Share
Embed
Transcript
57 slides
-
ENABLING REPRODUCIBLE IN-SILICO DATA ANALISES WITH NEXTFLOW Paolo Di
Tommaso, CRG Wellcome Trust Sanger Institute, 1 May 2018, Cambridge -
WHO IS THIS CHAP? @PaoloDiTommaso Research software engineer Comparative Bioinformatics,
Notredame Lab Center for Genomic Regulation (CRG) Author of Nextflow project -
AGENDA • The challenges with computational workflows • Nextflow main
principles • Handling parallelisation and portability • Deployments scenarios • Comparison with other tools • Future plans -
GENOMIC WORKFLOWS • Data analysis applications to extract information from
(large) genomic datasets • Embarrassingly parallelisation, can spawn 100s-100k jobs over distributed cluster • Mash-up of many different tools and scripts • Complex dependency trees and configuration → very fragile ecosystem -
To reproduce the result of a typical computational biology
paper requires 280 hours. ≈1.7 months! -
Platform Amazon Linux Debian Linux Mac OSX Number of chromosomes
36 36 36 Overall length (bp) 32,032,223 32,032,223 32,032,223 Number of genes 7,781 7,783 7,771 Gene density 236.64 236.64 236.32 Number of coding genes 7,580 7,580 7570 Average coding length (bp) 1,764 1,764 1,762 Number of genes with multiple CDS 113 113 111 Number of genes with known function 4,147 4,147 4,142 Number of t-RNAs 88 90 88 Comparison of the Companion pipeline annotation of Leishmania infantum genome executed across different platforms * * Di Tommaso P, et al., Nextflow enables computational reproducibility, Nature Biotech, 2017 -
CHALLENGES • Reproducibility, replicate results over time • Portability, run
across different platforms • Scalability ie. deploy big distributed workloads • Usability, streamline execution and deployment of complex workloads ie. remove complexity instead of adding new one • Consistency ie. track changes and revisions consistently for code, config files and binary dependencies -
HOW? • Fast prototyping 㱺 custom DSL that enables tasks
composition, simplifies most use cases + general purpose programming lang. for corner cases • Easy parallelisation 㱺 declarative reactive programming model based on dataflow paradigm, implicit portable parallelism • Self-contained 㱺 functional approach, a task execution is idempotent ie. cannot modify the state of other tasks + isolate dependencies with containers • Portable deployments 㱺 executor abstraction layer + deployment configuration from implementation logic -
process align_sample { input: file 'reference.fa' from genome_ch file 'sample.fq'
from reads_ch output: file 'sample.bam' into bam_ch script: """ bwa mem reference.fa sample.fq \ | samtools sort -o sample.bam """ } TASK EXAMPLE bwa mem reference.fa sample.fq \ | samtools sort -o sample.bam -
TASKS COMPOSITION process index_sample { input: file 'sample.bam' from bam_ch
output: file 'sample.bai' into bai_ch script: """ samtools index sample.bam """ } process align_sample { input: file 'reference.fa' from genome_ch file 'sample.fq' from reads_ch output: file 'sample.bam' into bam_ch script: """ bwa mem reference.fa sample.fq \ | samtools sort -o sample.bam """ } -
DATAFLOW • Declarative computational model for parallel process executions •
Processes wait for data, when an input set is ready the process is executed • They communicate by using dataflow variables i.e. async FIFO queues called channels • Parallelisation and tasks dependencies are implicitly defined by process in/out declarations -
HOW PARALLELISATION WORKS samples_ch = Channel.fromPath('data/sample.fastq') process FASTQC { input:
file reads from samples_ch output: file 'fastqc_logs' into fastqc_ch """ mkdir fastqc_logs fastqc -o fastqc_logs -f fastq -q ${reads} """ } -
samples_ch = Channel.fromPath('data/*.fastq') process FASTQC { input: file reads from
samples_ch output: file 'fastqc_logs' into fastqc_ch """ mkdir fastqc_logs fastqc -o fastqc_logs -f fastq -q ${reads} """ } HOW PARALLELISATION WORKS -
LOCAL EXECUTION • Common development scenario • Dependencies can be
managed using a container runtime • Parallelisations is managed spawning posix processes • Can scale vertically using fat server / shared mem. machine nextflow OS local storage docker/singularity laptop / workstation -
CENTRALISED ORCHESTRATION computer cluster • Nextflow orchestrates workflow execution submitting
jobs to a compute cluster eg. SLURM • It can run in the head node or a compute node • Requires a shared storage to exchange data between tasks • Ideal for corse-grained parallelisms NFS/Lustre cluster node cluster node cluster node cluster node submit jobs cluster node nextflow -
DISTRIBUTED ORCHESTRATION login node NFS/Lustre job request cluster node cluster
node launcher wrapper nextflow cluster nextflow driver nextflow worker nextflow worker nextflow worker HPC cluster • A single job request allocates the desired computes nodes • Nextflow deploys its own embedded compute cluster • The main instance orchestrate the workflow execution • The worker instances execute workflow jobs (work stealing approach) -
KUBERNETES • Next generation native cloud clustering for containerised workloads
• There's the need of workflow orchestration • Latest NF version includes a new command that streamline the workflow deployment to K8s -
PORTABILITY process { executor = 'slurm' queue = 'my-queue' memory
= '8 GB' cpus = 4 container = 'user/image' } -
PORTABILITY process { executor = 'awsbatch' queue = 'my-queue' memory
= '8 GB' cpus = 4 container = 'user/image' } -
GALAXY vs. NEXTFLOW • Command line oriented tool • Can
incorporate any tool w/o any extra adapter • Fine control over tasks parallelisation • Scalability 100㱺1M jobs • One liner installer • Suited for production workflows + experienced bioinformaticians • Web based platform • Built-in integration with many tools and dataset • Little control over tasks parallelisation • Scalability 10㱺1K jobs • Complex installation and maintenance • Suited for training + not experienced bioinformaticians -
SNAKEMAKE vs. NEXTFLOW • Command line oriented tool • Push
model • Can manage any data structure • Compute DAG at runtime • All major container runtimes • Built-in support for clusters and cloud • No (yet) support for sub-workflows • Built-in support for Git/GitHub, etc., manage pipeline revisions • Groovy/JVM based • Command line oriented tool • Pull model • Rules defined using file name patterns • Compute DAG ahead • Built-in support for Singularity • Custom scripts for cluster deployments • Support for sub-workflows • No support for source code management system • Python based -
CWL vs. NEXTFLOW • Language + app. runtime • DSL
on top of a general purpose programming lang. • Concise, fluent (at least try to be!) • Community driven • Single implementation, quick iterations • Language specification • Declarative meta-language (YAML/JSON) • Verbose • Committee driven • Many vendors/implementations (and specification version) -
CONTAINERISATION • Nextflow envisioned the use of software containers to
fix computational reproducibility • Mar 2014 (ver 0.7), support for Docker • Dec 2016 (ver 0.23), support for Singularity Nextflow job job job -
SINGULARITY FEATURES Kurtzer et al. Singularity: Scientific containers for mobility
of compute. PLoS ONE 12(5): e0177459 -
BENCHMARK* * Di Tommaso P, Palumbo E, Chatzou M, Prieto
P, Heuer ML, Notredame C. (2015) The impact of Docker containers on the performance of genomic pipelines. PeerJ 3:e1273 https://dx.doi.org/10.7717/peerj.1273 container execution can have an impact on short running tasks ie. < 1min -
SINGULARITY BENCHMARK https://github.com/wresch/python_import_problem Singularity image format speeds up Python execution
having many imports from a shared file system ! -
BEST PRACTICES • Helps to isolate dependencies from dev or
local deployment environment • Provides a reproducibles sandbox for third party users • Binary images preserve against software decay • Make it transparent ie. always include the Dockefile • Docker image format is de-facto standard, it can be executed by different runtime eg. Singularity, Shifter, uDocker, etc. -
ERROR RECOVERY • Each task outputs are saved in a
separate directory • This allows to safely record interrupted executions discarding • Dramatically simplify debugging ! • Computing resources can be defined in a *dynamic* manner, so that a failing task can be automatically re- execute with more memory, longer timeout, etc. -
IMPROVEMENTS • Built-in support for Bioconda recipies • Better meta-data
and provenance handling • Workflow composition aka sub-workflows • More clouds support ie. Azure and GCP -
APACHE SPARK • Native support for Apache Spark clusters and
execution model • Allow hybrid Nextflow and Spark applications • Mix the best of the two worlds, Nextflow for legacy tools/corse grain parallelisation and Spark for fine grain/distributed execution eg. GATK4 -
• Partecipate in Cloud Work Stream working group • TES:
Task Execution API (working prototype) • WES: Workflow Execution API • Enable interoperability with GA4GH complaint platforms eg. Cancer Genomics Cloud and Broad FireCloud -
• Community effort to collect production ready analysis pipelines built
with Nextflow • Initially supported by SciLifeLab, QBiC and A*Star Genome Institute Singapore • https://nf-core.github.io Alexander Peltzer Phil Ewels Andreas Wilm -
CONCLUSION • Data analysis reproducibility is hard and it's often
underestimated. • Nextflow does not provide a magic solution but enables best-practices and provide support for community and industry standards. • It strictly separates the application logic from the configuration and deployment logic, enabling self-contained workflows. • Applications can be easily deployed across different environment in a reproducible manner with a single command. • The functional/reactive model allows applications to scale to millions of jobs with ease.