Notes for Developers

Upgrading the gRPC protobuf files

Buildgrid’s gRPC stubs are not built as part of installation. Instead, they are precompiled and shipped with the source code. The protobufs are prone to compatibility-breaking changes, so we update them manually.

First bring the updated proto file into the source tree. For example, if updating the remote execution proto, replace the old buildgrid/_protos/build/bazel/remote/execution/v2/remote_execution.proto with the newer one.

Then, compile the protobufs. Continuing with the remote_execution.proto example…

pip install grpcio grpcio-tools grpc-stubs mypy-protobuf
python -m grpc_tools.protoc -Ibuildgrid/_protos/ --python_out=buildgrid/_protos --grpc_python_out=buildgrid/_protos --mypy_out=buildgrid/_protos build/bazel/remote/execution/v2/remote_execution.proto

The most important thing here is to make sure that buildgrid/_protos is in your include path with -Ibuildgrid/_protos. If all goes well, the new files should have been generated.

Implementing a data store for the Scheduler

The buildgrid.server.scheduler.Scheduler can be configured to use one of multiple backends. Currently in-memory and SQL-based backends are available and supported.

It is possible to implement a new backend, by implementing the data store interface (buildgrid.server.persistence.interface.DataStoreInterface) and adding a class to buildgrid/_app/settings/ to allow the interface to be configured.

The implementation is free to decide how to persist the data, as long as all of the abstract methods of DataStoreInterface are implemented. The implementations are only required to exist as specified; the details are left up to the author. This allows implementations to make unneccessary methods do nothing for example, as long as the expected data type is returned. There may be nothing to do for enabling/disabling monitoring for some implementations for example.

Backend implementations are also required to implement some way of triggering the events that are used when streaming updates to clients. These are instances of buildgrid.utils.TypedEvent, and notify_change should be used to indicate an update message should be sent, and notify_stop should be used to indicate that the thread handling the stream should check whether the client is connected, and stop if not. Implementations should only really need to use notify_change, as the disconnect logic is in the DataStoreInterface already.

Both existing implementations start a thread which periodically checks the state of the data for jobs that are being watched and compares it with the previous state. If it detects a change, then it calls notify_change on the relevant event, which is stored in the buildgrid.utils.JobWatchSpec for the affected job, which is in the self.watched_jobs dictionary. Adding/removing entries from this dictionary is handled by the DataStoreInterface class, so doesn’t need to be a concern for implementations.

The class to parse a YAML tag to allow the data store implementation to be configured should inherit from YamlFactory and have a __new__ method which returns an instance of the implementation. There are no other limitations on what it should and shouldn’t do. In order to allow the parser to understand the tag, the get_parser function at the bottom of should also be modified to add a constructor for the new tag.