Skip to content

Latest commit

 

History

History
234 lines (132 loc) · 9.49 KB

CHANGELOG.md

File metadata and controls

234 lines (132 loc) · 9.49 KB

Changelog

The project uses Semantic Versioning: major.minor.patch i.e., breaking.feature.bugfix.

We use annotated git tags for each release commit:

  1. Create a new tag locally: git tag -a <tag-name> -m"<annotation>" [commit] (Example for Head: git tag -a v1.0.0 -m "release 1.0.0")
  2. Push the tag to remote repo: git push origin tag <tag_name>
  3. Create a release from the tag on Github

[1.2.0] - 2024-08-09

Added

  • Add DOES_RESULTS_DIR environment variable to change the default results dir <DOES_PROJECT_DIR>/doe-suite-results (can be useful for artefact evaluation)

Changed

  • Extend etl_debug.py to include an option to specify overwrite_suite_id_map that will overwrite specific suite ids in a super_etl. The overwrite_suite_id_map can now also be used to filter pipeline.experiments in super etl designs.

[1.1.1] - 2024-05-30

Fixed

  • update Python dependencies
  • allow referencing suite names in (super) ETL that do not have a design but instead have an existing result.

1.1.0 - 2024-04-30

Added

  • Introduce Docker cloud (run suites locally):

    • uses the Docker API to create hosts as docker containers. cloud=docker
    • uses Ubunutu 20.04 LTS as default image (but custom Dockerfiles can be used)

  • Add docker-test.yml for CI via Github Action: Runs all the example design tests with cloud=docker within the Github checks.

Changed

  • runs the local tests with GNU parallel (instead of make -j).

Fixed

  • fixed non-determinism issues in the ETL of the test suite by adding an explicit sorting transformer in certain steps.

1.0.0 - 2024-04-16

Upgrade Process

Existing project should update the following in their doe-suite-config to use this version of doe-suite:

  • Add inventory folder containing euler.yml (can be found in the inventory dir inside the cookiecutter-doe-suite-config or the demo_project). See demo_project/doe-suite-config/inventory/euler.yml.

  • Update poetry.lock file in doe-suite-config using poetry update doespy, to update the lockfile with new dependencies added to the doespy project

  • [Optional] In order to use the new custom machine cloud backend, update the path definition in groups_vars/all/main.yml to:

    remote:
dir: "{{ remote_dir | default('/home/' + ansible_user | mandatory('ansible_user must be set') + '/doe-suite/' + prj_id + '/' + suite) }}"
results_dir: "{{remote_results_dir | default('/home/' + ansible_user + '/doe-suite-results') }}"

Added

  • Support cloud definitions based on an Ansible Inventory File:

    • In addition to running experiments on AWS and ETHZ Euler, an Ansible inventory file can now also define the remote experiment environment. This allows the doe-suite to run experiments on any set of machines that are reachable with SSH.

    • With introducing more potential clouds/backends, we need support for cloud-specific (potentially even host-specific) variables in the experiment design, e.g., defining a path. This is supported via the existing variable exp_host_lst, which already contained the dns_name/ip of all hosts in the experiment. Each host in this list now also has an entry hostvars, which provides access to all variables on that host. The new jinja2 filter [% 'test' | at_runtime(exp_host_lst) %] allows accessing the value of a variable at runtime. Without additional arguments, it will raise an error if the variable has different values on different hosts participating in the experiment.

    • The euler cloud is now also implemented as a cloud by inventory file. (see doe-suite/cookiecutter-doe-suite-config/{{cookiecutter.repo_name | default("doe-suite-config") }}/inventory/euler.yml)

    • We have renamed/removed a lot of euler specific task files (because now if experiments can be run not only on aws and euler).

  • Custom jinja filter in experiment design: To simplify the experiment design, we added the possibility to define project-specific custom Jinja2 filters under doe-suite-config/design/filter_plugins, which can then be used in the experiment design. For example, when accessing a host-specific (cloud-dependent) variable at runtime, we have an at_runtime Jinja filter.

  • Multi-Command runs (main and background tasks on same host): We support multi-command runs (see, example05-complex). We always have a main cmd, but then we can have an arbitrary number of background tasks that are bound to the lifetime of the main command. If the main command finishes, we also ensure that the background tasks are terminated.

  • Demo Notebook developing ETL steps: We added can example for developing etl steps (with caching the inputs to the step) demo_project/doe-suite-config/super_etl/debug_etl.ipynb.

  • Introduce ColCrossPlotLoader:

    • We add the ColCrossPlotLoader which is an advanced loader that allows building complex plots based on the data. (see the documentation for more details)
    • Add super etl examples that use the ColCrossPlotLoader (see doe-suite/demo_project/doe-suite-config/super_etl)
    • We added a new example design: example08-superetl, which generates some fake data that is used to showcase etl plotting features.

Changed

  • in setup roles for doe-suite, ensure that sudo is only required if a package is missing (otherwise) non-sudo access is sufficient.

  • update the way poetry is installed in role setup-small.

  • For extractors, the file_regex defaults can now be defined by overriding the class property.

  • By default, super etls have the output path doe-suite-results-super next to doe-suite-results

Fixed

  • in docs: for etl steps replace the broken automodule with manual docs.

  • Bugfix tsp pattern in case the command contains a [.

0.4.1 - 2024-01-23

Changed

  • update python dependencies

0.4.0 - 2024-01-23

Added

  • use pydantic for super etl (+ support include)

  • add save data option to plot loader

Changed

  • only run docs deploy action on main branch

Fixed

  • tsp status module consistency of jobs between servers: In tsp status module, add check to see if there is an inconsistent set of jobs between servers, which can happen if one of the servers loses the job in tsp for some reason. Before doe-suite would get in some kind of inexplicable loop, now it checks if there are inconsistencies. In some cases there are still issues when some jobs are still running, but this should fix most of the issues.

  • tsp_info regex: Update regex in tsp_info to handle negative exit code

0.3.0 - 2023-12-22

Added

  • Add support for $SUITE_ID$ dict replacement of super etl via command line: Each super ETL configuration starts with a $SUITE_ID$ dictionary, specifying the run ids for the results. With this update, we introduce support to replace this dictionary via the command line.

  • ETL Extractors now receive the run config as part of their options: This enhancement enables more intricate extractors to implement advanced functionalities, such as implementing filters during the extraction stage. As this does not change the Extractors' interface, we ensure backward compatibility.

0.2.0 - 2023-08-11

Added

  • New Experiment Design Feature except_filters: We added a basic implementation of #71 that allows filtering out certain combinations.

  • New Super ETL Feature: pipelines filter: We add the possibility when running a super_etl via make to run only a subset of pipelines (see pipelines="a b")

  • ETL Step: We add a new ETL Loader that allows storing a data frame as a pickle.

  • Option to save ETL results to Notion: We add a utility function that allows storing the results of a loader to a Notion page.

Changed

  • Optimize config/dir creation: As part of the setup, before running experiments, we create all working directories and place the config.json in the folder. Until now, this relied on pure Ansible. However, for many jobs, this creation becomes a bottleneck. Now, there is a custom module that does this more efficiently.

  • Optimize result fetching: We replaced the slow fetching of results with an additional custom module. More efficiency is possible because now tsp or slurm do not need to process a completing job id one-by-one. Instead, they can report that a list of job ids finished and can be downloaded.

  • ETL Extract Optimization: We only flatten results that require flattening, which speeds up the processing.

0.1.0 - 2023-05-16

Added

  • Documentation: start sphinx documentation

Changed

  • Design Validation: switch design validation to pydantic models (experiment design schema)