servicex-databinder

ServiceX DataBinder

Release v0.5.0

servicex-databinder is a user-analysis data management package using a single configuration file. Samples with external data sources (e.g. RucioDID or XRootDFiles) utilize ServiceX to deliver user-selected columns with optional row filtering.

The following table shows supported ServiceX transformers by DataBinder

Input format	Code generator	Transformer	Output format
ROOT Ntuple	func-adl	uproot	root or parquet
ATLAS Release 21 xAOD	func-adl	atlasr21	root
ROOT Ntuple	python function	python	root or parquet

Prerequisite

• Access to a ServiceX instance
• Python 3.7+

Installation

pip install servicex-databinder

Configuration file

The configuration file is a yaml file containing all the information.

The following example configuration file contains minimal fields. You can also download servicex-opendata.yaml file (rename to servicex.yaml) at your working directory, and run DataBinder for OpenData without an access token.

General:
  ServiceXName: servicex-opendata
  OutputFormat: parquet
  
Sample:
  - Name: ggH125_ZZ4lep
    XRootDFiles: "root://eospublic.cern.ch//eos/opendata/atlas/OutreachDatasets\
                  /2020-01-22/4lep/MC/mc_345060.ggH125_ZZ4lep.4lep.root"
    Tree: mini
    Columns: lep_pt, lep_eta

General block requires two mandatory options (ServiceXName and OutputFormat) as in the example above.

Input dataset for each Sample can be defined either by RucioDID or XRootDFiles or LocalPath.

ServiceX query can be constructed with either TCut syntax or func-adl.

• Options for TCut syntax: Filter¹ and Columns
• Option for Func-adl expression: FuncADL

      ¹ Filter works only for scalar-type of TBranch.

Output format can be either Apache parquet or ROOT ntuple for uproot backend. Only ROOT ntuple format is supported for xAOD backend.

The followings are available options:

Option for General block	Description	DataType
ServiceXName*	ServiceX backend name in your servicex.yaml file	String
OutputFormat*	Output file format of ServiceX delivered data (parquet or root for uproot / root for xaod)	String
Transformer	Set transformer for all Samples. Overwrites the default transformer in the servicex.yaml file.	String
Delivery	Delivery option; LocalPath (default) or LocalCache or ObjectStore	String
OutputDirectory	Path to a directory for ServiceX delivered files	String
WriteOutputDict	Name of an ouput yaml file containing Python nested dictionary of output file paths (located in the OutputDirectory)	String
IgnoreServiceXCache	Ignore the existing ServiceX cache and force to make ServiceX requests	Boolean

*Mandatory options

Option for Sample block	Description	DataType
Name	Sample name defined by a user	String
Transformer	Transformer for the given sample	String
RucioDID	Rucio Dataset Id (DID) for a given sample; Can be multiple DIDs separated by comma	String
XRootDFiles	XRootD files (e.g. root://) for a given sample; Can be multiple files separated by comma	String
Tree	Name of the input ROOT TTree; Can be multiple TTrees separated by comma (uproot ONLY)	String
Filter	Selection in the TCut syntax, e.g. jet_pt > 10e3 && jet_eta < 2.0 (TCut ONLY)	String
Columns	List of columns (or branches) to be delivered; multiple columns separately by comma (TCut ONLY)	String
FuncADL	Func-adl expression for a given sample	String
LocalPath	File path directly from local path (NO ServiceX tranformation)	String

A config file can be simplified by utilizing Definition block. You can define placeholders under Definition block, which will replace all matched placeholders in the values of Sample block. Note that placeholders must start with DEF_.

You can source each Sample using different ServiceX transformers. The default transformer is set by type of servicex.yaml, but Transformer in the General block overwrites if present, and Transformer in each Sample overwrites any previous transformer selection.

The following example configuration shows how to use each Options.

General:
  ServiceXName: servicex-uc-af
  Transformer: uproot
  OutputFormat: root
  OutputDirectory: /Users/kchoi/data_for_MLstudy
  WriteOutputDict: fileset_ml_study
  IgnoreServiceXCache: False
  
Sample:  
  - Name: Signal
    RucioDID: user.kchoi:user.kchoi.signalA,
              user.kchoi:user.kchoi.signalB,
              user.kchoi:user.kchoi.signalC
    Tree: nominal
    FuncADL: DEF_ttH_nominal_query
  - Name: Background1
    XRootDFiles: DEF_ggH_input
    Tree: mini
    Filter: lep_n>2
    Columns: lep_pt, lep_eta
  - Name: Background2
    Transformer: atlasr21
    RucioDID: DEF_Zee_input
    FuncADL: DEF_Zee_query
  - Name: Background3
    LocalPath: /Users/kchoi/Work/data/background3
  - Name: Background4
    Transformer: python
    RucioDID: user.kchoi:user.kchoi.background4
    Function: |
      def run_query(input_filenames=None):
          import awkward as ak, uproot
          tree_name = "nominal"
          o = uproot.lazy({input_filenames:tree_name})
          return {"nominal: o}

Definition:
  DEF_ttH_nominal_query: "Where(lambda e: e.met_met>150e3). \
              Select(lambda event: {'el_pt': event.el_pt, 'jet_e': event.jet_e, \
              'jet_pt': event.jet_pt, 'met_met': event.met_met})"
  DEF_ggH_input: "root://eospublic.cern.ch//eos/opendata/atlas/OutreachDatasets\
                  /2020-01-22/4lep/MC/mc_345060.ggH125_ZZ4lep.4lep.root"
  DEF_Zee_input: "mc15_13TeV:mc15_13TeV.361106.PowhegPythia8EvtGen_AZNLOCTEQ6L1_Zee.\
                merge.DAOD_STDM3.e3601_s2576_s2132_r6630_r6264_p2363_tid05630052_00"
  DEF_Zee_query: "SelectMany('lambda e: e.Jets(\"AntiKt4EMTopoJets\")'). \
              Where('lambda j: (j.pt() / 1000) > 30'). \
              Select('lambda j: j.pt() / 1000.0'). \
              AsROOTTTree('junk.root', 'my_tree', [\"JetPt\"])"

Deliver data

from servicex_databinder import DataBinder
sx_db = DataBinder('<CONFIG>.yml')
out = sx_db.deliver()

The function deliver() returns a Python nested dictionary that contains delivered files.

Input configuration can be also passed in a form of a Python dictionary.

Delivered Samples and files in the OutputDirectory are always synced with the DataBinder config file.

Error handling

failed_requests = sx_db.get_failed_requests()

If failed ServiceX request(s), deliver() will print number of failed requests and the name of Sample, Tree if present, and input dataset. You can get a full list of failed samples and error messages for each by get_failed_requests() function. If it is not clear from the message you can browse Logs in the ServiceX instance webpage for the detail.

Useful tools

Create Rucio container for multiple DIDs

The current ServiceX generates one request per Rucio DID. It's often the case that a physics analysis needs to process hundreds of DIDs. In such cases, the script (scripts/create_rucio_container.py) can be used to create one Rucio container per Sample from a yaml file. An example yaml file (scripts/rucio_dids_example.yaml) is included.

Here is the usage of the script:

usage: create_rucio_containers.py [-h] [--dry-run DRY_RUN]
                                  infile container_name version

Create Rucio containers from multiple DIDs

positional arguments:
  infile             yaml file contains Rucio DIDs for each Sample
  container_name     e.g. user.kchoi:user.kchoi.<container-name>.Sample.v1
  version            e.g. user.kchoi:user.kchoi.fcnc_ana.Sample.<version>

optional arguments:
  -h, --help         show this help message and exit
  --dry-run DRY_RUN  Run without creating new Rucio container

Acknowledgements

Support for this work was provided by the the U.S. Department of Energy, Office of High Energy Physics under Grant No. DE-SC0007890