Commit graph

6 commits

Author SHA1 Message Date
jianglai
b19563f69d Use Maven repository on GCS for Cloud Build
This CL changes the Cloud Build flows to retrieve dependencies from our self-hosted GCS repository, to ensure that the release build are reproducible and hermetic (Note that it is still not truely reproducible as the dependency publishing process will override any existing artifacts in GCS with the current artifacts in Maven central. This is an issue that we should fix later).

There are a couple of changes involved to get this working:

1. Changed internal repo location to pull from the new repo.

2. Remove jcenter repo. It is only used to pull in the docker gradle plugin, which is not used. We instead build the deploy jar file with Gradle and build the docker image with a Dockerfile. The docker gradle plugin artifacts uploaded to GCS cannot be read because it is using some special classifier which seems to not be preserved when uploading. The java application plugin is also removed because it is only used by the docker gradle plugin.

3. Removed netty tcnative library classifier. It does not appear to be actually used (the jar downloaded from Maven central is an uber jar) and the classifier again interferes with downloading the artifacts from GCS.

4. Removed the cyclic dependency of the util project on itself. It was added because the nebula linter wanted it, which I think is an erroneous warning which should be reported upstream. The cyclic dependency was not a problem before (for yet unknown reasons), but it seems like when we force the dependency resolution (by calling project.generateDependencyPublications during configuration stage) it exacerbated the hidden issue and caused a cyclic task dependency in the util project, which is fatal. Now Nebula will complain again, but the warning is considered benign and will not cause the build to fail.

5. Added the nebula dependency lock files. We need these files when using the GCS maven repo because the we only upload artifacts after conflict resolution to GCS. If both v1 and v2 of the same library are requested in the dependency graph, only one will be uploaded. If we do not have the lock files in place, when building from GCS maven repo, Gradle will try to first find both v1 and v2 in the repo (which fails because v1 is not present in the repo), before proceeding to select v2 to use.

6. Refactored the code to upload Maven artifacts to GCS. We need to manually edit the POM file to reproduce the dependencies for each artifact so that they are all put in the classpath during compilation. Before, the POM files do not have any dependency information, which causes compilation to fail because transitive dependencies are not loaded (even though they are present in the GCS repo).

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=233408051
2019-02-11 11:24:12 -05:00
shicong
4866955c76 Unify dependency declarations into one place
This change unified declarations of third party dependencies into
one dependencies.gradle file by reference to this blog
https://medium.com/freelancer-engineering/managing-dependencies-in-multi-project-builds-with-gradle-7626d9c6448d

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=228713862
2019-01-10 16:23:35 -05:00
shicong
296acf80bb Add build logic to upload dependency JARs
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=225539482
2018-12-14 11:28:10 -05:00
jianglai
a85544b3f6 Use gson to make JSON string in proxy log formatter
This is simpler than using fasterxml.jackson.

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=224583713
2018-12-12 13:22:34 -05:00
jianglai
f46f817f9e Create proxy Docker image with Gradle
This CL adds two ways to build an docker image with gradle:

1) Adds a :proxy:deployJar task that builds an uber jar that contains all runtime dependencies. The jar can then be add to a docker image by calling docker with the added Dockerfile. The base image for this image can be both distroless java or openjdk:alpine.

2) Uses the Gradle distribution plugin to build a distribution tar file that contains all dependencies (as separate jar files) and a run script that sets up the classpath before calling the main class. Then the docker application plugin can build a docker image (with the dockerBuildImage task) using the application tar file. This only works with openjdk:alpline base image as the distroless java image does not contain a shell and therefore the script created by the distribution plugin cannot be launched.

We may later decide to use one of the method and remove the other.

Also adds an outcast test pattern that caused the tests to be flaky.

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=222145192
2018-11-19 18:19:27 -05:00
jianglai
118367dc6b Move proxy to its separate project
It can be run directly by "gradle :proxy:run"

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=222102849
2018-11-19 18:17:47 -05:00