Verifier: running and monitoring Blender process

[cameel]

This is a part of the implementation of the #193 blueprint. This task depends on #225.

Settings and checks

• BLENDER_MAX_RENDERING_TIME: defines the maximum time (in seconds) rendering a Blender project can take. Choose a sensible default value.
  • Must be defined if 'verifier' is in CONCENT_FEATURES. Otherwise must be undefined or None.
  • Must be a positive integer

Sequence of operation

• Verifier starts a child process in such a way that it can monitor its stdout and stderr. The output should be logged.
  • Don't use shell to start the process. Start it directly. We don't need the overhead and we don't need shell features here anyway.
• Verifier checks the exit code from the child process.
  • If Blender finishes with errors, verification ends here. verification_result informing about the error is sent to the work queue.
  • If rendering takes more than BLENDER_MAX_RENDERING_TIME Verifier kills Blender process. Verification ends here and verification_result informing about the error is sent to the work queue.
  • If Blender finishes successfully, verification proceeds.
• Verifier deletes source files of the Blender project from its storage.

Setup for local development

Work queue

For local development you'll need to run a RabbitMQ server. You can run RabbitMQ in a docker container. See https://github.com/golemfactory/concent/issues/197#issuecomment-375620295.

Blender

You can install blender on your machine or run it in the Docker container created by Golem team. Just make the command to start it configurable.

[rwrzesien]

@cameel Are there any docs on how should blender command looks like ?

[cameel]

Update: We need the default value for BLENDER_MAX_RENDERING_TIME. That's for now. Now that I think about it, we should probably have a deadline for blender based on message timings instead of a constant value.

[cameel]

@rwrzesien

Are there any docs on how should blender command looks like ?

Unfortunately no. You have to look at what Golem does. It embeds the command inside src_code in TaskToCompute. Please look at the example message in #235.

Fortunately we don't have to parse it out of this message and we definitely should not run code from the client. We agreed with @lukasz-glen that Concent will ignore this field until TaskToCompute is changed to store all information needed to run Blender in a more declarative and portable way.

[dybi]

BLENDER_MAX_RENDERING_TIME - default value is one week. Local testing has been conducted (without real run of blender - we still do not have files to run verification) and the flow works correctly.

[cameel]

I think that we should not close this issue until we at least check that we can run Blender and it can successfully complete its rendering. I think there still may be some undiscovered problems here. I don't consider the task finished until it can pass at least a simple smoke test.

[cameel]

Here are some test files:

• golem issue #132 > Test Blender Scenes
• Golem Resources (Unofficial)
• blender.org > Demo Files