Dendro

Installation

pip install --upgrade dendro

Or install from source.

Getting started

Hello world on public resources

Let's run a simple example job using remote public resources.

First, you'll need a Dendro API key.

• Visit https://dendro.vercel.app/settings and log in using GitHub.
• Click "Regenerate API key" and copy the key.
• Set the environtment variable DENDRO_API_KEY to the key you copied. For example, in bash:

export DENDRO_API_KEY=your-api-key

Now (after cloning this repo) run the first example:

python examples/example1.py

The output should look something like this:

https://dendro.vercel.app/job/9yS0phSTf1lqgLdKdFa5 completed

Well, that was fast! What happened?

It turns out that this exact job was already run by someone else, and the result was stored in the public resource. By default, Dendro will check to see if the job has already been run and return the result if it has.

Click on that link to see the job details, including the console output which should include the string "Hello, world!" This job has no input or output files, so the console output is the only interesting thing to see.

Let's examine the code:

import os
from dendro.client import submit_job, DendroJobDefinition, DendroJobRequiredResources, DendroJobParameter


service_name = os.getenv('DENDRO_SERVICE_NAME', 'hello_world_service')

def main():
    job_def = DendroJobDefinition(
        appName='hello_world',
        processorName='hello_world_1',
        inputFiles=[],
        outputFiles=[],
        parameters=[
            DendroJobParameter(
                name='name',
                value='world'
            )
        ]
    )
    required_resources = DendroJobRequiredResources(
        numCpus=1,
        numGpus=0,
        memoryGb=4,
        timeSec=60
    )
    job = submit_job(
        service_name=service_name,
        job_definition=job_def,
        required_resources=required_resources,
        tags=['example'],
        rerun_failing=True,
        delete_failing=True
    )
    print(job.job_url, job.status)

if __name__ == '__main__':
    main()

The job_def is the job definition which uniquely defines job in terms of the Dendro app, the processor within that app, the inputs, outputs, and parameters. The "hello_world" app is installed on the service, and the processor "hello_world_1" is defined within that app. See here for the source code defining its functionality. The hash of this job definition is what is used to determine whether the job has already been run on the service.

The required_resources is self-explanatory. This determines whether the job can be run on a given compute client. If the resources are too demanding, the job may never run.

Then submit_job is the function that actually queues the Dendro job in the central database (or returns an existing job). Here's an explanation of the arguments:

• service_name is the name of the service that will run the job. A service is a piece of the Dendro network that has a collection of available Apps, a collection of compute clients, and a list of privileged users. In this case the service is hello_world_service, which is a public service (with limited resources) that anyone can use.
• job_definition is the job definition.
• required_resources is the required resources.
• tags is a list of tags that can be used to filter jobs.
• rerun_failing is a boolean that determines whether to rerun the job if a failing job with the same job definition already exists. If this is set to False (the default) the failing job will be returned.
• delete_failing is a boolean that determines whether to delete the failing job if it is rerun.

Abother possible argument not mentioned is the skip_cache boolean.

Now, create a new script example1_test.py and modify this example with a custom "name" parameter. Perhaps set it to your own name. Assuming nobody has run that particular job definition, you should see something like

https://dendro.vercel.app/job/your-job-id pending

This will stay in the pending state until a compute client picks it up and runs it. You can check the status of the job by visiting the link or by running the script again.

Running a job that produces an output file

The second example is a bit more interesting because it creates an output file, namely a text file with the contents "Hello, world!". Try it out!

python examples/example2.py

Here's a breakdown of the code:

import os
from dendro.client import submit_job, DendroJobDefinition, DendroJobRequiredResources, DendroJobOutputFile, DendroJobParameter


service_name = os.getenv('DENDRO_SERVICE_NAME', 'hello_world_service')

def main():
    job_def = DendroJobDefinition(
        appName='hello_world',
        processorName='hello_world_2',
        inputFiles=[],
        outputFiles=[
            DendroJobOutputFile(
                name='output',
                fileBaseName='output.txt'
            )
        ],
        parameters=[
            DendroJobParameter(
                name='name',
                value='world'
            )
        ]
    )
    required_resources = DendroJobRequiredResources(
        numCpus=1,
        numGpus=0,
        memoryGb=4,
        timeSec=60
    )
    job = submit_job(
        service_name=service_name,
        job_definition=job_def,
        required_resources=required_resources,
        tags=['example']
    )
    print(job.job_url, job.status)

if __name__ == '__main__':
    main()

This code should be fairly self-explanatory. In addition to the parameter in the job defintion, we have an output file named "output".

After running the script, you should see something like this:

https://dendro.vercel.app/job/Pjc2mcych7MOPz5nI1Up completed

Opening that link once again brings you to job details where you can click on the URL of the generated output file to see the expected content "Hello, world!".

Running a job that depends on a previous job

The third example (example3.py) is a bit more complex because it involves a job that depends on the output of a previous job. Dendro will automatically handle the job orchestration for you in that it will not run the dependent job until the previous job has completed successfully.

Let's take a look at the code.

import os
from dendro.client import submit_job, DendroJobDefinition, DendroJobRequiredResources, DendroJobInputFile, DendroJobOutputFile, DendroJobParameter


service_name = os.getenv('DENDRO_SERVICE_NAME', 'hello_world_service')

def main():
    file1 = DendroJobOutputFile(
        name='output',
        fileBaseName='output.txt'
    )
    job_def = DendroJobDefinition(
        appName='hello_world',
        processorName='hello_world_2',
        inputFiles=[],
        outputFiles=[
            file1
        ],
        parameters=[
            DendroJobParameter(
                name='name',
                value='world'
            )
        ]
    )
    required_resources = DendroJobRequiredResources(
        numCpus=1,
        numGpus=0,
        memoryGb=4,
        timeSec=60
    )
    job1 = submit_job(
        service_name=service_name,
        job_definition=job_def,
        required_resources=required_resources,
        tags=['example']
    )
    print(job1.job_url, job1.status)

    job2_def = DendroJobDefinition(
        appName='hello_world',
        processorName='count_characters',
        inputFiles=[
            DendroJobInputFile(
                name='input',
                fileBaseName='input.txt',
                url=file1
            )
        ],
        outputFiles=[
            DendroJobOutputFile(
                name='output',
                fileBaseName='output.json'
            )
        ],
        parameters=[
            DendroJobParameter(
                name='include_whitespace',
                value=True
            )
        ]
    )
    job2 = submit_job(
        service_name=service_name,
        job_definition=job2_def,
        required_resources=required_resources,
        tags=['example'],
        rerun_failing=True
    )
    print(job2.job_url, job2.status, job2.isRunnable)

if __name__ == '__main__':
    main()

Running this will produce two jobs and the output will look something like this:

python examples/example3.py

https://dendro.vercel.app/job/Pjc2mcych7MOPz5nI1Up completed
https://dendro.vercel.app/job/Ctv7xyt66vHvrwIKVasu completed True

Click on that second link and you should see the output file "output.json" which contains the number of characters in the input file. The input file is the output of the first job, which is a text file with the contents "Hello, world!".

Neurosift integration

Neurosift is a web-based tool designed for visualizing NWB files, particularly those hosted in the DANDI Archive. While many of Neurosift’s visualizations are processed client-side and require minimal computational resources, certain visualizations necessitate more intensive data processing that must be handled server-side. The integration of Neurosift with Dendro enables the offloading of these computationally demanding tasks to Dendro’s distributed computing environment.

Upon submitting a job to Dendro via the Neurosift interface, users can monitor its progress directly within Neurosift. Once completed, the visualization’s output is automatically downloaded to the user’s browser for rendering. Because Dendro caches the results of these jobs, subsequent identical requests by any user will be served from the cache, without the need to recompute the visualization.

For a guided tour of Neurosift and Dendro, see this workshop presentation.

Installing from source

# clone the repo, then
cd dendro
cd python
pip install -e .

Dendro system architecture

Dendro system architecture

Important note to self (do not delete)

When using Cloudflare R2 with Range headers and large files, it's important to configure the website to bypass the cache.

See: https://community.cloudflare.com/t/public-r2-bucket-doesnt-handle-range-requests-well/434221/4