0.3.0

Release Notes — v0.3.0

Breaking Changes

job_name_prefix is now truly optional. Previously, omitting the prefix caused a random 5-character prefix to be auto-generated. Now, when no prefix is configured, job names are submitted without a prefix and bjobs queries are not filtered by name. Clients that relied on the implicit random prefix for isolation between concurrent sessions should now set an explicit job_name_prefix in their config. reconnect() still requires a prefix to be set.

Unknown config profiles now raise ValueError. Requesting a profile name that doesn't exist in the YAML file previously failed silently (the profile was just ignored). It now raises with a message listing available profiles. Clients passing speculative profile names will need to handle this or ensure the profile exists.

Array job aggregate status: FAILED now takes precedence over KILLED. When an array job has a mix of failed and killed elements, the overall status is now FAILED rather than KILLED. This better reflects that something went wrong rather than just being cancelled.

New Features

done flag on cancel() and cancel_all(). Pass done=True to mark cancelled jobs as DONE instead of KILLED. On LSF this translates to bkill -d. Useful for gracefully retiring jobs you don't want flagged as killed.

Bug Fixes

• Paths with spaces no longer break LSF submission. Stdout, stderr, and cwd paths in #BSUB directives are now double-quoted. The bsub invocation itself switched from shell execution (create_subprocess_shell) to create_subprocess_exec with stdin file redirection, eliminating shell-escaping issues.
• cancel_by_name() no longer raises when no jobs match. LSF's "No matching job" / "No unfinished job" errors are now caught and logged at debug level instead of propagating.
• reconnect() skips completed jobs. Terminal (DONE/EXIT) jobs returned by bjobs -a are now filtered out so reconnect only picks up jobs that are still running or pending.
• Zombie processes reaped after local cancel. The local executor now properly awaits killed child processes to avoid zombies.

Improvements

• Faster local array job cancellation. SIGTERM is now sent to all element processes concurrently, with a single collective wait, instead of terminating and waiting on each one sequentially.
• Warning when using max_concurrent with LocalExecutor. Since the local executor doesn't support concurrency throttling, it now logs a warning instead of silently ignoring the parameter.
• Cancel logic refactored into _cancel_job(). Cancellation is now a proper overridable method on the base Executor, making it easier to customize in subclasses.
• Comprehensive docstrings added to Executor, ResourceSpec, and internal methods. New docs/API.md reference document.

0.2.4

Fix to convert timestamps to UTC.

Full Changelog: 0.2.3...0.2.4

0.2.3

• Sanitize job names

Full Changelog: 0.2.2...0.2.3

0.2.2

New Features

• Job reconnection — LSFExecutor.reconnect() queries bjobs by
  name prefix and reconstructs JobRecord instances so
  monitoring can resume after a process restart. Supports both
  single jobs and array jobs. Requires job_name_prefix to be
  set in config.
• Debug logging — Added logger.debug() calls for all
  scheduler CLI invocations (bsub, bkill, bjobs) so command
  lines are visible at DEBUG level. Info-level logging for
  submit, cancel, and reconnect outcomes.

Improvements

• Simplified bsub submission — Removed the use_stdin code
  path; script submission now always uses shell redirection
  (bsub < script.sh), which correctly enables #BSUB directive
  parsing.

Cleanup

• Removed unused lsf_detect_units() function (dead code,
  never called).
• Removed unused register_executor() function from the
  executor registry.
• Deduplicated _ARRAY_ELEMENT_RE regex — LocalExecutor now
  imports the pattern from core.py instead of defining its own
  copy.

Full Changelog: 0.2.1...0.2.2

0.2.1

New Features

• extra_args on config and ResourceSpec — extra CLI arguments
  appended to the submit command (e.g. bsub -P myproject
  script.sh). Useful for flags that must appear on the command
  line rather than as script directives. Supports both
  config-level (all jobs) and per-job via ResourceSpec.
• Real-time log output — LocalExecutor now streams
  stdout/stderr directly to log files during execution instead
  of buffering until completion.

Breaking Changes

• ResourceSpec.account removed — use extra_args=["-P",
  "myproject"] or extra_directives=["-P myproject"] instead.
• ResourceSpec.cluster_options renamed to extra_directives —
  aligns with the same field on ClusterConfig. Both now
  auto-prepend the directive prefix (#BSUB), so values should
  be raw flags (e.g. "-G mygroup") not full directives.
• ClusterConfig.account removed — same as above.
• ClusterConfig.extra_directives no longer requires the #BSUB
  prefix — the executor prepends it automatically. Change
  "#BSUB -env 'all'" to "-env 'all'".

0.2.0

New Features

• LocalExecutor array job support — submit_array() now spawns
  one subprocess per array element, each with an ARRAY_INDEX
  environment variable. Poll, cancel, and output file writing
  all handle array elements correctly. Per-element log files
  use the pattern stdout.{job_id}.{index}.log /
  stderr.{job_id}.{index}.log.

Breaking Changes

• ResourceSpec.work_dir is now required (non-nullable) —
  changed from str | None (default None) to str (default
  os.getcwd()). Code passing work_dir=None explicitly will
  break; code omitting it will get cwd automatically.
• ResourceSpec gained stdout_path / stderr_path fields — new
  optional fields for overriding default log file paths. Not
  breaking per se, but changes the dataclass shape.
• ClusterConfig.log_directory removed — log files are now
  written to ResourceSpec.work_dir instead of a separate global
  log directory. Config files with log_directory will trigger
  an "unknown config keys" warning.
• Executor._submit_job() signature changed — now takes
  (command, name, resources, prologue, epilogue, env, *, cwd)
  and returns tuple[str, str | None] instead of taking
  (script_path, name, env) and returning str. Custom executor
  subclasses must be updated.
• Executor._submit_array_job() signature changed — same
  pattern as _submit_job(), now receives full submission
  parameters and returns tuple[str, str | None].
• Executor.render_script() and Executor.build_header()
  removed from base class — script rendering moved to
  standalone render_script() / write_script() functions in the
  new cluster_api.script module. build_header() is now only on
  executor subclasses, not abstract on the base.
• Executor._write_script() removed — replaced by
  cluster_api.script.write_script(). Scripts are now written to
  work_dir instead of log_directory.
• Executor.directive_prefix removed from base class — only
  defined on subclasses that need it.
• Log file naming changed — LSF logs changed from
  {name}.out/{name}.err to stdout.%J.log/stderr.%J.log; local
  logs changed from {name}.out/{name}.err to
  stdout.{job_id}.log/stderr.{job_id}.log.

Bug Fixes

• Race condition in JobMonitor.wait_for() — completion events
  are now registered before checking is_terminal, preventing a
  hang when a job finishes between the check and registration.
• Non-atomic script counter — LSFExecutor and LocalExecutor
  now use itertools.count() instead of manual += 1 increment,
  avoiding counter collisions on concurrent submits.

Improvements

• Unknown config keys now logged — load_config() emits a
  warning for unrecognized keys instead of silently ignoring
  them.
• Unmapped LSF statuses now logged — unknown bjobs status
  strings trigger a warning instead of silently mapping to
  UNKNOWN. Added mappings for UNKWN, WAIT, and PROV.
• Better poll error logging — failed status queries now
  include the exception message and traceback in the warning.
• Script rendering extracted to cluster_api.script module —
  render_script() and write_script() are now standalone
  functions, decoupled from the executor class hierarchy.

0.1.1

Full Changelog: 0.1.0...0.1.1

0.1.0

Initial release