Cleaned up old files and reorganized navigation order.

[arvados.git] / doc / user / tutorial-new-pipeline.textile

---
layout: default
navsection: userguide
title: "Tutorial 5: Construct a new pipeline"
navorder: 15
---

h1. Tutorial 5: Construct a new pipeline

Here you will write two new crunch scripts, incorporate them into a new pipeline template, run the new pipeline a couple of times using different parameters, and compare the results. One of the new scripts will use the Arvados API to look up trait→human→data relations and use this information to compile a collection of data to analyze.

_Like the previous tutorial, this needs more of a basis in some actual
clinical/research question to motivate it_

h3. Prerequisites

* Log in to a VM "using SSH":ssh-access.html
* Put an "API token":api-tokens.html in your @ARVADOS_API_TOKEN@ environment variable
* Put the API host name in your @ARVADOS_API_HOST@ environment variable

If everything is set up correctly, the command @arv -h user current@ will display your account information.

h3. Git repository access

Pushing code to your git repository involves using your private key. There are a few ways to arrange this:

*Option 1:* Use an SSH agent, and log in to your VM with agent forwarding enabled. With Linux, BSD, MacOS, etc., this looks something like:

<pre>
ssh-add -l
eval `ssh-agent` # (only if "ssh-add -l" said it could not open a connection)
ssh-add          # (this adds your private key to the agent)
ssh -A my_vm.arvados
ssh-add -l       # (run this in your VM account to confirm forwarding works)
</pre>

_This discussion about ssh should probably go under the "ssh" section_

With PuTTY under Windows, run "pageant", add your key to the agent, and turn on agent forwarding in your PuTTY settings.

*Option 2:* Edit code on your workstation and push code to your Arvados repository from there instead of your VM account. Depending on your @.ssh/config@ file, you will use names like @my_vm_name.arvados@ instead of @my_vm_name.{{ site.arvados_api_host }}@ in git and ssh commands.

<pre>
git clone git@git.arvados:my_repo_name.git
cd my_repo_name
[...]
git push
</pre>

*Option 3:*  Edit code in your VM, and use git on your workstation as an intermediary.

<pre>
git clone git@my_vm_name.arvados:my_repo_name.git
cd my_repo_name
git remote add arvados git@git.arvados:my_repo_name.git

[...make edits and commits in your repository on the VM...]

git pull                        #(update local copy)
git push arvados master:master  #(update arvados hosted copy)
</pre>

Whichever setup you choose, if everything is working correctly, this command should give you a list of repositories you can access:

<pre>
ssh -T git@git.{{ site.arvados_api_host }}
</pre>

&darr;

<pre>
hello your_user_name, the gitolite version here is v2.0.2-17-g66f2065
the gitolite config gives you the following access:
     R   W      your_repo_name
</pre>

_You need to have a git repository set up already, which is not
necessarily the case for new users, so this should link to the git
section about setting up a new repo_

h3. Set some variables

Adjust these to match your login account name and the URL of your Arvados repository. The Access&rarr;VMs and Access&rarr;Repositories pages on Workbench will show the specifics.

<pre>
repo_url=git@git.{{ site.arvados_api_host }}:my_repo_name.git
repo_name=my_repo_name
</pre>

h3. Set up a new branch in your Arvados git repository

We will create a new empty branch called "pipeline-tutorial" and add our new crunch scripts there.

<pre>
mkdir pipeline-tutorial
cd pipeline-tutorial
git init
git checkout -b pipeline-tutorial
git remote add origin $repo_url
</pre>

_Should explain each step_
_Creating an empty branch in an empty repository makes git do weird
things, need to fix using
<pre>
git branch --set-upstream pipeline_tutorial origin/pipeline_tutorial
</pre>
but I don't know what this means._

h3. Write the create-collection-by-trait script

<pre>
mkdir -p crunch_scripts
touch crunch_scripts/create-collection-by-trait
chmod +x crunch_scripts/create-collection-by-trait
nano crunch_scripts/create-collection-by-trait
</pre>

_the -p to mkdir isn't necessary here_

Here is the script:

<pre>
#!/usr/bin/env python

import arvados
import re
import json

trait_name = arvados.current_job()['script_parameters']['trait_name']

# get UUIDs of all matching traits
trait_uuids = map(lambda t: t['uuid'],
                  filter(lambda t: re.search(trait_name, t['name'], re.IGNORECASE),
                         arvados.service.traits().list(limit=1000).execute()['items']))

# list humans linked to these conditions
trait_links = arvados.service.links().list(limit=10000, where=json.dumps({
            'link_class': 'human_trait',
            'tail_kind': 'arvados#human',
            'head_uuid': trait_uuids
            })).execute()['items']
human_uuids = map(lambda l: l['tail_uuid'], trait_links)

# find collections linked to these humans
provenance_links = arvados.service.links().list(where=json.dumps({
    "link_class": "provenance",
    "name": "provided",
    "tail_uuid": human_uuids
  })).execute()['items']
collection_uuids = map(lambda l: l['head_uuid'], provenance_links)

# pick out all of the "var" files, and build a new collection
out_manifest = ''
for locator in collection_uuids:
    for f in arvados.CollectionReader(locator).all_files():
        if re.search('var-.*\.tsv.bz2', f.name()):
            out_manifest += f.as_manifest()

# output the new collection
arvados.current_task().set_output(arvados.Keep.put(out_manifest))
</pre>

h3. Write the find-dbsnp-id script

<pre>
touch crunch_scripts/find-dbsnp-id
chmod +x crunch_scripts/find-dbsnp-id
nano crunch_scripts/find-dbsnp-id
</pre>

Here is the script:

<pre>
#!/usr/bin/env python

import arvados
import re

arvados.job_setup.one_task_per_input_file(if_sequence=0, and_end_task=True)

this_job = arvados.current_job()
this_task = arvados.current_task()
this_task_input = this_task['parameters']['input']
dbsnp_search_pattern = re.compile("\\bdbsnp\\.\\d+:" +
                                  this_job['script_parameters']['dbsnp_id'] +
                                  "\\b")

input_file = list(arvados.CollectionReader(this_task_input).all_files())[0]
out = arvados.CollectionWriter()
out.set_current_file_name(input_file.decompressed_name())
out.set_current_stream_name(input_file.stream_name())
for line in input_file.readlines():
    if dbsnp_search_pattern.search(line):
        out.write(line)

this_task.set_output(out.finish())
</pre>

_This should probably match the code we ran the user through in the
previous tutorial, with the only difference being that the prior
tutorial is interactive, and this tutorial is demonstrating how to
create a job.

h3. Commit your new code

<pre>
git add crunch_scripts/create-collection-by-trait
git add crunch_scripts/find-dbsnp-id
git commit -m 'add scripts from tutorial'
</pre>

h3. Push your new code to your Arvados git repository

Push the new "pipeline-tutorial" branch to your Arvados hosted repository.

<pre>
git push origin pipeline-tutorial
</pre>

h3. Note the commit ID of your latest code

Show the latest commit. The first line includes a 40-digit hexadecimal number that uniquely identifies the content of your git tree. You will specify this in your pipeline template in the next step to ensure that Arvados uses the correct version of your git tree when running jobs.

<pre>
git show | head
</pre>

&darr;

<pre>
commit 37c7faef1b066a2dcdb0667fbe82b7cdd7d0be93
[...]
</pre>

h3. Write the pipeline template

Make a directory called @pipeline_templates@ and create a file called @find-dbsnp-by-trait.json@.

<pre>
mkdir pipeline_templates
nano pipeline_templates/find-dbsnp-by-trait.json
</pre>

Copy the following pipeline template.

<pre>
{
  "name":"find_dbsnp_by_trait",
  "components":{
    "create_collection":{
      "script":"create-collection-by-trait",
      "script_parameters":{
        "trait_name":"Non-melanoma skin cancer"
      },
      "script_version":"YOUR_GIT_COMMIT_SHA1_HERE"
    },
    "find_variant":{
      "script":"find-dbsnp-id",
      "script_parameters":{
        "input":{
          "output_of":"create_collection"
        },
        "dbsnp_id":"rs1126809"
      },
      "script_version":"YOUR_GIT_COMMIT_SHA1_HERE"
    }
  }
}
</pre>

_This desparately needs to be explained, since this is the actual
pipeline definition_

h3. Store the pipeline template in Arvados

<pre>
read -rd $'\000' the_pipeline < pipeline_templates/find-dbsnp-by-trait.json

arv pipeline_template create --pipeline-template "$the_pipeline"
</pre>

@arv@ will output the UUID of the new pipeline template.

<pre>
qr1hi-p5p6p-uf9gi9nolgakm85
</pre>

The new pipeline template will also appear on the Workbench&rarr;Compute&rarr;Pipeline&nbsp;templates page.

_Storing the pipeline in arvados as well as in git seems redundant_

h3. Invoke the pipeline using "arv pipeline run"

Replace the UUID here with the UUID of your own new pipeline template:

<pre>
arv pipeline run --template qr1hi-p5p6p-uf9gi9nolgakm85
</pre>

This instantiates your pipeline template: it submits the first job, waits for it to finish, submits the next job, etc.

h3. Monitor pipeline progress

The "arv pipeline run" command displays progress in your terminal until the pipeline instance is finished.

<pre>
2013-07-17 05:06:15 +0000 -- pipeline_instance qr1hi-d1hrv-8i4tz440whvwf2o
create_collection qr1hi-8i9sb-haibhu51olihlwp 9e2e489a73e1a918de8ecfc6f59ae5a1+1803+K@qr1hi
find_variant      qr1hi-8i9sb-sqduc932xb1tpff cad082ba4cb174ffbebf751bbe3ed77c+506+K@qr1hi
</pre>

The new pipeline instance will also show up on your Workbench&rarr;Compute&rarr;Pipeline&nbsp;instances page.

h3. Find the output collection UUID

The output of the "find_variant" component is shown in your terminal with the last status update from the "arv pipeline run" command.

It is also displayed on the pipeline instance detail page: go to Workbench&rarr;Compute&rarr;Pipeline&nbsp;instances and click the UUID of your pipeline instance.

_There needs to be an easier way to get the output from the workbench_

h3. Compute a summary statistic from the output collection

For this step we will use python to read the output manifest and count how many of the inputs produced hits.

Type @python@ at a command prompt and paste this script at the prompt:

<pre>
import arvados

hash = 'cad082ba4cb174ffbebf751bbe3ed77c+506+K@qr1hi' # Use your result here!

collection = arvados.CollectionReader(hash)
hits = len(filter(lambda f: f.size() > 0, collection.all_files()))
misses = len(filter(lambda f: f.size() == 0, collection.all_files()))
print "%d had the variant, %d did not." % (hits, misses)
</pre>

&darr;

<pre>
4 had the variant, 3 did not.
</pre>

_Explain each step_

_h3. Run the pipeline again using different parameters

We can use the same pipeline template to run the jobs again, this time overriding the "trait_name" parameter with a different value:

<pre>
arv pipeline run --template qr1hi-p5p6p-uf9gi9nolgakm85 create_collection::trait_name=cancer
</pre>

When this template instance finishes, run the same Python script on the new output collection to summarize the results.