELOPER_GUIDE.md

File metadata and controls

478 lines (353 loc) · 20.2 KB
Raw

Developer Guide

So you want to contribute code to OpenSearch k-NN? Excellent! We're glad you're here. Here's what you need to do.

Getting Started

Fork OpenSearch k-NN Repo

Fork opensearch-project/OpenSearch k-NN and clone locally.

Example:

git clone https://github.com/[your username]/OpenSearch.git

Install Prerequisites

JDK 11

OpenSearch builds using Java 11 at a minimum. This means you must have a JDK 11 installed with the environment variable JAVA_HOME referencing the path to Java home for your JDK 11 installation, e.g. JAVA_HOME=/usr/lib/jvm/jdk-11.

One easy way to get Java 11 on *nix is to use sdkman.

curl -s "https://get.sdkman.io" | bash
source ~/.sdkman/bin/sdkman-init.sh
sdk install java 11.0.2-open
sdk use java 11.0.2-open

Team has to replace minimum JDK version 14 as it was not an LTS release. JDK 14 should still work for most scenarios. In addition to this, the plugin has been tested with JDK 17, and this JDK version is fully supported.

CMake

The plugin requires that cmake >= 3.24.0 is installed in order to build the JNI libraries.

One easy way to install on mac or linux is to use pip:

pip install cmake==3.24.0

On Mac M series machines, install cmake using:

brew install cmake

Faiss Dependencies

To build the faiss JNI library, you need to have openmp, lapack and blas installed. For more information on faiss dependencies, please refer to their documentation.

Openblas can be used for both lapack and blas. To install on Mac, run:

brew install openblas

Additionally, the gcc toolchain needs to be installed on Mac. To install, run:

brew install gcc

Additional setup for Mac M series Machines

The following commands enable running/building k-NN on M series machines:

// Go to k-NN folder
cd k-NN

// Build to generate the necessary files to be modified below (will fail)
./gradlew build

//Go to jni folder
cd jni

// File changes required
sed -i -e 's/\/usr\/local\/opt\/libomp\//\/opt\/homebrew\/opt\/llvm\//g' cmake/init-faiss.cmake
sed -i -e 's/__aarch64__/__undefine_aarch64__/g' external/faiss/faiss/utils/distances_simd.cpp
sed -i -e 's/pragma message WARN/pragma message /g' external/nmslib/similarity_search/src/distcomp_scalar.cc
// Change -mcpu value to use chip version according to your M series, for example, -mcpu=apple-m1. You can see the supported mcpu values via gcc --print-supported-cpus
sed -i -e 's/-march=native/-mcpu=apple-m1/g' external/nmslib/similarity_search/CMakeLists.txt
sed -i -e 's/-mcpu=apple-a14/-mcpu=apple-m1/g' external/nmslib/python_bindings/setup.py

// Install llvm
brew install llvm
echo 'export PATH="/opt/homebrew/opt/llvm/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc

// Set compiler path for CMAKE
export CC=/opt/homebrew/opt/llvm/bin/clang
export CXX=/opt/homebrew/opt/llvm/bin/clang++

// In case of linking issues with the external libraries and clang, you can try setting the CMAKE compiler to gcc/g++ instead through the following commands:
export CC=gcc
export CXX=g++
sed -i '' '/set(CMAKE_CXX_STANDARD_REQUIRED True)/a\'$'\n''set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xclang -fopenmp -L/opt/homebrew/opt/libomp/lib -I/opt/homebrew/opt/libomp/include -lomp -arch arm64 -fexceptions")' CMakeLists.txt

// Build
cmake . --fresh
make

Next, obtain a minimum distribution tarball of the k-NN version you want to build:

  1. Fork the OpenSearch Repo into your github account.
  2. Clone the repository locally
  3. Run the following commands: cd OpenSearch && ./gradlew -p distribution/archives/darwin-tar assemble
  4. You should see a opensearch-min--SNAPSHOT-darwin-x64.tar.gz file present in distribution/archives/darwin-tar/build/distributions/
  5. Build k-NN by passing the OpenSearch distribution path in ./gradlew <integTest/run> -PcustomDistributionUrl="<Full path to .tar.gz file you noted above>"

If you want to start OpenSearch directly on Mac M series, make sure to use JDK for ARM. Otherwise, you will see the following error: mach-o file, but is an incompatible architecture (have 'arm64', need 'x86_64'). It is better to start OpenSearch by running bash opensearch-tar-install.sh instead of ./bin/opensearch. To run ./bin/opensearch, the environment variable JAVA_LIBRARY_PATH needs to be set correctly so that OpenSearch can find the JNI library:

export OPENSEARCH_HOME=the directory of opensearch...
export JAVA_LIBRARY_PATH=$JAVA_LIBRARY_PATH:$OPENSEARCH_HOME/plugins/opensearch-knn/lib

Environment

Currently, the plugin only supports Linux on x64 and arm platforms.

Use an Editor

IntelliJ IDEA

When importing into IntelliJ you will need to define an appropriate JDK. The convention is that this SDK should be named "11", and the project import will detect it automatically. For more details on defining an SDK in IntelliJ please refer to this documentation. Note that SDK definitions are global, so you can add the JDK from any project, or after project import. Importing with a missing JDK will still work, IntelliJ will report a problem and will refuse to build until resolved.

You can import the OpenSearch project into IntelliJ IDEA as follows.

  1. Select File > Open
  2. In the subsequent dialog navigate to the root build.gradle file
  3. In the subsequent dialog select Open as Project

Java Language Formatting Guidelines

Taken from OpenSearch's guidelines:

Java files in the OpenSearch codebase are formatted with the Eclipse JDT formatter, using the Spotless Gradle plugin. The formatting check can be run explicitly with:

./gradlew spotlessJavaCheck

The code can be formatted with:

./gradlew spotlessApply

Please follow these formatting guidelines:

  • Java indent is 4 spaces
  • Line width is 140 characters
  • Lines of code surrounded by // tag::NAME and // end::NAME comments are included in the documentation and should only be 76 characters wide not counting leading indentation. Such regions of code are not formatted automatically as it is not possible to change the line length rule of the formatter for part of a file. Please format such sections sympathetically with the rest of the code, while keeping lines to maximum length of 76 characters.
  • Wildcard imports (import foo.bar.baz.*) are forbidden and will cause the build to fail.
  • If absolutely necessary, you can disable formatting for regions of code with the // tag::NAME and // end::NAME directives, but note that these are intended for use in documentation, so please make it clear what you have done, and only do this where the benefit clearly outweighs the decrease in consistency.
  • Note that JavaDoc and block comments i.e. /* ... */ are not formatted, but line comments i.e // ... are.
  • There is an implicit rule that negative boolean expressions should use the form foo == false instead of !foo for better readability of the code. While this isn't strictly enforced, it might get called out in PR reviews as something to change.

Build

OpenSearch k-NN uses a Gradle wrapper for its build. Run gradlew on Unix systems.

Tests use JAVA11_HOME environment variable, make sure to add it in the export path else the tests might fail. e.g

echo "export JAVA11_HOME=<JDK11 path>" >> ~/.zshrc
source ~/.zshrc

Build OpenSearch k-NN using gradlew build

./gradlew build

For Mac M series machines use

./gradlew build -PcustomDistributionUrl="<Full path to .tar.gz file file you noted above>"

JNI Library

The plugin relies on 2 JNI libraries to perform approximate k-NN search. ./gradlew build will first build the libraries before running the plugins tests. If you see errors related to library linkage failure, make sure all libraries are in the Java library path.

To build the JNI Library manually, follow these steps:

cd jni
cmake .

# To build everything, including tests. If your computer has multiple cores you can speed it up by building in parallel using make -j 2 (or a higher number for more parallelism)
make

# To just build the libraries
make opensearchknn_faiss opensearchknn_nmslib

The libraries will be placed in the jni/release directory.

Our JNI uses Google Tests for the C++ unit testing framework. To run the tests, run:

# To run all tests
./bin/jni_test

# To run nmslib tests
./bin/jni_test --gtest_filter='Nmslib*'

# To run faiss tests
./bin/jni_test --gtest_filter='Faiss*'

JNI Library Artifacts

We build and distribute binary library artifacts with OpenSearch. We build the library binaries in this script. In it, we package the libraries together with an openmp shared object. For blas and lapack, we statically link them into the faiss library. We use Centos 7 with g++ 4.8.5 to build. Additionally, in order to provide as much general compatibility as possible, we compile the libraries without some of the optimized instruction sets. For users that want to get the most out of the libraries, they should follow this section and build the libraries from source in their production environment, so that if their environment has optimized instruction sets, they take advantage of them.

Custom patch on JNI Library

If you want to make a custom patch on JNI library

  1. Make a change on top of current version of JNI library and push the commit locally.
  2. Create a patch file for the change using git format-patch -o patches HEAD^
  3. Place the patch file under jni/patches
  4. Make a change in jni/CmakeLists.txt, .github/workflows/CI.yml to apply the patch during build

By default, in the cmake build system, these patches will be applied and committed to the native libraries. In order to successfully make the commits the user.name and user.email git configurations need to be setup. If you cannot set these in your environment, you can disable committing the changes to the library by passing gradle this flag: build.lib.commit_patches=false. For example, gradlew build -Dbuild.lib.commit_patches=false. If the patches are not committed, then the full library build process will run each time cmake is invoked. In a development environment, it is recommended to setup the user git configuration to avoid this cost.

Parallelize make

When we are building the plugin for the first time, it takes some time to build the JNI libraries. We can parallelize make and speed up the build time by setting and passing this flag to gradle, nproc.count if your computer has more number of cores (greater than or equal to 2).

# While building OpenSearch k-NN
./gradlew build -Dnproc.count=4

# While running OpenSearch k-NN
./gradlew run -Dnproc.count=4

# When building the JNI library manually
cd jni
cmake .
# Pass the processor count with make using `-j`
make -j 4

Enable SIMD Optimization

SIMD(Single Instruction/Multiple Data) Optimization is enabled by default on Linux and Mac which boosts the performance by enabling AVX2 on x86 architecture and NEON on ARM64 architecture while building the Faiss library. But to enable SIMD, the underlying processor should support this (AVX2 or NEON). It can be disabled by setting the parameter simd.enabled to false. As of now, it is not supported on Windows OS.

# While building OpenSearch k-NN
./gradlew build -Dsimd.enabled=true

# While running OpenSearch k-NN
./gradlew run -Dsimd.enabled=true

# While building the JNI libraries
cd jni
cmake . -DSIMD_ENABLED=true

Run OpenSearch k-NN

Run Single-node Cluster Locally

Run OpenSearch k-NN using gradlew run. For Mac M series add -PcustomDistributionUrl= argument.

./gradlew run

That will build OpenSearch and start it, writing its log above Gradle's status message. We log a lot of stuff on startup, specifically these lines tell you that plugin is ready.

[2020-05-29T14:50:35,167][INFO ][o.e.h.AbstractHttpServerTransport] [runTask-0] publish_address {127.0.0.1:9200}, bound_addresses {[::1]:9200}, {127.0.0.1:9200}
[2020-05-29T14:50:35,169][INFO ][o.e.n.Node               ] [runTask-0] started

It's typically easier to wait until the console stops scrolling, and then run curl in another window to check if OpenSearch instance is running.

curl localhost:9200

{
  "name" : "runTask-0",
  "cluster_name" : "runTask",
  "cluster_uuid" : "oX_S6cxGSgOr_mNnUxO6yQ",
  "version" : {
    "number" : "1.0.0-SNAPSHOT",
    "build_type" : "tar",
    "build_hash" : "0ba0e7cc26060f964fcbf6ee45bae53b3a9941d0",
    "build_date" : "2021-04-16T19:45:44.248303Z",
    "build_snapshot" : true,
    "lucene_version" : "8.7.0",
    "minimum_wire_compatibility_version" : "6.8.0",
    "minimum_index_compatibility_version" : "6.0.0-beta1"
  }
}

Additionally, it is also possible to run a cluster with security enabled:

./gradlew run -Dsecurity.enabled=true -Dhttps=true -Duser=admin -Dpassword=<admin-password>

Then, to access the cluster, we can run

curl https://localhost:9200 --insecure -u admin:<admin-password>

{
  "name" : "integTest-0",
  "cluster_name" : "integTest",
  "cluster_uuid" : "kLsNk4JDTMyp1yQRqog-3g",
  "version" : {
    "distribution" : "opensearch",
    "number" : "3.0.0-SNAPSHOT",
    "build_type" : "tar",
    "build_hash" : "9d85e566894ef53e5f2093618b3d455e4d0a04ce",
    "build_date" : "2023-10-30T18:34:06.996519Z",
    "build_snapshot" : true,
    "lucene_version" : "9.8.0",
    "minimum_wire_compatibility_version" : "2.12.0",
    "minimum_index_compatibility_version" : "2.0.0"
  },
  "tagline" : "The OpenSearch Project: https://opensearch.org/"
}

Run Multi-node Cluster Locally

It can be useful to test and debug on a multi-node cluster. In order to launch a 3 node cluster with the KNN plugin installed, run the following command:

./gradlew run -PnumNodes=3

In order to run the integration tests, run this command:

./gradlew :integTest -PnumNodes=3

Additionally, to run integration tests with security enabled, run

./gradlew :integTest -Dsecurity.enabled=true -PnumNodes=3

Integration tests can be run with remote cluster. For that run the following command and replace host/port/cluster name values with ones for the target cluster:

./gradlew :integTestRemote -Dtests.rest.cluster=localhost:9200 -Dtests.cluster=localhost:9200 -Dtests.clustername="integTest-0" -Dhttps=false -PnumNodes=1

In case remote cluster is secured it's possible to pass username and password with the following command:

./gradlew :integTestRemote -Dtests.rest.cluster=localhost:9200 -Dtests.cluster=localhost:9200 -Dtests.clustername="integTest-0" -Dhttps=true -Duser=admin -Dpassword=<admin-password>

Debugging

Sometimes it is useful to attach a debugger to either the OpenSearch cluster or the integration test runner to see what's going on. For running unit tests, hit Debug from the IDE's gutter to debug the tests. For the OpenSearch cluster, first, make sure that the debugger is listening on port 5005. Then, to debug the cluster code, run:

./gradlew :integTest -Dcluster.debug=1 # to start a cluster with debugger and run integ tests

OR

./gradlew run --debug-jvm # to just start a cluster that can be debugged

The OpenSearch server JVM will connect to a debugger attached to localhost:5005 before starting. If there are multiple nodes, the servers will connect to debuggers listening on ports 5005, 5006, ...

To debug code running in an integration test (which exercises the server from a separate JVM), first, setup a remote debugger listening on port 8000, and then run:

./gradlew :integTest -Dtest.debug=1

The test runner JVM will connect to a debugger attached to localhost:8000 before running the tests.

Additionally, it is possible to attach one debugger to the cluster JVM and another debugger to the test runner. First, make sure one debugger is listening on port 5005 and the other is listening on port 8000. Then, run:

./gradlew :integTest -Dtest.debug=1 -Dcluster.debug=1

Backwards Compatibility Testing

The purpose of Backwards Compatibility Testing and different types of BWC tests are explained here

Use these commands to run BWC tests for k-NN:

  1. Rolling upgrade tests: ./gradlew :qa:rolling-upgrade:testRollingUpgrade
  2. Full restart upgrade tests: ./gradlew :qa:restart-upgrade:testRestartUpgrade
  3. ./gradlew :qa:bwcTestSuite is used to run all the above bwc tests together.

Use this command to run BWC tests for a given Backwards Compatibility Version:

./gradlew :qa:bwcTestSuite -Dbwc.version=1.0.0

Here, we are testing BWC Tests with BWC version of plugin as 1.0.0.

Adding new tests

Before adding any new tests to Backward Compatibility Tests, we should be aware that the tests in BWC are not independent. While creating an index, a test cannot use the same index name if it is already used in other tests. Also, adding extra operations to the existing test may impact other existing tests like graphCount.

Codec Versioning

Starting from 2.0 release the new versioning for codec has been introduced. Two positions will be used to define the version, in format 'X.Y', where 'X' corresponds to underlying version of Lucene and 'Y' is the version of the format. Please note that Lucene version along with corresponding Lucene codec is part of the core OpenSearch. KNN codec should be in sync with Lucene codec version from core OpenSearch.

Codec version is used in following classes and methods:

  • org.opensearch.knn.index.codec.KNNXYCodec.KNNXYCodec
  • org.opensearch.knn.index.codec.KNNXYCodec.KNNXYPerFieldKnnVectorsFormat
  • org.opensearch.knn.index.codec.KNNCodecVersion

These classes and methods are tied directly to Lucene version represented by 'X' part. Other classes use the delegate pattern so no direct tie to Lucene version are related to format and represented by 'Y'

  • BinaryDocValues
  • CompoundFormat
  • DocValuesConsumer
  • DocValuesReader

Version '910' is going to be the first such new version. It corresponds to Lucene 9.1 that is used by the underlying OpenSearch 2.0 and initial version of the format classes. If in future we need to adjust something in format logic, we only increment the 'Y' part and version became '911'.

Submitting Changes

See CONTRIBUTING.

Backports

The Github workflow in backport.yml creates backport PRs automatically when the original PR with an appropriate label backport <backport-branch-name> is merged to main with the backport workflow run successfully on the PR. For example, if a PR on main needs to be backported to 1.x branch, add a label backport 1.x to the PR and make sure the backport workflow runs on the PR along with other checks. Once this PR is merged to main, the workflow will create a backport PR to the 1.x branch.