caliper 0.0.13

Caliper is a tool for measuring and assessing change in packages.

Caliper

Caliper is a tool for measuring and assessing change in packages.

Getting Started

Installation

You can easily install from pypi:

pip install caliper

Concepts

• Manager: a handle to interact with a package manager
• Extractor: a controller to use a manager to extract metrics of interest
• Analysis: A caliper analysis means attempting to build containers across versions of a library, and run against scripts for tests to assess functionality.
• Version repository: a repository created by an extractor that tagged commits for package releases
• Metrics: are a type of classes that can extract a single timepoint, or a change over time (e.g., lines changed). You can see example metrics that have been extracted under examples/metrics or in the vsoch/caliper-metrics repository.

Managers

A manager is a handle to interact with a package.

Pypi

The first kind of package we are interested in is one from pypi. We might quickly extract all metrics to an output folder from the command line for a Pypi package:

caliper extract --outdir caliper-metrics/ pypi:sregistry

Pypi Details

or we can instantiate a manager from Python, and walk through the steps that the client takes. First we create the manager.

from caliper.managers import PypiManager
manager = PypiManager("sregistry")

The manager specs include the source archive, version, and hash for each version of the package. The schema of the spec is a subset of the spack package schema. Every manager exposes this metadata.

manager.specs[0]
Found 82 versions for sregistry
{'name': 'sregistry',
 'version': '0.0.1',
 'source': {'filename': 'https://files.pythonhosted.org/packages/ef/2f/ccc36e816dc081abbe0932c422586eda868719025ec07ac206ed254d6a3c/sregistry-0.0.1.tar.gz',
  'type': 'source'},
 'hash': 'd4ee6933321b5a3da13e0b1657ca74f90477f670e59096a6a0a4dbb30a0b1f07'}

manager.specs[-1]
{'name': 'sregistry',
 'version': '0.2.36',
 'source': {'filename': 'https://files.pythonhosted.org/packages/75/6c/2b5bcf0191c0ddc9b95dd156d827c8d80fa8fe86f01f7a053fdd97eaea41/sregistry-0.2.36.tar.gz',
  'type': 'source'},
 'hash': '238ebd3ca0e0408e0be6780d45deca79583ce99aed05ac6981da7a2b375ae79e'}

If you just interact with manager.specs, you'll get a random architecture for each one. This can be okay if you want to do static file analysis, but if you want to choose a specific python version, your best bet is to call the get package metadata function directly and provide your preferences. For example, here we want Tensorflow for Python 3.5 and a specific linux architecture:

manager.get_package_metadata(python_version="35", arch="manylinux1_x86_64")

To derive these search strings, you can look at examples of wheels provided. This isn't the default because not all packages provide such rich choices. Here is an example from an early version of tensorflow.

tensorflow-0.12.0-cp27-cp27m-macosx_10_11_x86_64.whl
tensorflow-0.12.0-cp27-cp27mu-manylinux1_x86_64.whl
tensorflow-0.12.0-cp34-cp34m-manylinux1_x86_64.whl
tensorflow-0.12.0-cp35-cp35m-macosx_10_11_x86_64.whl
tensorflow-0.12.0-cp35-cp35m-manylinux1_x86_64.whl
tensorflow-0.12.0-cp35-cp35m-win_amd64.whl

For more recent versions you would see Python 3.8 and 3.9, and definitely not 2.x. The above function still selects one release based on your preferences. You can also choose to return a subset of all versions with the filter function. For example, here let's narrow down the set to include those that can be installed on Linux.

releases = manager.filter_releases('manylinux1_x86_64')

You can also get a set of unique Python versions across packages:

python_versions = manager.get_python_versions()
# {'cp27', 'cp33', 'cp34', 'cp35', 'cp36', 'cp37', 'cp38'}

Not all package versions are guaranteed to have these Python versions, but that's something interesting to consider. And you can always interact with the raw package metadata at manager.metadata.

GitHub

We might also be interested in releases from GitHub. Extracting metrics from the command line would look like this:

caliper extract --outdir caliper-metrics/ github:vsoch/pull-request-action

GitHub Details

And we could do the same steps as above (as we did with the pypi manager to create an interactive manager client.

from caliper.managers import GitHubManager
manager = GitHubManager("vsoch/pull-request-action")

GitManager

A GitManager is a special kind of manager that exists to interact with a git repository. It will be possible to use it as a manager proper (not yet developed) but it can also serve to create and interact with local git repositories. For example, let's create a temporary directory, add stuff to it, commit and then tag it.

from caliper.managers import GitManager
import tempfile
git = GitManager(tempfile.mkdtemp())
git.init()

# write some content (file.txt)

git.add("file.txt")
git.commit("Adding new content!")
git.tag("tag")

Note that when you run git.init() a dummy username and email will be added to the .git/config file so we can continue interactions without needing a global setting. This is done intentionally based on the idea that the user likely won't keep the version repository, however if you do want to keep it, feel free to change or remote these settings in favor of global ones.

You can imagine how this might be used - we can have a class that can take a manager, and then iterate over versions/releases and create a tagged commit for each. We can then easily extract metrics about files changed between versions. This is the metrics extractor discussed next.

Caliper Analyze

Caliper supports analyzing package functionality, which means that we take a configuration file (see an example), a caliper.yaml with a package name, manager, Dockerfile template to build, and a list of tests. We do this with the Caliper analyze command:

$ caliper analyze --help
usage: caliper analyze [-h] [--config CONFIG] [--no-progress] [--serial] [--force] [--nprocs NPROCS]

optional arguments:
  -h, --help       show this help message and exit
  --config CONFIG  A caliper.yaml file to use for the analysis (required)
  --no-progress    Do not show a progress bar (defaults to unset, showing progress)
  --serial         Run in serial instead of parallel
  --force          If an output file exists, force re-write (default will not overwrite)
  --nprocs NPROCS  Number of processes. Defaults to cpu count.

For example, we might use the example and do:

$ caliper analyze --config examples/analyze/caliper.yaml 

to do a docker system prune --all after each build (recommended) add --cleanup

$ caliper analyze --config examples/analyze/caliper.yaml --cleanup

And if your caliper.yaml is in the same folder as you are running caliper from, you don't need to supply it (it will be auto-detected):

caliper analyze --cleanup

and run the builds in serial. A parallel argument is supported, but in practice it doesn't work well building multiple containers at once.

caliper.yml

The caliper.yml file is a small configuration file to run caliper. Currently, it's fairly simply and we need to define the dependency to run tests over (e.g., tensorflow), the Dockerfile template, a name, and then a list of runs:

analysis:
  name: Testing tensorflow
  packagemanager: pypi
  dockerfile: Dockerfile
  dependency: tensorflow
  versions:
    - 0.0.11
  python_versions:
    - cp27
  tests:
    - tensorflow_v0.11/5_MultiGPU/multigpu_basics.py
    - tensorflow_v0.11/1_Introduction/basic_operations.py
    - tensorflow_v0.11/1_Introduction/helloworld.py
    - tensorflow_v0.11/4_Utils/tensorboard_advanced.py

If you don't define a list of python_versions all will be used by default. If you don't define a list of versions all versions of the library will be tested. If you want to add custom arguments for your template (beyond a base image that is derived for your Python software, and the dependency name to derive wheels to install) you can do this with args:

analysis:
  name: Testing tensorflow
  packagemanager: pypi
  dockerfile: Dockerfile
  args:
     additionaldeps: 
       - scikit-learn

The functionality of your arguments is up to you. In the example above, additionaldeps would be a list, so likely you would loop over it in your Dockerfile template (which uses jinja2).

Dockerfile

The Dockerfile template (specified in the caliper.yaml) should expect the following arguments from the caliper analysis script:

• base: The base python image, derived from the wheel we need to install
• filename: the url filename of the wheel to download with wget
• basename: the basename of that to install with pip

Additional arguments under args will be handed to the template, and are up to you to define and render appropriately.

Metrics Extractor

Finally, a metrics extractor provides an easy interface to iterate over versions of a package, and extract some kind of metric. There are two ways to go about it - starting with a repository that already has tags of interest, or starting with a manager that will be used to create it.

Extraction Using Client

When installed, caliper comes with an executable, caliper that can make it easy to extract a version repository.

$ caliper --help
usage: caliper [-h] [--version] {version,metrics,analyze,extract,view} ...

Caliper is a tool for measuring and assessing changes in packages.

optional arguments:
  -h, --help            show this help message and exit
  --version             suppress additional output.

actions:
  actions

  {version,metrics,analyze,extract,view}
                        actions
    version             show software version
    metrics             see metrics available
    analyze             analyze functionality of a package.
    extract             extract one or more metrics for a software package.
    view                extract a metric and view a plot.

LOGGING:
  --quiet               suppress logging.
  --verbose             verbose output for logging.
  --log-disable-color   Disable color for snakeface logging.
  --log-use-threads     Force threads rather than processes.

The extract command allows to extract metrics for a package:

$ caliper extract --help
usage: caliper extract [-h] [--metric METRIC] [--outdir OUTDIR] [--no-cleanup] [packages [packages ...]]

positional arguments:
  packages         packages to extract, e.g., pypi, GitHub, or (eventually) spack.

optional arguments:
  -h, --help       show this help message and exit
  --metric METRIC  one or more metrics to extract (comma separated), defaults to all metrics
  --outdir OUTDIR  output directory to write files (defaults to temporary directory)
  --no-cleanup     do not cleanup temporary extraction repositories.
  --force          if a metric file exists, do not overwrite.

But first we might want to see metrics available:

$ caliper metrics
         totalcounts: caliper.metrics.collection.totalcounts.metric.Totalcounts
        changedlines: caliper.metrics.collection.changedlines.metric.Changedlines

Let's say we want to extract the changedlines metric for a pypi repository, sif, which will return insertions, deletions, and overall change for each tagged release. That would look like this:

caliper extract --metric changedlines pypi:sif
Found 2 versions for sif
Repository for [manager:sif] is created at /tmp/sif-26hqifbm
Results written to /tmp/caliper-p633odvg

By default, if you don't specify an output directory, the metrics will be saved to the present working directory. The organizaion is by package type, name, and then results files:

$ tree /tmp/caliper-p633odvg
/tmp/caliper-p633odvg
└── pypi
    └── sif
        └── changedlines
            ├── changedlines-file-results.json
            └── changedlines-summed-results.json

3 directories, 2 files

but you can instead save to an output folder of your choosing (with the same structure).

mkdir -p examples/metrics/
caliper extract --metric changedlines --outdir examples/metrics/ pypi:sif
Found 2 versions for sif
Repository for [manager:sif] is created at /tmp/sif-0vpe767q
Results written to examples/metrics/

These files are provided for inspection under examples/metrics. For a change metric (a type that looks at change across tagged commits) you'll see a range of version like EMPTY..0.0.1. For a metric specific to a commit you will see just the tag (e.g., 0.0.1).

Extraction Using Manager

The manager knows all the files for a release of some particular software, so we can use it to start an extraction. For example, let's say we have the Pypi manager above:

from caliper.managers import PypiManager
manager = PypiManager("sregistry")

manager
# [manager:sregistry]

We can then hand it off to the extractor:

from caliper.metrics import MetricsExtractor
extractor = MetricsExtractor(manager)

# This repository will have each release version represented as a tagged commit
repo = extractor.prepare_repository()

...
[master b45263b] 0.2.34
 8 files changed, 60 insertions(+), 65 deletions(-)
[master 555962b] 0.2.35
 4 files changed, 4 insertions(+), 4 deletions(-)
[master ead9302] 0.2.36
 117 files changed, 141 insertions(+), 141 deletions(-)
Repository for [manager:sregistry] is created at /tmp/sregistry-j63wuvei

At this point you'll see the extractor iterating through each repository version, and commiting changes based on the version. It's fun to open the repository folder (in /tmp named based on the package) and watch the changes happening in real time. At this point we would have our version repository that we can calculate metrics over. For example, we can see commits that correspond to versions:

$ git log
commit ead9302cec47e62f8eabc5aefc0e55eeb3b8d717 (HEAD -> master, tag: 0.2.36)
Author: vsoch <[e-mail]>
Date:   Fri Dec 18 14:51:34 2020 -0700

    0.2.36

commit 555962bad5f9e6d0d8ea4c4ea6bb0bdcb92d12f3 (tag: 0.2.35)
Author: vsoch <[e-mail]>
Date:   Fri Dec 18 14:51:34 2020 -0700

    0.2.35

commit b45263b9c4da6aef096d49cc222bb9c64d2f6997 (tag: 0.2.34)
Author: vsoch <[e-mail]>
Date:   Fri Dec 18 14:51:34 2020 -0700

    0.2.34

commit 113bc796acbffc593d400a19471c56c36289d764 (tag: 0.2.33)
Author: vsoch <[e-mail]>
Date:   Fri Dec 18 14:51:33 2020 -0700
...

We can see the tags:

$ git tag
0.0.1
0.0.2
0.0.3
...
0.2.34
0.2.35
0.2.36

This is really neat! Next we can use the extractor to calculate metrics.

Extraction from Existing

As an alternative, if you create a repository via a manager (or have another repository you want to use that doesn't require one) you can simply provide the working directory to the metrics extractor:

from caliper.metrics import MetricsExtractor
extractor = MetricsExtractor(working_dir="/tmp/sregistry-j63wuvei")

You can see that we've created a git manager at this root:

extractor.git
<caliper.managers.git.GitManager at 0x7ff92a66ca60>

And we then might want to see what metrics are available for extraction.

extractor.metrics
{'totalcounts': 'caliper.metrics.collection.totalcounts.metric.Totalcounts',
 'changedlines': 'caliper.metrics.collection.changedlines.metric.Changedlines'}

Without going into detail, there are different base classes of metrics - a MetricBase expects to extract some metric for one timepoint (a tag/commit) and a ChangeMetricBase expects to extract metrics that compare two of these timepoints. The metric changedlines above is a change metric, and totalcounts is a base metric (for one commit timepoint). We can then run the extraction:

extractor.extract_metric("changedlines")

Note that you can also extract all metrics known to the extractor.

extractor.extract_all()

Parsing Results

For each extractor, you can currently loop through them and extract either data on the level of individual files, or summary results:

for name, metric in extractor:
    # Changedlines <caliper.metrics.collection.changedlines.metric.Changedlines at 0x7f7cd24f4940>

    # A lookup with file level changes
    metric.get_file_results()

    # A lookup with group or summed changed
    metric.get_group_results()

    # A lookup with "by-file" and "by-group" that includes both the above
    metric.get_results()

For example, an entry in group results might look like this:

{'0.2.34..0.2.35': {'size': 0, 'insertions': 4, 'deletions': 4, 'lines': 8}}

To say that between versions 0.2.34 and 0.2.35 there were 4 insertions, 4 deletions, and 8 lines changed total, and there was no change in overall size. We will eventually have more examples for how to parse and use this data.

Metrics View

To extract and view metrics, you can use caliper view

usage: caliper view [-h] [--metric METRIC] [--title TITLE] [--outdir OUTDIR] [--force] input

positional arguments:
  input            input data file to visualize.

optional arguments:
  -h, --help       show this help message and exit
  --metric METRIC  a metric to extract
  --title TITLE    the title for the graph (defaults to one set by metric)
  --outdir OUTDIR  output directory to write files (defaults to temporary directory)
  --force          if a file exists, do not overwrite.

For example, let's say we want to view an already extracted metric. We would provide the file as input:

$ caliper view ../caliper-metrics/github/spack/spack/changedlines/changedlines-results.json

We might also add a custom title:

$ caliper view ../caliper-metrics/github/spack/spack/changedlines/changedlines-results.json --title "Spack Version Changes"

Note that caliper will attempt to derive the metric name from the file. If you've renamed the file, then you'll need to provide it directly:

$ caliper view --metric changedlines mystery-file.json

Note from the usage that you can also select an output directory. Caliper tries to derive the name of the metric from the filename (e.g., <metric>-results.json however if you rename the file, you can specify the metric directly with --metric. You can see an example in docs. We expect to have more examples when we improve the documentation.

Use Cases

Assess Version Changes

Using the MetricsExtractor, we can start with a package and then calculate metrics for each version change, and ask questions like:

• What is the degree of change between minor/major versions?
• How much do dependencies change over time?
• How quickly does the package grow?

We might then be able to say that one package is more volatile than another, and use the metrics in other kinds of analyses.

Break a Workflow

An interesting use case for caliper is to use metrics to figure out if we can predict breaking. For example, we might have:

1. A Dockerfile with an entrypoint and command that generates some output
2. A list of requirements defined in a requirements.txt file (or similar)

And then we might derive a baseline container to run the workflow in question, and then vary package versions to determine if the container is still able to run and produce the same result, or if the dependency cannot be resolved all together. We can then assess, based on ranges of package versions that work vs. not and the degree of changes for each:

1. The degree to which some version increment is likely to break a build or run and
2. How close the developer was to representing a "correct" set of versions.

"Correct" is in quotes because we cannot easily assess dependency interaction (but perhaps can make some kind of proxy for it eventually).

Note this is all still being developed, and likely to change!

TODO

• create official docs in docs folder alongside code
• write tests to discover and test all metrics (type, name, etc.)
• think of common functions to run metric
• create visualization of grid of errors / scripts, and mouesover to see codes (server rendered)
• stats: try to estimate types of codes (e.g., error name)

Analysis Ideas

1. Start with a bunch of unit tests for a library (or example scripts) and build a model that can predict success (return code 0) vs fail (any other code) based on the scripts (tokens is a simple idea). Then given a new script without declared versions, predict which will work.
2. (A slightly different project) - given a set of known "optimal" containers for a library like tensorflow, parse dependencies, versions, and library versions, run across same set of unit tests / example scripts, and try to say which combos (architecture and dependencies) works for different kinds of scripts.

License

• Free software: MPL 2.0 License