harens/AnomaLog: v0.4.0

0.4.0 (2026-06-03)

Features

• add next-event prediction diagnostics to DeepCASE and DeepLog models (1b48275)
• Add OpenStackDeepLogParser and SpellTemplateParser for OpenStack log processing (2892c72)
• add optional per-event anomaly labels to TemplateSequence and validate alignment (3820888)
• add Slurm job wrappers and configuration management (5ca223b)
• ait-ads: suppress sequence-level metrics in reports and add related tests (a0f6808)
• anomalog: add preprocessed DeepLog session source (a2aeffc)
• baselines: run with all ML models (8638cf2)
• bgl: split CCS 2017 and How Far Are We protocols (51a2e4f)
• build_templated_dataset: coarse lock over dataset name + cache path (5ab01b5)
• datasets: add AIT-ADS scenario support (1a7ec8c)
• deepcase: add cluster-labelling ablations to registry (d451235)
• deepcase: add event-level prediction metrics, not just sequence (436e154)
• deepcase: enhance training feedback by reporting progress per epoch (7693743)
• deepcase: update documentation and tests for zero-query iterations in scoring (c2d0629)
• deeplog: add HDFS short-session padding variant (e63086e)
• deeplog: add OpenStack parameter CI approximation (ce38437)
• deeplog: add OpenStack regression coverage (b2d8470)
• deeplog: add progress reporting for key model and parameter schema preparation (a631783)
• deeplog: add short-session padding fidelity mode (e5ada94)
• deeplog: align BGL 2022 path with Drain3 and CI highlights (834e308)
• deeplog: carry parameter history across openstack entity windows (d7d4a13)
• deeplog: disable parameter model by default, correct g default value (fe77e1c)
• deeplog: emit parameter ci report artifacts (c243b1f)
• deeplog: left pad by default (c41195d)
• deeplog: make top-g replay configurable (2b55195)
• deeplog: merge continuous context series in parameter dataset construction (5967617)
• deeplog: separate parameter summary from debug trace (3556dca)
• deeplog: support parameter-only scoring and registry updates (ea44148)
• deeplog: surface OpenStack Figure 9 metadata in parameter reports (945ae45)
• dependencies: add spellpy as a git dependency and update related markers (b0d079c)
• detectors: implement BatchExperimentDetector for bulk scoring and enhance evaluation logic (b1f0ff8)
• docs: add reference documentation for experiments package (80e20ae)
• Enhance continuous context handling in DeepLog (e464816)
• enhance materialization to handle stale Prefect cache paths and update related tests (9feaa97)
• entity_chronological: add AIT-ADS entity-chronological dataset and update related configurations (dcd771a)
• experiment_logger: enhance logging with concrete run names (413f52a)
• experiment_runner: add --write-predictions option to persist predictions.jsonl (ef06f46)
• experiment_runner: enhance failure handling to log bundle failures without stopping execution (a9a97e4)
• experiments: add DeepLog and DeepCASE detectors (7b3745e)
• experiments: add event-level baseline reporting (8cf2c13)
• experiments: add HDFS Table-IV compatibility suite (488bf5e)
• experiments: add Prefect telemetry configuration to Slurm wrapper (cd70adb)
• experiments: allow for multi-process model sweeps (46e0a05)
• experiments: commit registry-backed experiment overhaul (af97052)
• experiments: enhance error handling in submit_experiments for sbatch failures (c0d6e9b)
• experiments: lazy load models for optional extra dependencies (a58e50c)
• experiments: make data/cache root configurable for slurm jobs (caf620e)
• experiments: publish scoped metric blocks in run outputs (1fcae3d)
• experiments: separate run_groups to execute (5759e81)
• experiments: share progress totals and logging (f0019a6)
• experiments: streamline run metric metadata (c8ce126)
• experiments: submit Slurm arrays in one job (18bd438)
• group missing suite reruns by experiment (0133d2e)
• markov: introduce markov baseline (d53a315)
• metadata: record DeepCASE version in environment metadata (0250d97)
• models: add SingleFitMixin for single fit state management in detectors (9f4247b)
• models: support bounded train progress hints (2b24a2a)
• modify train/test fractions in tandem (28a21f0)
• naive_bayes: add type annotations and improve docstrings (b611c63)
• openstack: make the Figure 9 anomaly slice explicit (2ca83e0)
• parquet: implement chronological entity grouping and persist entity chronology index (28d7f48)
• parsers: add context manager for spellpy logging and update tests (16a01f5)
• parsers: implement lightweight mode for SpellTemplateParser and add corresponding test (47869ca)
• parsing: add thunderbird parser and timing logs (f750391)
• registry: add HDFS DeepLog Drain3 ablations (0be1600)
• run_suite: add check for missing concrete runs and update tests (6b5e28e)
• run_suite: enhance output handling for missing runs and update rerun command format (d824176)
• sequences: infer split labels from preprocessed entity prefixes (c276b7c)
• set paper-faithful deepcase iterations, expose next event prediction metrics (0131ccf)
• slurm: add initial sbatch experiment scripts (a2aa3ba)
• slurm: enhance job scripts to dynamically set repository root and cache directory (674f476)
• slurm: simplify REPO_ROOT assignment in Slurm scripts and tests (26b0f0d)
• slurm: update wrap script to export EXPERIMENT_NAME for nested commands (d82fa79)
• slurm: update wrap script to use loop for experiment indexing (f038698)
• slurm: update wrap script to use set -- for experiment indexing (9b84a8d)
• sources: add preprocessed and thunderbird tests (c80f5be)
• sources: add tar.gz support (4964c07)
• template_frequency: clarify documentation on threshold calibration and update label checks (ff609ea)
• tests: add fixtures to mirror Prefect API URL in subprocess environments (71c1c23)
• tests: add test for entity counting using grouped rows in ParquetStructuredSink (236ea9e)
• tests: add test for parquet dataset schema exposing timestamp and partition fields (bbbc128)
• tests: add unit tests for DeepLog preprocessed dataset source helpers (60f9be6)
• tests: update wrap script assertions to reflect changes in error handling (1f55820)
• thunderbird: add entity-grouped DeepCASE extension and update registry (d3c2561)
• thunderbird: consolidate registry entries (3336713)
• thunderbird: enhance parser to handle message tails and update entity ID extraction (a684bdf)
• thunderbird: normalise benchmark slice input (0d96003)
• torch_runtime: add shared helpers for managing torch device and seed (b7c0485)
• update experiment configurations to use chronological datasets and remove obsolete files (0c34300)
• update metrics structure to use canonical metric blocks and remove legacy fields (7256c05)

Bug Fixes

• accept all-normal inline-labelled structured datasets (dde55eb)
• ait-ads: split by chronology before entity grouping (7268f8a)
• align BGL 10% DeepLog with normal-entry split (bc349b3)
• align Thunderbird fixed-window contract (75bffbe)
• bump deepcase commit version for batching fix (1043917)
• cache: enhance result storage handling and cache policy configuration (5025ef5)
• cache: stop cached materialisation results pointing to stale per-run paths (fbac953)
• deepcase: align HDFS compat split with entity grouping (9030c01)
• deepcase: align query defaults with paper (3f5a000)
• deepcase: align workload reduction with paper semantics (9b6b520)
• deepcase: cache training chunks and separate query budget (913bd90)
• deepcase: chunk adapter batches to bound memory (1a07946)
• deepcase: don't treat abstained scores as anomalous (0b99151)
• deepcase: exclude evaluation_event_mask events from score (c736cd8)
• deepcase: optimise template access and label resolution in training batch (9155033)
• deepcase: preserve optimiser state in chunked training (0e27c5f)
• deepcase: reduce fit-time memory pressure (08352da)
• deepcase: use finer grained anomaly labels where available (03741f0)
• deeplog: evaluator no longer repeatedly charging warm-up penalty (8d9b32f)
• deeplog: harden OpenStack Figure 9 parameter CI (7aa7bae)
• deeplog: keep compatibility variant on deeplog_default (95df577)
• deeplog: microbatch key-model training (6a391c2)
• deeplog: next-event predictions over all logs, not just latest one (c20858e)
• deeplog: remove g=11 from default replay cutoffs (742bb69)
• deeplog: remove short-session padding fallback (bf94416)
• deeplog: stream Spell input during training (7f740a0)
• deeplog: use increased epochs/batch size for hdfs pre-processed (d137477)
• dependencies: add diff-cover to dev dependency group (66f41b4)
• don't consume log sequences between different models (7a53d32)
• drop incompatible baselines from paper configs (f020967)
• drop obsolete spellpy persist_state argument (e536a43)
• evaluator no longer treats every DeepCASE outcome as abstained (9e8a1ef)
• experiments: apply deepcase model-set overrides (64f0f12)
• experiments: correct supervised entity split fractions (0e27773)
• experiments: invalidate dataset cache on force reruns (826abfb)
• experiments: raise Prefect startup timeout for dataset builds (e38d2e5)
• hdfs_v1: add support for non-integer csv anomaly labels (7a2f4df)
• keep test set the same across different train splits (d27fc42)
• make OpenStack DeepLog preprocessing paper-faithful (6c37d8d)
• markov: accept mixed BGL training chunks (823c94f)
• markov: remove quadratic calibration scan (7638d2b)
• openstack-deeplog: normalise volatile session tokens (958e522)
• openstack: group by instance id (c9b75c7)
• paper-faithful split contract changes (da2d6c5)
• parquet-sink: rebuild missing structured cache directories (19557ca)
• parquet: sort rows within each entity by line order for accurate grouping (24e4253)
• parquet: stale parquet cache making entity scan look empty (4c85634)
• parquet: validate cached parquet fragments before reuse (409d915)
• parsers: fix openstack dropping traceback/error continuation rows (6407caf)
• parsers: simplify SpellTemplateParser after spellpy fix (f383178)
• record file-boundary split provenance in manifests (56a2c47)
• recover stale Prefect cache paths (b01eae2)
• repair chronological stream and audit typing (4d7f94b)
• resolve paths in ExperimentBundle for consistent path handling (029fe01)
• ruff: update external linting rules to include only "DOC" (f68660f)
• run_bundle: improve error handling for existing result paths (0a283de)
• spell: collapse back to direct spellpy parsing (3b55bbf)
• template: clear stale spell cache before retraining (f142c57)
• tests: keep deeplog/case runners independent of dataset.toml (f003c94)
• thunderbird: invalidate template cache by raw slice (33afd82)
• thunderbird: skip empty Thunderbird message rows (a225a79)

Performance Improvements

• cache registry bundle decoding (fce28e2)
• deeplog: speed up Thunderbird key training (3272761)

Documentation

• anomalog: apply strict pydoclint across modules (8429e76)
• deeplog: audit templates (82eede4)
• deeplog: document duplicate-session findings in HDFS preprocessed test files (74b389a)
• experiments: document caching and reruns (5240b0d)
• experiments: document DeepLog and DeepCASE support (8b5bda5)
• experiments: pydoclint docstrings (8c082df)
• experiments: refresh Slurm registry examples (3b73ca0)
• experiments: satisfy pydoclint for Slurm backend (d594306)
• tests: apply strict pydoclint (0978c04)
• thunderbird: note cache-key fix and template contract (541ecc7)
• thunderbird: record parser skips and DeepLog comparison (c75c635)
• thunderbird: refresh reproduction notes and config (a3f3172)
• tighten docstrings for pydoclint (85db1be)