Refine.bio harmonizes petabytes of publicly available biological data into ready-to-use datasets for cancer researchers and AI/ML scientists.
This README file is about building and running the refine.bio project source code.
If you're interested in simply using the service, you should go to the website or read the documentation.
Refine.bio currently has four sub-projects contained within this repo:
- common Contains code needed by both
foremanandworkers. - foreman Discovers data to download/process and manages jobs.
- workers Runs Downloader and Processor jobs.
- infrasctructure Manages infrastructure for Refine.bio.
- Refine.bio
refinebio uses a
feature branch
based workflow. New features should be developed on new feature branches, and
pull requests should be sent to the dev branch for code review. Merges into
master happen at the end of sprints, and tags in master correspond to
production releases.
To run Refine.bio locally, you will need to have the prerequisites installed onto your local machine. This will vary depending on whether you are developing on a Mac or a Linux machine. Linux instructions have been tested on Ubuntu 16.04 or later, but other Linux distributions should be able to run the necessary services. Microsoft Windows is currently unsupported by this project.
Note: The install_all.sh script will configure a git pre-commit hook to auto-format your python code. This will format your code in the same way as the rest of the project, allowing it to pass our linting check.
The easiest way to run Refine.bio locally is to run ./scripts/install_all.sh
to install all of the necessary dependencies. As long as you are using a recent
version of Ubuntu or macOS it should work. If you are using another version of
Linux it should still install most of the dependencies as long as you give the
appropriate INSTALL_CMD environment variable, but some dependencies may be
named differently in your package manager than in Ubuntu's.
The following services will need to be installed:
- Python3 and Pip:
sudo apt-get -y install python3-pip - Docker: Be sure to follow the post installation steps so Docker does not need sudo permissions.
- Terraform
- pip3 can be installed on Linux clients with
sudo apt-get install python3-pip - black can be installed on Linux clients with
pip3 install black - jq
- iproute2
- shellcheck
Instructions for installing Docker and Terraform can be found by
following the link for each service. jq and iproute2 can be installed via
sudo apt-get install jq iproute2 shellcheck.
The following services will need to be installed:
Instructions for installing Docker and Homebrew can be found by on their respective homepages.
Once Homebrew is installed, the other required applications can be installed by running: brew install iproute2mac terraform jq black shellcheck.
Many of the computational processes running are very memory intensive. You will need to raise the amount of virtual memory available to Docker from the default of 2GB to 12GB or 24GB, if possible.
Run ./scripts/create_virtualenv.sh to set up the virtualenv. It will activate the dr_env
for you the first time. This virtualenv is valid for the entire refinebio
repo. Sub-projects each have their own environments managed by their
containers. When returning to this project you should run
source dr_env/bin/activate to reactivate the virtualenv.
refinebio also depends on Postgres. Postgres can be
run in a local Docker container
To start a local Postgres server in a Docker container, use:
./scripts/run_postgres.shThen, to initialize the database, run:
./scripts/install_db_docker.shIf you need to access a psql shell for inspecting the database, you can use:
./scripts/run_psql_shell.shor if you have psql installed this command will give you a better shell experience:
source scripts/common.sh && PGPASSWORD=mysecretpassword psql -h $(get_docker_db_ip_address) -U postgres -d data_refinery
The common sub-project contains common code which is
depended upon by the other sub-projects. So before anything else you
should prepare the distribution directory common/dist with this
script:
./scripts/update_models.sh(Note: This step requires the postgres container to be running and initialized.)
Note: there is a small chance this might fail with a can't stat, error. If this happens, you have
to manually change permissions on the volumes directory with sudo chmod -R 740 volumes_postgres
then re-run the migrations.
One of the API endpoints is powered by ElasticSearch. ElasticSearch must be running for this functionality to work. A local ElasticSearch instance in a Docker container can be executed with:
./scripts/run_es.shAnd then the ES Indexes (akin to Postgres 'databases') can be created with:
./scripts/rebuild_es_index.shTo run the entire test suite:
./scripts/run_all_tests.sh(Note: Running all the tests can take some time, especially the first time because it downloads a lot of files.)
For more granular testing, you can just run the tests for specific parts of the system.
To just run the API tests:
./api/run_tests.shTo just run the common tests:
./common/run_tests.shTo just run the foreman tests:
./foreman/run_tests.shTo just run the workers tests:
./workers/run_tests.shIf you only want to run tests with a specific tag, you can do that too. For example, to run just the salmon tests:
./workers/run_tests.sh -t salmonAll of our worker tests are tagged, generally based on the Docker image required to run them. Possible values for worker test tags are:
- affymetrix
- agilent
- downloaders
- illumina
- no_op
- qn (short for quantile normalization)
- salmon
- smasher
- transcriptome
R files in this repo follow Google's R Style Guide. Python Files in this repo follow PEP 8. All files (including Python and R) have a line length limit of 100 characters.
In addition to following pep8, python files must also conform to the formatting style enforced by black.
black is a highly opinionated auto-formatter.
(black's highly opinionated style is a strict sub-set of pep8.)
The easiest way to conform to this style is to run black . --line-length=100.
This will auto-format your code.
Running the ./scripts/install_all.sh script will install a pre-commit git hook that will run this formatter on every commit you make locally. Under the hood this uses pre-commit, which you can also install directly by running pip3 install pre-commit & pre-commit install. Then, if you want to run pre-commit without making a git commit, you can use pre-commit run --all-files.
To install black see the installation instructions.
Any Pull Requests that do not conform to the style enforced by black will be flagged by our continous integration and will not be accepted until that check passes.
All user-facing scripts have been linted with shellcheck for common
warnings and POSIX-correctness. If a script is user-facing, it should ideally
be POSIX-compliant and have the extension .sh, but if bashisms are necessary
it should have the extension .bash. To install shellcheck, you can run
apt-get install shellcheck or brew install shellcheck. Then, you can lint
scripts with shellcheck FILE.
During development, you make encounter some occasional strangeness. Here's some things to watch out for:
- Since we use multiple Docker instances, don't forget to
./scripts/update_models - If builds are failing, increase the size of Docker's memory allocation. (Mac only.)
- If Docker images are failing mysteriously during creation, it may
be the result of Docker's
Docker.qcow2orDocker.rawfile filling. You can prune old images withdocker system prune -a. - If it's killed abruptly, the containerized Postgres images can be left in an unrecoverable state. Annoying.
We have created some utilities to help us keep R stable, reliable, and from periodically causing build errors related to version incompatibilites.
The primary goal of these is to pin the version for every R package that we have.
The R package devtools is useful for this, but in order to be able to install a specific version of it, we've created the R script common/install_devtools.R.
There is another gotcha to be aware of should you ever need to modify versions of R or its packages.
In Dockerfiles for images that need the R language, we install apt packages that look like r-base-core=3.4.2-1xenial1.
It's unclear why the version for these is so weird, but it was determined by visiting the package list here: https://cran.revolutionanalytics.com/bin/linux/ubuntu/xenial/
If it needs to be updated then a version should be selected from that list.
Additionally there are two apt packages, r-base and r-base-core, which seem to be very similar except that r-base-core is slimmed down some by not including some additional packages. For a while we were using r-base, but we switched to r-base-core when we pinned the version of the R language because the r-base package caused an apt error.
Once you've built the common/dist directory and have
the Postgres service running, you're ready to run jobs.
To run the API you also need the elasticsearch service running.
There are three kinds of jobs within Refine.bio.
The API can be run with:
./api/serve.shSurveyor Jobs discover samples to download/process along with recording metadata about the samples.
A Surveyor Job should queue Downloader Jobs to download the data it discovers.
However, at the moment there is no automated way for the downloader jobs to be run.
This will be resolved ASAP, see AlexsLemonade#2775 for more information.
The Surveyor can be run with the ./foreman/run_management_command.sh script.
The first argument to this script is the type of Surveyor Job to run, which will always be survey_all.
Details on these expected arguments can be viewed by running:
./foreman/run_management_command.sh survey_all -hThe Surveyor can accept a single accession code from any of the source data repositories (e.g., Sequencing Read Archive, ArrayExpress, Gene Expression Omnibus):
./foreman/run_management_command.sh survey_all --accession <ACCESSION_CODE>Example for a GEO experiment:
./foreman/run_management_command.sh survey_all --accession GSE85217Example for an ArrayExpress experiment:
./foreman/run_management_command.sh survey_all --accession E-MTAB-3050 # AFFY
./foreman/run_management_command.sh survey_all --accession E-GEOD-3303 # NO_OPTranscriptome indices are a bit special. For species within the "main" Ensembl division, the species name can be provided like so:
./foreman/run_management_command.sh survey_all --accession "Homo sapiens"However for species that are in other divisions, the division must follow the species name after a comma like so:
./foreman/run_management_command.sh survey_all --accession "Caenorhabditis elegans, EnsemblMetazoa"The possible divisions that can be specified are:
- Ensembl (this is the "main" division and is the default)
- EnsemblPlants
- EnsemblFungi
- EnsemblBacteria
- EnsemblProtists
- EnsemblMetazoa
If you are unsure what division a species falls into, unfortunately the only way to tell is go to check ensembl.com. (Although googling the species name + "ensembl" may work pretty well.)
You can also supply a newline-deliminated file to survey_all which will
dispatch survey jobs based on accession codes like so:
./foreman/run_management_command.sh survey_all --file MY_BIG_LIST_OF_CODES.txtThe main foreman job loop can be started with:
./foreman/run_management_command.sh retry_jobsThis must actually be running for jobs to move forward through the pipeline.
When surveying SRA, you can supply either run accession codes (e.g.,
codes beginning in SRR, DRR, or ERR) or study accession codes
(SRP, DRP, ERP).
Run example (single read):
./foreman/run_management_command.sh survey_all --accession DRR002116Run example (paired read):
./foreman/run_management_command.sh survey_all --accession SRR6718414Study example:
./foreman/run_management_command.sh survey_all --accession ERP006872Building transcriptome indices used for quantifying RNA-seq data requires us to retrieve genome information from Ensembl. The Surveyor expects a species' scientific name in the main Ensembl division as the accession:
./foreman/run_management_command.sh survey_all --accession "Homo Sapiens"See the Ensembl Transcriptome Index section for additional usage examples inclduing surveying additional Ensembl divisions.
Downloader Jobs will be queued automatically when Surveyor Jobs
discover new samples. However, if you just want to queue a Downloader Job
yourself rather than having the Surveyor do it for you, you can use the ./workers/run_job.sh
script:
./workers/run_job.sh run_downloader_job --job-name=<EXTERNAL_SOURCE> --job-id=<JOB_ID>For example:
./workers/run_job.sh run_downloader_job --job-name=SRA --job-id=12345or
./workers/run_job.sh run_downloader_job --job-name=ARRAY_EXPRESS --job-id=1Or for more information run:
./workers/run_job.sh -hProcessor Jobs will be queued automatically by successful Downloader Jobs.
However, if you just want to run a Processor Job without yourself without having
a Downloader Job do it for you, the following command will do so:
./workers/run_job.sh -i <IMAGE_NAME> run_processor_job --job-name=<JOB_NAME> --job-id=<JOB_ID>For example
./workers/run_job.sh -i affymetrix run_processor_job --job-name=AFFY_TO_PCL --job-id=54321or
./workers/run_job.sh -i no_op run_processor_job --job-name=NO_OP --job-id=1or
./workers/run_job.sh -i salmon run_processor_job --job-name=SALMON --job-id=1or
./workers/run_job.sh -i transcriptome run_processor_job --job-name=TRANSCRIPTOME_INDEX_LONG --job-id=1Or for more information run:
./workers/run_job.sh -hIf you want to quantile normalize combined outputs, you'll first need to create a reference target for a given organism or organisms. This can be done in a production environment by running the following on the Foreman instance:
./run_management_command.sh dispatch_qn_jobs --organisms=DANIO_RERIO,HOMO_SAPIENSTo create QN targets for all organisms with enough processed samples:
./run_management_command.sh dispatch_qn_jobsThis will at some point move to the foreman and then it will take a list of organisms to create QN targets for.
Creating species-wide compendia for a given species can be done in a production environment by running the following on the Foreman instance:
./run_management_command.sh create_compendia --organisms=DANIO_RERIO --svd-algorithm=ARPACKor for a list of organisms:
./run_management_command.sh create_compendia --organisms=DANIO_RERIO,HOMO_SAPIENS --svd-algorithm=ARPACKor for all organisms with sufficient data:
./run_management_command.sh create_compendia --svd-algorithm=ARPACKAlternatively a compendium can be created which only includes quant.sf files by using the create_quantpentida command:
./run_management_command.sh create_quantpendia --organisms=DANIO_RERIO
Compendia jobs run on the smasher instance.
However they require a very large amount of RAM to be able to complete.
Our smasher instance does not generally have enough RAM to be able to run them, so if you need to run a smasher job you should temporarily increase the size of the smasher instance.
This can be done by changing the terraform variable smasher_instance_type which can be found in infrastructure/variables.tf.
Select an AWS instance type that has enough RAM to run the compendia jobs.
At the time of writing, compendia jobs require 180GB of RAM and m5.12xlarge has 192GM of RAM so it is sufficiently large to run the jobs.
Normally we wait until ever sample in an experiment has had Salmon run on it before we run Tximport. However Salmon won't work on every sample, so some experiments are doomed to never make it to 100% completion. Tximport can be run on such experiments by running the follow on the Foreman instance:
To run tximport on all eligible experiments:
./run_management_command.sh run_tximportTo run tximport on a single experiment if it is eligible:
./run_management_command.sh run_tximport --accession-codes=SRP095529To run tximport on a the eligible experiments in a list:
./run_management_command.sh run_tximport --accession-codes=SRP095529,ERP006872Note that if the experiment does not have at least 25 samples with at least 80% of them processed, this will do nothing.
It can be useful to have an interactive Python interpreter running within the
context of the Docker container. The scripts/run_shell.sh script has been provided
for this purpose. It is in the top level directory so that if you wish to
reference it in any integrations its location will be constant. However, it
is configured by default for the Foreman project. The interpreter will
have all the environment variables, dependencies, and Django configurations
for the Foreman project. There are instructions within the script describing
how to change this to another project.
Refine.bio requires an active, credentialed AWS account with appropriate permissions to create network infrastructure, users, compute instances and databases.
Deploys are automated to run via CirlceCI whenever a signed tag starting with a v is pushed to either the dev or master branches (v as in version, i.e. v1.0.0).
Tags intended to trigger a staging deploy MUST end with -dev, i.e. v1.0.0-dev.
CircleCI runs a deploy on a dedicated AWS instance so that the Docker cache can be preserved between runs.
Instructions for setting up that instance can be found in the infrastructure/deploy_box_instance_data.sh script.
To trigger a new deploy, first see what tags already exist with git tag --list | sort --version-sort
We have two different version counters, one for dev and one for master so a list including things like:
- v1.1.2
- v1.1.2-dev
- v1.1.3
- v1.1.3-dev
However you may see that the dev counter is way ahead, because we often need more than one staging deploy to be ready for a production deploy.
This is okay, just find the latest version of the type you want to deploy and increment that to get your version.
For example, if you wanted to deploy to staging and the above versions were the largest that git tag --list output, you would increment v1.1.3-dev to get v1.1.4-dev.
Once you know which version you want to deploy, say v1.1.4-dev, you can trigger the deploy with these commands:
git checkout dev
git pull origin dev
git tag -s v1.1.4-dev
git push origin v1.1.4-devgit tag -s v1.1.4-dev will prompt you to write a tag message; please try to make it descriptive.
We use semantic versioning for this project so the last number should correspond to bug fixes and patches, the second middle number should correspond to minor changes that don't break backwards compatibility, and the first number should correspond to major changes that break backwards compatibility.
Please try to keep the dev and master versions in sync for major and minor versions so only the patch version gets out of sync between the two.
Refine.bio uses a number of different Docker images to run different pieces of the system.
By default, refine.bio will pull images from the Dockerhub repo ccdlstaging.
If you would like to use images you have built and pushed to Dockerhub yourself you can pass the -r option to the deploy.sh script.
To make building and pushing your own images easier, the scripts/update_docker_images.sh has been provided.
The -r option will allow you to specify which repo you'd like to push to.
If the Dockerhub repo requires you to be logged in, you should do so before running the script using docker login.
The -v option allows you to specify the version, which will both end up on the Docker images you're building as the SYSTEM_VERSION environment variable and also will be the docker tag for the image.
scripts/update_docker_images.sh will not build the dr_affymetrix image, because this image requires a lot of resources and time to build.
It can instead be built with ./scripts/prepare_image.sh -i affymetrix -r <YOUR_DOCKERHUB_REPO>.
WARNING: The affymetrix image installs a lot of data-as-R-packages and needs a lot of disk space to build the image.
It's not recommended to build the image with less than 75GB of free space on the disk that Docker runs on.
There are a few extra things that you need to install before deploying the stack:
- Terraform, if you haven't already installed it
- awscli, for interacting with AWS
- boto3, which is necessary for some of our deployment scripts
- PostgreSQL, which is necessary for some of our deployment scripts
The easiest way to install Terraform is by running ./scripts/install_all.sh, or you can also install it manually by following the directions on the website. We currently use version 0.13.5.
For awscli and boto3, you need to install them using pip3 install awscli boto3. Ubuntu's repositories contain outdated versions of both packages which do not work with our deploy script.
Postgres can be installed using either apt install psql or brew install postgresql as appropriate.
Once you have all of the dependencies installed, you're almost ready to deploy a dev stack.
The only thing remaining is to make sure that you can authenticate properly.
To authenticate awscli, you need to run awscli configure and follow the directions.
For ssh access to the servers, which is used during the deploy, copy the RefinebioSSHKey from LastPass and save it to the file: infrastructure/data-refinery-key.pem.
If you do not have access to this key in LastPass, ask another developer.
The correct way to deploy to the cloud is by running the deploy.sh script. This script will perform additional
configuration steps, such as setting environment variables, setting up Batch job specifications, and performing database migrations. It can be used from the infrastructure directory like so:
./deploy.sh -u myusername -e dev -d us-east-1 -v v1.0.0 -r my-dockerhub-repoThis will spin up the whole system. It will usually take about 15 minutes, most of which is spent waiting for the Postgres instance to start.
The command above would spin up a development stack in the us-east-1 region where all the resources' names would end with -myusername-dev.
All of the images used in that stack would come from my-dockerhub-repo and would be tagged with v1.0.0.
The -e specifies the environment you would like to spin up. You may specify, dev, staging, or prod. dev is meant for individuals to test infrastructure changes or to run large tests. staging is to test the overall system before re-deploying to prod.
To see what's been created at any time, you can:
terraform state list
If you want to change a single entity in the state, you can use
terraform taint <your-entity-from-state-list>
And then rerun deploy.sh with the same parameters you originally ran it with.
refine.bio relies on AWS Batch as its job queue and uses it to provision instances. AWS Batch has three primary components:
- Compute Environments: These are what provision EC2 instances for refineb.bio. In this project each Compute Environment can either have one or zero instances. The goal is to have jobs that are run in the same compute environment be run on the same instance, so that data stored on the local disk by a downloader job will be available to the processor job. Only allowing a maximum of one instance per compute environment almost ensures this, however it is possible for an instance to be cycled in between jobs so sometimes the downloader job has to be rerun.
- Job Queues: These are what track the jobs submitted to AWS Batch and assign them to compute environments. In refine.bio each Job Queue uses a single compute environment, so if two jobs are placed in the same job queue they will be run in the same Compute Environment.
- Job Definitions: These are what specify the configuration to be used for each job type including what Docker Image will be used, what environment variables will be passed to it, what secrets it can access, and how many vCPUs and RAM it requires.
refine.bio uses three types of job queues:
- Compendia Job Queue: This job queue is for running very large compendia-building jobs that require a large instance. The Compute Environment assigned to this queue is configured to provision very large instances.
- Smasher Job Queue: This job queue is used for running smashing jobs. Having a dedicated queue for smasher jobs is useful because it ensure they won't be blocked by processing jobs and the instance provisioned by its compute environment has enough resources to run one of these jobs at a time and no more.
- Worker Job Queues: This is the only job queue with multiple instances. These do the general processing, so if there is a sufficient volume of work to necessitate more than one instance the Foreman will distribute jobs to more and more queues until all the queues are in use. The lowest index queue will be assigned Surveyor and Downloader jobs if it has capacity for them, if not the next lowest index queue with capacity will be chosen. Processor jobs will always be assigned to the same job queue that ran their downloader job.
Jobs can be submitted by running the following commands on the Foreman instance.
To start a job for a single accession code::
./run_management_command.sh survey_all --accession E-GEOD-3303You can also supply a newline-deliminated file which resides in S3 to survey_all which will
dispatch survey jobs based on accession codes like so:
./run_management_command.sh surveyor_dispatcher --file s3://data-refinery-test-assets/MY_BIG_LIST_OF_CODES.txtThe surveyor_dispatcher command will submit SurveyJobs to the AWS Batch queue, so it's more appropriate for running a large number of survey jobs in production.
See the Running Locally section for additional examples of survey_all usage.
Note that there is a run_management_command.sh included in the foreman directory that is completely different than the one that is created on the Foreman instance.
These two scripts share a name to make the commands work in either place.
All of the different Refine.bio subservices log to the same AWS CloudWatch Log
Group. If you want to consume these logs, you can use the awslogs tool, which
can be installed from pip like so:
pip install awslogsor, for OSX El Capitan:
pip install awslogs --ignore-installed sixOnce awslogs is installed, you can find your log group with:
awslogs groupsThen, to see all of the logs in that group for the past day, watching as they come in:
awslogs get <your-log-group> ALL --start='1 days' --watchYou can also apply a filter on these logs like so:
awslogs get <your-log-group> ALL --start='1 days' --watch --filter-pattern="DEBUG"Or, look at a named log stream (with or without a wildcard.) For instance: (Unfortunately this feature seems to be broken at the moment: jorgebastida/awslogs#158)
awslogs get data-refinery-log-group-myusername-dev log-stream-api-nginx-access-* --watchwill show all of the API access logs made by Nginx.
Automatic snapshots are created automatically by RDS. Manual database dumps can be created by priveledged users with these instructions. Postgres versions on the host (I suggest the PGBouncer instance) must match the RDS instance version:
sudo add-apt-repository "deb http://apt.postgresql.org/pub/repos/apt/ $(lsb_release -sc)-pgdg main"
wget --quiet -O - https://www.postgresql.org/media/keys/ACCC4CF8.asc | sudo apt-key add -
sudo apt-get update
sudo apt-get install postgresql-9.6Archival dumps can also be provided upon request.
Dumps can be restored locally by copying the backup.sql file to the volumes_postgres directory, then executing:
docker exec -it drdb /bin/bash
psql --user postgres -d data_refinery -f /var/lib/postgresql/data/backup.sqlThis can take a long time (>30 minutes)!
A stack that has been spun up via deploy.sh -u myusername -e dev can be taken down with destroy_terraform.sh -u myusername -e dev -d us-east-1.
The same username and environment must be passed into destroy_terraform.sh as were used to run deploy.sh either via the -e and -u options or by specifying TF_VAR_stage or TF_VAR_user so that the script knows which to take down.
Note that this will prompt you for confirmation before actually destroying all of your cloud resources.
Refine.bio is supported by Alex's Lemonade Stand Foundation, with some initial development supported by the Gordon and Betty Moore Foundation via GBMF 4552 to Casey Greene.
The table of contents for this README is generated using doctoc.
doctoc can be installed with: sudo npm install -g doctoc
Once doctoc is installed the table of contents can be re-generated with: doctoc README.md
BSD 3-Clause License.