Release 2.0.0

[AlessandroPomponio]

🏅 Highlights

• Major CLI overhaul — Several deprecated commands have been removed and renamed as part of the 2.0 cleanup: ado show summary, ado show details, ado get measurementrequest, and the legacy migrator in ado upgrade are gone. --query is renamed to --filter, and ado show entities is now ado show measurements. See the migration guide for details.
• Provenance & versioning — ado now records distribution package provenance on resource creation and ships full experiment and operator versioning support, giving every measurement a clear traceable lineage.
• Rich statistics everywhere — A new ado show stats command and -o stats output mode are available across operations, spaces, sample stores, and data containers, making it easy to get a quick quantitative summary of any resource.
• Trace support — ado show trace lets you inspect the execution trace of a space or store, surfacing how resources relate across operations.

Migration guide for ado 2.0.0: https://ibm.github.io/ado/migration/1x-to-2x/

✨ Features

CLI enhancements

• Statistics output — ado show stats provides aggregated statistics across resource types. A new -o stats flag is supported by ado get operation, ado get space, ado get samplestore, and ado get datacontainer.
• Trace commands — ado show trace and ado show trace space/store expose the execution provenance graph for spaces and stores.
• Non-interactive patching — ado edit now supports --patch and --patch-file flags for scripted, non-interactive metadata updates.
• --use-latest in ado get — Convenience flag extended to ado get commands so the most recent resource is selected automatically.

Core improvements

• StandardActuator base class — A new base class simplifies authoring actuator plugins with minimal boilerplate.
• Experiment and operator versioning — Full versioning support for experiments and operators, with an upgrade mechanism for orchestrator references in modules.
• Provenance recording — Distribution package provenance is captured at resource creation time and exposed through ado provenance.
• Statistics API — New functions to calculate statistics for operations (including multi-operation aggregation), spaces, sample stores, and data containers, surfaced through DiscoverySpaceStatistics.
• Measurement filtering — Sample stores now support server-side filtering of measurements.
• Script operation support — DiscoverySpace gains support for script-based operations.
• Improved Ray resilience — Detection of Ray measurement tasks that never start, plus new remote runtime setup options; improved SIGTERM shutdown handling.
• Conditional plugin model validation — Resources can now disable plugin model validation when needed.
• Multi-operation query support — operation_measurement_statistics and measurement_requests_for_operation both accept multiple operation IDs.
• Relationship queries — New get_resources_by_relationship function replaces the legacy camelCase equivalents.

🐛 Fixes

• CLI deletion — ado delete no longer truncates output when related resources prevent deletion.
• Python 3.10 compatibility — Fixed scanning of RECORD files for distributions on Python 3.10.
• SIGTERM handling — Operation shutdown is now handled correctly when SIGTERM is received mid-run.

🧰 Build

• Entry-point-based actuator loading — Actuators are now loaded via Python entry points (ado.actuators), replacing the previous discovery mechanism.
• Decoupled plugin versioning — All in-tree plugins (anomalous-series, autoconf, cplex-mip, example-actuator, profile-space, ray-tune, sfttrainer, trim, vllm-performance) now have independent version tags and are no longer coupled to the ado-core version.
• Python 3.14 support — ado now supports Python 3.14.
• Pinned Python 3.11 — CI is pinned to Python 3.11 as the baseline test version.
• Git tag handling — Forward slashes in git tags are handled correctly for component versioning.

📝 Docs

• Migration guide — A new migration guide is published on the website covering all breaking changes in 2.0.
• Installation instructions — Updated to reflect the new entry-point plugin model.
• JOSS paper — A citation file and JOSS badge have been added to the repository.
• Website updates — README, website index, local context example, and roadmap have all been refreshed.

🧹 Refactoring

The following are breaking changes in this release:

• BREAKING --query renamed to --filter in all ado get and ado show commands.
• BREAKING ado show details removed — use ado show with the appropriate resource type.
• BREAKING ado show summary removed — replaced by ado show stats.
• BREAKING Legacy migrator system in ado upgrade removed.
• BREAKING ado get measurementrequest removed.
• BREAKING ado show entities renamed to ado show measurements.
• BREAKING Import package renamed from orchestrator to ado — any code importing from orchestrator must be updated.

Internal refactoring (no user impact):

• dict_to_measurements renamed to observed_property_values_from_dict.
• search operations renamed to explore internally.
• getRelatedResourceIdentifiers / getRelatedResources replaced by get_resources_by_relationship.
• Sample store instantiation moved to class methods.
• configuration_model_default renamed to example_configuration in the plugin interface.
• Experiment ID parsing unified across CLI commands.

⚙️ Autoconf

• Fixed a bug where avoid_oom_recommender did not correctly consider the original number_gpus when evaluating GPU configurations.
• Fixed an incorrect model mapping for Llama 3.1-8b.
• Internal: avoid_oom_recommender now skips GPU counts fewer than the original GPU count.

🧠 SFTTrainer

• Weight download helper no longer uses Ray for execution, improving reliability in environments without a Ray cluster.

📈 vLLM Performance

• IOProcessor thread pool — Added thread pool support for vLLM IOProcessors, improving concurrency under load.
• OpenTelemetry traces — Full OpenTelemetry trace support for vLLM benchmark runs.
• Dependency installation in workers — Dependencies are now installed in benchmark execution workers at runtime.
• Fixed propagation of max_batch_tokens, gpu_memory_utilization, dtype, cpu_offload, and max_num_seq when creating a deployment.
• Fixed deletion of deployments that time out during creation.
• Fixed TypeError when parsing renderer_num_workers.

✂️ TRIM

No user-visible changes in this release cycle beyond build and versioning decoupling.

🧮 CPLEX-MIP

• New CPLEX-MIP operator — Added a custom experiment for solving Mixed-Integer Programming problems using CPLEX.
• Warm start & memory handling — Added warm start support and improved memory management for large MIP problems.

🚨 Detect Anomalous Series

No user-visible changes in this release cycle beyond build and versioning decoupling.

Full Changelog: 1.8.0...2.0.0

---

This discussion was created from the release Release 2.0.0.