Release History#

Unreleased#

Breaking Changes#

  • The Queue Server runner no longer listens to the Bluesky document stream. Right after submitting each plan, it calls the evaluation function in a background thread with the queue item UID returned by item_add. The evaluation function should wait for its data and raise if the acquisition failed. OptimizationResult.uids holds queue item UIDs (#76).

  • QueueserverAgent and QueueserverClient no longer accept document_dispatcher. acquisition_run_filter, ConsumerCallback, CORRELATION_UID_KEY, and QueueserverClient.start_listener/stop_listener are removed, and submitted plans no longer carry blop_correlation_uid.

  • QueueserverOptimizationRunner.stop() and QueueserverAgent.stop() stop further submissions without interrupting the in-flight evaluation. The future resolves with the partial result after that evaluation returns and is ingested, instead of immediately.

  • optimize, sample_suggestions, and acquire_baseline send the record of each evaluated point to the callbacks passed in the new callbacks argument instead of writing it through the RunEngine. Agent and Scipy pass the callbacks registered with subscribe. optimize and sample_suggestions no longer open an outer RunEngine run, so callbacks subscribed with RE.subscribe see only acquisition runs and RE(agent.optimize()) returns only acquisition run UIDs. To store optimization runs in Tiled, also subscribe the TiledWriter to the agent; find them by start["bluesky-adaptive"]["optimization_uid"] (#82).

  • Optimization-run events name their bookkeeping fields adaptive_suggestion_id (a string), adaptive_iteration, and adaptive_acquisition_uid instead of suggestion_ids, iteration, and acquisition_uid. Acquisition UIDs that are not strings, including tuples and arrays, are stored as their repr. The descriptor’s object_keys name the parameter ("parameters") and outcome ("outcomes") fields instead of DataKey source values. Parameter and outcome names must not start with adaptive_, and a parameter and an outcome must not share a name. The optimization stream of optimize_in_run is unchanged.

  • Removed OptimizationCallbackRouter, optimize_step (use optimize with iterations=1), the readable_cache argument of optimize and sample_suggestions, and the plans.OPTIMIZE_RUN_KEY and plans.SAMPLE_SUGGESTIONS_RUN_KEY constants.

Features#

  • OptimizationProblem and QueueserverOptimizationProblem gain optimization_uid, which the optimize, optimize_in_run, and sample_suggestions runs record as start["bluesky-adaptive"]["optimization_uid"]. Agent and Scipy generate one, expose it, and accept it to continue an optimization (#82).

  • Optimization run metadata is also recorded under "bluesky-adaptive" in start documents, duplicating evaluation_function, acquisition_plan, optimizer, sensors, and actuators. The top-level keys are kept for backwards compatibility.

  • default_acquire records the suggestion IDs in scan order as start["bluesky-adaptive"]["acquisition_order"]. "blop_acquisition_order" is kept for backwards compatibility.

  • New bluesky_adaptive.evaluation package. per_suggestion(evaluate, fetch) builds an evaluation function from a function that evaluates the data of one suggestion and sets the "_id" of each outcome. bluesky_adaptive.evaluation.documents.RunEventBuffer, subscribed with RE.subscribe(buffer), buffers the "primary" event data of default_acquire runs so that per_suggestion(evaluate, buffer.pop) can evaluate them (#73).

  • bluesky_adaptive.evaluation.tiled.TiledRunPoller(client) reads the "primary" data of default_acquire runs from Tiled so that per_suggestion(evaluate, TiledRunPoller(client)) can evaluate them. Each call polls Tiled until the run has a stop document, raising TimeoutError after timeout seconds. It needs the new tiled extra, which all includes (#73).

  • QueueserverOptimizationRunner runs its loop as an asyncio coroutine instead of in a dedicated thread guarded by a lock. run_async() and submit_suggestions_async() on the runner and on QueueserverAgent run it on the caller’s event loop. run() and submit_suggestions() still return a concurrent.futures.Future and run it on a shared background event loop; cancelling that future cancels the loop. Interpreter exit now waits for an in-flight evaluation to return (#98).

Bug Fixes#

  • optimize marks a step’s trials failed on a TrialFaultAware optimizer when its outcomes fail validation, as it already did for acquisition and evaluation errors.

  • sample_suggestions validates its suggestions and outcomes, and acquire_baseline its outcomes, like optimize instead of ingesting mismatched outcomes.

  • sample_suggestions marks its suggestions failed on a TrialFaultAware optimizer when acquisition, evaluation, or outcome validation fails, like optimize.

  • optimize records the step that meets the stopping criterion.

  • OptimizationLogger omits the Iterations line for runs whose start document has no iterations, such as sample_suggestions runs, instead of showing “Until stopping criterion”.

v1.0.1 (2026-11-07)#

Breaking Changes#

  • CanRegisterSuggestions.register_suggestions returns already-registered suggestions unchanged instead of assigning them a new "_id". AxOptimizer and XoptOptimizer raise ValueError if an "_id" is no longer awaiting outcomes or its parameters differ, and AxOptimizer also rejects an unknown "_id". Remove the "_id" to register the suggestion as new (#83).

  • QueueserverAgent and QueueserverClient no longer enable queue autostart, and QueueserverClient no longer accepts autostart. Start the queue or call re_manager_api.queue_autostart(True) yourself (#84).

Features#

  • QueueserverAgent and QueueserverOptimizationProblem accept acquisition_run_filter, a predicate on the start document that selects the run to evaluate when the acquisition plan opens more than one run with the submitted metadata (for example, a dark-frame run before the measurement). Previously the first such run completed the iteration and the next one stopped the campaign with current_uid did not match start document (#86).

Bug Fixes#

  • AxOptimizer and XoptOptimizer no longer register suggestions from suggest() a second time (#83).

  • The Queue Server runner validates suggestions and outcomes like optimize instead of ingesting mismatched outcomes (#85).

  • XoptOptimizer.from_checkpoint restores Bayesian generators with a fitted model. checkpoint() previously saved the model’s repr, which restore rejected (#78).

Full Changelog

v1.0.0 (2026-10-26)#

First release of Bluesky Adaptive built on the Blop v1.2.0 codebase (bluesky/blop). Entries below prefixed with “blop” are Blop releases.

Breaking Changes#

  • The package is renamed: replace pip install blop[...] with pip install bluesky-adaptive[...] and import blop with import bluesky_adaptive.

  • Extra names (ax, queueserver, scipy, xopt, all, cpu, dev) and the run metadata keys (blop_suggestions, blop_acquisition_order, blop_correlation_uid) are unchanged.

  • The bluesky-adaptive 0.x APIs (agents, adjudicators, server, and lock-step recommenders) are removed. Pin bluesky-adaptive<1 to keep them.

blop v1.2.0 (2026-10-23)#

This is the final feature release of Blop. Development continues in bluesky-adaptive 1.0.

Features#

  • Single-run optimization (experimental): blop.plans.optimize_in_run runs the optimization loop inside one Bluesky run, using blop.plan_stubs.list_scan_in_run for acquisition (#353).

  • optimize and Agent.optimize accept iterations=None to run until the optimizer’s stopping criterion is met (#364).

  • default_acquire records suggestion IDs in scan order as "blop_acquisition_order" start metadata. Use it to match acquired data to suggestions (#365).

  • optimize checkpoints after each iteration, so a deferred pause stops between iterations (#378).

Breaking Changes#

  • Optimization runs record one event per suggestion, with an iteration field, instead of one event per iteration with array values. read_step no longer takes n_points (#376).

Compatibility Notes#

  • EvaluationFunction and AcquisitionPlan are generic over the uid type (#370).

  • Suggestion "_id" values must be unique and hashable, and evaluation functions must return exactly one outcome per suggestion ID (#353).

Bug Fixes#

  • String fields such as suggestion IDs are no longer truncated to the length of their first value (#363).

  • default_acquire accepts md, which the Queue Server runner requires (#359).

Dependency Changes#

  • New scipy extra, included in all (#374).

Full Changelog

blop v1.1.0 (2026-08-26)#

Features#

  • SciPy backend: The new blop.scipy namespace provides Bluesky-compatible wrappers for SciPy’s local minimization, SHGO, and dual annealing routines (#339).

  • Global stopping criteria: The new SupportsStoppingCriteria protocol allows an optimizer to stop an optimization run early. Agent and AxOptimizer now accept an Ax global stopping strategy directly (#337, #338, #341).

  • Queueserver checkpointing: QueueserverAgent.run() and QueueserverOptimizationRunner.run() now accept checkpoint_interval. The checkpoint path can also be set on an existing Agent or AxOptimizer (#327, #328).

  • Flexible string actuators: Ax DOFs backed by a queueserver device string can now specify a separate name, supporting components whose queueserver path and device name differ (#346).

  • Generalized XRT simulations: blop-sim can now construct arbitrary XRT beamlines from XML descriptions and infer Bluesky-compatible motors and detectors (#317).

Compatibility Notes#

  • Optimization event identifier renamed: The bluesky_uid field in optimization event documents is now acquisition_uid. Acquisition plans may return any hashable identifier, rather than only a Bluesky run UID (#350).

  • Protocol signatures generalized: Core protocol methods now use Sequence and Mapping instead of concrete list and dict types, and evaluation functions receive a Hashable acquisition identifier. Runtime behavior remains compatible, but custom protocol implementations may need their annotations updated to satisfy static type checkers (#350).

Bug Fixes#

  • Evaluation failures are now registered with fault-aware optimizers before the original exception is re-raised, preventing failed suggestions from remaining pending (#355).

  • Suggestion routing now guarantees that every suggested point is visited exactly once, fixing edge cases that produced cycles and duplicate visits (#356).

  • Queueserver submission now uses the manager’s native autostart API and relies on its allowed-device and allowed-plan validation, allowing remote errors to propagate correctly (#326, #344).

Documentation and Development#

  • Added a how-to guide for configuring built-in and custom Ax stopping conditions (#342).

  • Added a contributor guide and documented the project’s review and merge policy (#331).

  • Added a dedicated integration test suite and CI workflow for complete optimization runs (#349).

Full Changelog

blop v1.0.0 (2026-07-16)#

Features#

  • Xopt backend: New optional blop.xopt.XoptOptimizer backend, installable via the blop[xopt] extra, implementing the Optimizer, Checkpointable, CanRegisterSuggestions, and TrialFaultAware protocols on top of Xopt (#318).

  • Scalarized objectives in `Agent`: Agent can now be configured with ScalarizedObjective for weighted multi-objective optimization (#284).

  • Best point(s) navigation: New blop.plan_stubs.navigate_to_best plan stub that moves actuators to the best point(s) found so far, backed by Optimizer.get_best_points() (#293).

  • Queueserver plan keyword arguments: QueueserverOptimizationProblem now accepts additional keyword arguments to pass through to the acquisition plan (#306).

  • More reliable queueserver runner: The queueserver runner now starts its listener automatically on initialization (#316), tracks its state with concurrent.futures.Future for more robust status reporting (#305), and QueueserverAgent now exposes additional runner properties (actuators, sensors, acquisition_plan, current_iteration) and a stop() method directly (#291).

Breaking Changes#

  • `QueueserverClient` now takes a `RemoteDispatcher`: Replaces the previous raw ZMQ address argument, giving users full control over how documents are dispatched from the queueserver (#308).

  • `QueueserverAgent` moved to its own module: QueueserverAgent now lives in blop.ax.queueserver_agent instead of blop.ax.agent. The queueserver optional extra was also renamed from qs to queueserver (#321).

Bug Fixes#

  • Made ChoiceDOF hashable, fixing a bug when reconfiguring or fixing ChoiceDOF parameters (#301).

  • Fixed stream filtering in the optimization logging callback (#302).

  • Fixed the queueserver runner calling stop_listener from the wrong thread, and used the correct run_start key on the stop document (#291).

  • Queueserver runner now prevents concurrent/parallel runs (#296).

  • Fixed the import path for ContourPlot (#281).

Dependency Changes#

  • Reworked and clarified the optional-dependency structure, adding an all extra and renaming the qs extra to queueserver (#321).

Documentation#

  • Queueserver integration marked experimental: Clearly documented across module docstrings, the tutorial, and the API reference that the queueserver integration is experimental and subject to change (#309).

  • New tutorial for running Blop optimization against a remote Bluesky Queueserver (#292).

  • Improvements to the XRT KB mirror tutorial (#311) and the simple experiment tutorial (#303), plus a general documentation maintenance pass (#294).

  • New explanation content describing which components of an optimization workflow (evaluation functions, acquisition plans) users are responsible for implementing (#323).

  • Website redesign, including a version-switcher dropdown in the docs header (#269, #277, #279, #287, #288).

  • Various consistency improvements to the documentation.

Full Changelog

blop v1.0.0b1 (2026-04-17)#

Features#

  • Manual suggestions: A new method for the Agent that allows manual optimization control (#235).

  • Reconfigurable search spaces: Search spaces can now be modified between optimization steps, allowing dynamic adjustment of DOF bounds and constraints (#268).

  • Model checkpoints: Save and restore optimizer state across sessions (#233).

  • Fixed parameters: Hold specific parameters constant during optimization via Agent.fixed_dofs (#252).

  • Multi-point routing: When suggesting multiple points, suggestions are now routed to minimize actuator travel (#217).

  • Failed and abandoned suggestions: Properly handle and track failed optimization trials (#272).

  • Optimization logging callback: New OptimizationCallbackRouter and logging callback for monitoring optimization progress (#270).

  • Optimization step tracking via event-model: Optimization metadata is now emitted as Bluesky event-model documents (#236).

  • Queueserver support: New QueueserverAgent for running optimization through the Bluesky queueserver (#212, #264, #266).

  • Actuator and Sensor types: Expanded protocol types to support flyable and collectable devices (#211).

  • Python 3.13 support (#206).

Breaking Changes#

  • Plans moved to top-level package: Plans are now imported from blop.plans instead of blop.ax.plans (#259).

  • Simulation code separated: The simulation module has been extracted into a separate blop_sim package (#248, #256).

  • Deprecated code removed: Legacy APIs deprecated in earlier releases have been removed (#218).

  • Deprecated bayesian.acquisition removed: Use BoTorch’s built-in constrained acquisition functions instead (#207).

Dependency Changes#

  • Requires Ax Platform >= 1.2.3 (#241).

  • Requires BoTorch >= 0.16.0 (#221).

Documentation#

  • New how-to guide for using Tiled as a databroker (#215).

  • New how-to guide for using ophyd and ophyd-async devices (#210).

  • New explanation document for the Ax integration (#227).

  • Updated tutorials for the simple experiment and XRT KB mirrors (#222, #223, #224).

Full Changelog

blop v0.9.0 (2025-12-08)#

What’s Changed#

Full Changelog

blop v0.8.1 (2025-11-06)#

What’s Changed#

Full Changelog

blop v0.8.0 (2025-10-13)#

What’s Changed#

New Contributors#

Full Changelog

blop v0.7.5 (2025-06-18)#

What’s Changed#

New Contributors#

Full Changelog

blop v0.7.4 (2025-03-04)#

  • Add missing files for documentation

  • Fix trigger condition for releases on PyPI and documentation

blop v0.7.3 (2025-03-04)#

What’s Changed#

New Contributors#

Full Changelog

blop v0.7.2 (2025-01-31)#

  • Renamed package in PyPI to blop.

  • bloptools is still avaliable on PyPI.

blop v0.7.1 (2024-09-26)#

  • Add simulated hardware.

  • Added a method to prune bad data.

blop v0.7.0 (2024-05-13)#

  • Added functionality for Pareto optimization.

  • Support for discrete degrees of freedom.

blop v0.6.0 (2024-02-01)#

  • More sophisticated targeting capabilities for different objectives.

  • More user-friendly agent controls.

blop v0.5.0 (2023-11-09)#

  • Added hypervolume acquisition and constraints.

  • Better specification of latent dimensions.

  • Implemented Monte Carlo acquisition functions.

  • Added classes for DOFs and objectives.

blop v0.4.0 (2023-08-11)#

  • Easier-to-use syntax when building the agent.

  • Modular and stateful agent design for better usability.

  • Added the ability to save/load both data and hyperparameters.

  • Added passive degrees of freedom.

  • Added a number of test functions / artificial landscapes for optimization.

  • Updated the Sphinx documentation theme to furo.

blop v0.3.0 (2023-06-17)#

  • Implemented multi-task optimization.

  • Simplified the syntax on initializing the agent.

  • Resolved issues discovered at NSLS-II ISS.

blop v0.2.0 (2023-04-25)#

  • Rebased the Bayesian optimization models to be compatible with botorch code.

  • Optimization objectives can be customized with experiment modules.

  • Added optimization test functions for quicker testing and development.

blop v0.1.0 (2023-03-10)#

  • Changed from using SafeConfigParser to ConfigParser.

  • Implemented the initial version of the GP optimizer.

  • Updated the repo structure based on the new cookiecutter.

  • Added tests to the CI.

blop v0.0.2 (2021-05-14)#

Fixed _run_flyers() for sirepo optimization.

blop v0.0.1 - Initial Release (2020-09-01)#

Initial release of the Beamline Optimization library.

Used in:

Planned: