1. LSST software stack.

  2. Shared filesystem for data.

  3. Shared database.

    SQLite3 is fine for small runs like ci_hsc_gen3 if have POSIX filesystem. For larger runs, use PostgreSQL.

  4. A workflow management service.

    Currently, two workflow management services are supported HTCondor’s DAGMan and Pegasus WMS. Both of them requires an HTCondor cluster. NCSA hosts a few of such clusters, see this page for details.

  5. HTCondor’s Python bindings (if using HTCondor) or Pegasus WMS.

Installing Batch Processing Service

Starting from LSST Stack version w_2020_45, the package providing Batch Processing Service, ctrl_bps, comes with lsst_distrib. However, if you’d like to try out its latest features, you may install a bleeding edge version similarly to any other LSST package:

git clone
cd ctrl_bps
setup -k -r .

Creating Butler repository

You’ll need a pre-existing Butler dataset repository containing all the input files needed for your run. This repository needs to be on the filesystem shared among all compute resources (e.g. submit and compute nodes) you use during your run.


Keep in mind though, that you don’t need to bootstrap a dataset repository for every BPS run. You only need to do it when Gen3 data definition language (DDL) changes, you want to to start a repository from scratch, and possibly if want to add/change inputs in repo (depending on the inputs and flexibility of the bootstrap scripts).

For testing purposes, you can use pipelines_check package to set up your own Butler dataset repository. To make that repository, follow the usual steps when installing an LSST package:

git clone
cd pipelines_check
git checkout w_2020_45  # checkout the branch matching the software branch you are using
setup -k -r .

Defining a submission

BPS configuration files are YAML files with some reserved keywords and some special features. They are meant to be syntactically flexible to allow users figure out what works best for them. The syntax and features of a BPS configuration file are described in greater detail in BPS configuration file. Below is just a minimal example to keep you going.

There are groups of information needed to define a submission to the Batch Production Service. They include the pipeline definition, the payload (information about the data in the run), submission and runtime configuration.

Describe a pipeline to BPS by telling it where to find either the pipeline YAML file (recommended)

pipelineYaml: "${OBS_SUBARU_DIR}/pipelines/DRP.yaml#processCcd"

or a pre-made file containing a serialized QuantumGraph, for example

qgraphFile: pipelines_check_w_2020_45.qgraph


The file with a serialized QuantumGraph is not portable. The file must be crated by the same stack being used when running BPS and it can be only used on the machine with the same environment.

The payload information should be familiar too as it is mostly the information normally used on the pipetask command line (input collections, output collections, etc).

The remaining information tells BPS which workflow management system is being used, how to convert Datasets and Pipetasks into compute jobs and what resources those compute jobs need.

Listing 2 ${CTRL_BPS_DIR}/doc/lsst.ctrl.bps/pipelines_check.yaml
pipelineYaml: "${OBS_SUBARU_DIR}/pipelines/DRP.yaml#processCcd"
templateDataId: "{tract}_{patch}_{band}_{visit}_{exposure}_{detector}"
project: dev
campaign: quick
submitPath: ${PWD}/submit/{outCollection}
computeSite: ncsapool
requestMemory: 2048
requestCpus: 1

# Make sure these values correspond to ones in the bin/'s
# pipetask command line.
  runInit: true
  payloadName: pcheck
  butlerConfig: ${PIPELINES_CHECK_DIR}/DATA_REPO/butler.yaml
  inCollection: HSC/calib,HSC/raw/all,refcats
  output: "u/${USER}/pipelines_check"
  outCollection: "{output}/{timestamp}"
  dataQuery: exposure=903342 AND detector=10

    runQuantumCommand: "${CTRL_MPEXEC_DIR}/bin/pipetask --long-log run -b {butlerConfig} -i {inCollection} --output {output} --output-run {outCollection} --init-only --register-dataset-types --qgraph {qgraphFile} --clobber-partial-outputs --no-versions"
    requestMemory: 8192

wmsServiceClass: lsst.ctrl.bps.wms.htcondor.htcondor_service.HTCondorService
clusterAlgorithm: lsst.ctrl.bps.quantum_clustering_funcs.single_quantum_clustering
createQuantumGraph: '${CTRL_MPEXEC_DIR}/bin/pipetask qgraph -d "{dataQuery}" -b {butlerConfig} -i {inCollection} -p {pipelineYaml} -q {qgraphFile}'
runQuantumCommand: "${CTRL_MPEXEC_DIR}/bin/pipetask --long-log run -b {butlerConfig} -i {inCollection} --output {output} --output-run {outCollection} --extend-run --skip-init-writes --qgraph {qgraphFile} --clobber-partial-outputs --no-versions"

Submitting a run

Submit a run for execution with

bps submit example.yaml

If submission was successfully, it will output something like this:

Submit dir: /home/jdoe/tmp/bps/submit/shared/pipecheck/20201111T13h34m08s
Run Id: 176261

Adding --log-level INFO option to the command line outputs more information especially for those wanting to watch how long the various submission stages take.

Checking status

To check the status of the submitted run, you can use tools provided by HTCondor or Pegasus, for example, condor_status or pegasus-status. To get a more pipeline oriented information use

bps report

which should display run summary similar to the one below

     RUNNING   0   176270 jdoe       dev   quick    pcheck     shared_pipecheck_20201111T14h59m26s

To see results regarding past submissions, use bps report --hist X where X is the number of days past day to look at (can be a fraction). For example

$ bps report --hist 1
   FAILED   0   176263 jdoe       dev   quick    pcheck     shared_pipecheck_20201111T13h51m59s
SUCCEEDED 100   176265 jdoe       dev   quick    pcheck     shared_pipecheck_20201111T13h59m26s

Use bps report --help to see all currently supported options.

Canceling submitted jobs

The bps command to cancel bps-submitted jobs is

bps cancel --id <id>

or to cancel all of your runs use

bps cancel --user <username>

For example

$ bps submit pipelines_check.yaml
Submit dir: /scratch/mgower/submit/u/mgower/pipelines_check/20210414T190212Z
Run Id: 369

$ bps report
X      STATE  %S        ID OPERATOR   PRJ        CMPGN                PAYLOAD              RUN
     RUNNING   0       369 mgower     dev        quick                pcheck               u_mgower_pipelines_check_20210414T190212Z

$ bps cancel --id 369
Successfully canceled: 369.0

$ bps report
X      STATE  %S        ID OPERATOR   PRJ        CMPGN                PAYLOAD              RUN


Sometimes there may be a small delay between executing cancel and jobs disappearing from the WMS queue. Under normal conditions this delay is less than a minute.

This command tries to prevent someone using it to cancel non-bps jobs. It can be forced to skip this check by including the option --skip-require-bps. Use this at your own risk.

If bps cancel says “0 jobs found matching arguments”, first double check the id for typos. If you believe there is a problem with the “is it a bps job” check, add --skip-require-bps.

If jobs are hanging around in the queue with an X status in condor_q, you can add the following to force delete those jobs from the queue

--pass-thru "-forcex"

If bps cancel fails to delete the jobs, you can use direct WMS executables like condor_rm or pegasus-remove.


Using the WMS commands directly under normal circumstances is not advised as bps may someday include additional code.

Both take the runId printed by bps submit. For example

condor_rm 176270       # HTCondor
pegasus-remove 176270  # Pegasus WMS

bps report also prints the runId usable by condor_rm.

If you want to just clobber all of the runs that you have currently submitted, you can just do the following no matter if using HTCondor or Pegasus WMS plugin:

condor_rm <username>

BPS configuration file

The configuration file is in YAML format. One particular YAML syntax to be mindful about is that boolean values must be all lowercase.

Configuration file can include other configuration files using includeConfigs with YAML array syntax. For example

  - bps-operator.yaml
  - bps-site-htcondor.yaml

Values in the configuration file can be defined in terms of other values using {key} syntax, for example

patch: 69
dataQuery: patch = {patch}

Environment variables can be used as well with ${var} syntax, for example

submitRoot: ${PWD}/submit
runQuantumExec: ${CTRL_MPEXEC_DIR}/bin/pipetask


Note the difference, $ (dollar sign), when using an environmental variable, e.g. ${foo}, and plain config variable {foo}.

Section names can be used to store default settings at that concept level which can be overridden by settings at more specific concept levels.  Currently the order from most specific to general is: payload, pipetask, and site.

description of the submission including definition of inputs
subsections are pipetask labels where can override/set runtime settings for particular pipetasks (currently no Quantum-specific settings).

settings for specific sites can be set here. Subsections are site names which are matched to computeSite. The following are examples for specifying values needed to match jobs to glideins.

HTCondor plugin example:

        requirements: "(GLIDEIN_NAME == &quot;test_gname&quot;)"
        +GLIDEIN_NAME: "test_gname"

Pegasus plugin example:

    arch: x86_64
    os: LINUX
        path: /work/shared-scratch/${USER}
          operation: all
          url: file:///work/shared-scratch/${USER}
        style: condor
        auxillary.local: true
        universe: vanilla
        getenv: true
        requirements: '(ALLOCATED_NODE_SET == &quot;${NODESET}&quot;)'
        +JOB_NODE_SET: '&quot;${NODESET}&quot;'
        retry: 0
        PEGASUS_HOME: /usr/local/pegasus/current

Supported settings

Location of the Butler configuration file needed by BPS to create run collection entry in Butler dataset repository
A label used to group submissions together. May be used for grouping submissions for particular deliverable (e.g., a JIRA issue number, a milestone, etc.). Can be used as variable in output collection name. Displayed in bps report output.
Algorithm to use to group Quanta into single Python executions that can share in-memory datastore. Currently, just uses single quanta executions, but this is here for future growth.
Specification of the compute site where to run the workflow and which site settings to use in bps prepare).
The command line specification for generating QuantumGraphs.
Name of the Operator who made a submission. Displayed in bps report output. Defaults to the Operator’s username.
Location of the YAML file describing the science pipeline.
Another label for groups of submissions. May be used to differentiate between test submissions from production submissions. Can be used as a variable in the output collection name. Displayed in bps report output.
requestMemory, optional
Amount of memory, in MB, a single Quantum execution of a particular pipetask will need (e.g., 2048).
requestCpus, optional
Number of cpus that a single Quantum execution of a particular pipetask will need (e.g., 1).
Used when giving names to graphs, default names to output files, etc.  If not specified by user, BPS tries to use outCollection with ‘/’ replaced with ‘_’.
Directory where the output files of bps prepare go.
The command line specification for running a Quantum. Must start with executable name (a full path if using HTCondor plugin) followed by options and arguments. May contain other variables defined in the configuration file.

Whether to add a pipetask --init-only to the workflow or not. If true, expects there to be a pipetask section called pipetaskInit which contains the runQuantumCommand for the pipetask --init-only. For example

  runInit: true

    runQuantumCommand: "${CTRL_MPEXEC_DIR}/bin/pipetask --long-log run -b {butlerConfig} -i {inCollection} --output {output} --output-run {outCollection} --init-only --register-dataset-types --qgraph {qgraphFile} --clobber-partial-outputs --no-versions"
    requestMemory: 2048

The above example command uses both --output and --output-run. The --output option creates the chained collection if necessary and defines it to include both the --input and --output-run collections. The --output-run option saves the unique run collection that is also passed to all other compute jobs (i.e., one run collection per submission). If using both here, must include both --output and --output-run in the other runQuantumCommand.

Template to use when creating job names (and HTCondor plugin then uses for job output filenames).

Workload Management Service plugin to use. For example

wmsServiceClass: lsst.ctrl.bps.wms.htcondor.htcondor_service.HTCondorService  # HTCondor
Whether to put full submit-time path to QuantumGraph file in command line because the WMS plugin requires shared filesystem. Defaults to False. HTCondor and Pegasus plugins do not need this value.

When to output job QuantumGraph files (default = TRANSFORM).

  • NEVER = all jobs will use full QuantumGraph file. (Warning: make sure runQuantumCommand has --qgraph-id {qgraphId} --qgraph-node-id {qgraphNodeId}.)
  • TRANSFORM = Output QuantumGraph files after creating GenericWorkflow.
  • PREPARE = QuantumGraph files are output after creating WMS submission.

A boolean flag. If set to true, BPS will save serialized clustered quantum graph to a file called bps_clustered_qgraph.pickle using Python’s pickle module. The file will be located in the submit directory. By default, it is set to false.

Setting it to true will significantly increase memory requirements when submitting large workflows as pickle constructs a complete copy of the object in memory before it writes it to disk. [ref]

A boolean flag. If set to true, BPS will save serialized generic workflow called to a file called bps_generic_workflow.pickle using Python’s pickle module. The file will be located in the submit directory. By default, it is set to false.

A boolean flag. If set to true, BPS will generate graphical representations of both the clustered quantum graph and the generic workflow in DOT format. The files will be located in the submit directory and their names will be and, respectively. By default, it is set to false.

It is recommended to use this option only when working with small graphs/workflows. The plots will be practically illegible for graphs which number of nodes exceeds order of tens.

Reserved keywords


Name of the file with a pre-made, serialized QuantumGraph.

Such a file is an alternative way to describe a science pipeline. However, contrary to YAML specification, it is currently not portable.

Internal ID for the full QuantumGraph (passed as --qgraph-id on pipetask command line).
Comma-separated list of internal QuantumGraph node numbers to be executed from the full QuantumGraph (passed as --qgraph-node-id on pipetask command line).
Created automatically by BPS at submit time that can be used in the user specification of other values (e.g., in output collection names so that one can repeatedly submit the same BPS configuration without changing anything)


Any values shown in the example configuration file, but not covered in this section are examples of user-defined variables (e.g. inCollection) and are not required by BPS.

QuantumGraph Files

BPS can be configured to either create per-job QuantumGraph files or use the single full QuantumGraph file plus node numbers for each job. The default is using per-job QuantumGraph files.

To use full QuantumGraph file, the submit YAML must set whenSaveJobQgraph to “NEVER” and the pipetask run command must include --qgraph-id {qgraphId} --qgraph-node-id {qgraphNodeId}. For example:

whenSaveJobQgraph: "NEVER"
runQuantumCommand: "${CTRL_MPEXEC_DIR}/bin/pipetask --long-log run -b {butlerConfig} -i {inCollection} --output {output} --output-run {outCollection} --extend-run --skip-init-writes --qgraph {qgraphFile} --qgraph-id {qgraphId} --qgraph-node-id {qgraphNodeId} --clobber-partial-outputs --no-versions"

– warning:

Do not modify the QuantumGraph options in pipetaskInit's runQuantumCommand.  It needs the entire QuantumGraph.

– note:

If running on a system with a shared filesystem, you'll more than likely want to also set bpsUseShared
to true.


Where is stdout/stderr from pipeline tasks?

For now, stdout/stderr can be found in files in the submit run directory.


The names are of the format:

<run submit dir>/jobs/<task label>/<quantum graph nodeNumber>_<task label>_<templateDataId>[.<htcondor job id>.[sub|out|err|log]

Pegasus WMS

Pegasus does its own directory structure and wrapping of pipetask output.

You can dig around in the submit run directory here too, but try pegasus-analyzer command first.

Advanced debugging

Here are some advanced debugging tips:

  1. If bps submit is taking a long time, probably it is spending the time during QuantumGraph generation.  The QuantumGraph generation command line and output will be in quantumGraphGeneration.out in the submit run directory, e.g. submit/shared/pipecheck/20200806T00h22m26s/quantumGraphGeneration.out.

  2. Check the *.dag.dagman.out for errors (in particular for ERROR: submit attempt failed).

  3. The Pegasus runId is the submit subdirectory where the underlying DAG lives.  If you’ve forgotten the Pegasus runId needed to use in the Pegasus commands try one of the following:

    1. It’s the submit directory in which the braindump.txt file lives.  If you know the submit root directory, use find to give you a list of directories to try.  (Note that many of these directories could be for old runs that are no longer running.)o

      find submit  -name "braindump.txt"
    2. Use HTCondor commands to find submit directories for running jobs

      condor_q -constraint 'pegasus_wf_xformation == "pegasus::dagman"' -l | grep Iwd