Changelog

2.1.1 — 2026-06-09

Load from DB (new widget)

  • Lists every dataset registered in the datasets metadata table via SQLAlchemy and pulls the selected one back into Orange as an Orange.data.Table using pandas.read_sql + Orange.data.pandas_compat.table_from_frame.

  • Offers a Class column combo populated from the dataset’s columns, pre-selected with the user’s persisted choice or the class_name recorded by Save to DB. The output Table already exposes the chosen column as domain.class_var, so no Select Columns widget is needed downstream.

  • Same dialect selector (PostgreSQL / MySQL), connection-status label and QThread worker pattern as Save to DB. Both the metadata listing and the actual table read happen off the GUI thread.

  • Workflow-persisted settings: selected_dataset and selected_class are Setting(..., schema_only=True), restored as soon as the connection comes back up on reload.

  • The Dataset combo is a ComboBoxSearch — type to filter when the registry gets large.

  • ↻ button next to the Dataset combo re-runs the listing without dropping the connection (handy when something else just published a new dataset). Auto-load is suppressed on Refresh so the user never gets a Load they didn’t ask for.

  • Delete button drops the selected dataset’s table and removes its row in datasets. Runs through a _DeleteDatasetWorker on a background thread, guarded by a confirmation dialog. Idempotent: re-deleting a missing table is a no-op thanks to DROP TABLE IF EXISTS.

  • Auto-load on workflow reopen: if the persisted selected_dataset is still on the server when the connection succeeds and the listing returns, Load fires automatically — the data flows out without a single click. The flag is cleared on Refresh and on Delete to avoid surprises.

Variable Dependency Graph

  • New: edges now carry numeric weights equal to the largest |argument| among the temporal calls (shift, sum, mean, count, min, max, sd) in the source expression that reference the dependency. Plain (non-temporal) references default to weight 1. The weights live in network.edges[0].edges.data and are consumable by any downstream Network widget.

  • New: the graph is now directed (DirectedEdges). Previously the sparse matrix was passed straight to Network which auto-wrapped it as UndirectedEdges — a latent bug since A→B is not the same as B→A in a dependency graph.

  • New: nodes carry an extra expression meta (literal expression text, empty for original variables) so the Network Explorer can show the formula as the node label.

  • New: Warning.no_derived fires when every input row is an original variable — usually a sign the user wired the data output of Time Features Constructor instead of the variable-definitions one.

  • Refactor: flattened the from_row_col / grafo decorator pattern into a single, documented build_dependency_network function.

  • Performance: O(n³) → O(n²) by precomputing a name → index map and using a set to dedupe dependencies.

  • Bugfix: an IndexError was raised when the input domain had a single column. The widget now checks len(domain) >= 2 before reading the second column.

  • Removed 65 lines of dead commented-out code.

Time Features Constructor

  • New: chained descriptors. A descriptor can now reference another derived descriptor in its expression (e.g. X2 := shift(X1, -1) with X1 itself defined a few rows above). The widget topologically sorts the descriptors and cascades the transforms — each step runs against the table state produced by the previous one, so X2 sees X1 as a regular column. Cycles (e.g. X1 := X2 + 1 together with X2 := X1 + 1) raise a Circular dependency between descriptors: X1, X2 error. Errors during evaluation are reported per-descriptor so the failing row is obvious.

  • Fixed (critical): time-window functions used to lose context every 5 000 rows because Orange chunks tables during transform. The widget now caches the full source per FeatureFunc and returns the right slice for each chunk, so shift(x, -20) is correct across chunk boundaries on datasets of any size.

  • Fixed: “Variables to generate” was being cleared after every Send and on every input change, so workflow save / reload lost the definitions. Storage moved from ContextSetting to Setting(..., schema_only=True) (matching upstream Orange v4) and the editor is now restored from the persisted descriptors on every input. The legacy context handler stays around to migrate old workflows.

  • Each Send re-transforms the original input table instead of the cumulative output, removing the implicit “consume on Send” semantics and the duplicate-name errors that came with it.

  • eval hardened: __builtins__ replaced with an empty dict in the expression evaluator. Only the curated whitelist (safe builtins + math + selected random / numpy helpers) is exposed.

  • Refactor: modificar_expression collapsed from 7 near-identical loops to a single regex-driven pass.

  • Bugfix: FeatureEditor.editorData returned the variable name as the expression.

Save to DB

  • New: write mode selector with three options — Create new (default, fail if the target exists), Overwrite (drop and recreate the table and its datasets row), Append (keep existing rows and add the new ones). Re-running a workflow no longer breaks. The persisted write_mode Setting defaults to "create" so old workflows keep their previous behaviour. After the upload, the widget runs SELECT COUNT(*) and rewrites the datasets row with the actual total, so the registry stays accurate across appends.

  • New: MySQL support. The connection panel now exposes a database-type selector (PostgreSQL / MySQL). Per-dialect column types and identifier quoting ("name" vs \`name\`) live in a _Dialect abstraction.

  • Performance: uploads now go through Orange.data.pandas_compat.table_to_frame + DataFrame.to_sql(method='multi', chunksize=1000) over a SQLAlchemy engine. The old row-per-INSERT loop is gone — typical speedups are 50-100×, especially over the network. Identifier quoting and per-column DDL types are emitted by the SQLAlchemy dialect (DOUBLE_PRECISION on PostgreSQL, DOUBLE on MySQL, DATETIME instead of TIMESTAMP on MySQL to dodge the 2038 cap, VARCHAR(255) / TEXT everywhere else).

  • UX: the upload runs in a background QThread (_UploadWorker), so the canvas stays interactive on long writes. A status label under the connection box reports Not connected / Connected to … : … / Uploading rows N/M… / Upload completed in Xs / Upload failed: … with colour cues. Save, Connect and the form fields are disabled while the worker runs and re-enabled on success or failure; the widget aborts the thread cleanly on close.

  • Fixed (critical): SQL injection in the metadata INSERT. The query is now fully parametrised.

  • Added an identifier whitelist (^[A-Za-z_][A-Za-z0-9_]{0,62}$) and a quote_ident helper that wraps identifiers with PostgreSQL-standard double quotes (with any internal " doubled). Used in every CREATE TABLE / INSERT INTO that touches user-supplied names.

  • Replaced psycopg2 with psycopg2-binary in the install requirements so installation works on macOS, Linux and Windows without a C compiler.

  • setup.py now declares install_requires; pip install was previously not pulling in any runtime dependency.

Testing

The widget suite grew to 92 tests, covering:

  • Unit tests for modificar_expression, the time-window helpers, _sanitize_name / _expression_or_none and _temporal_weights.

  • A regression test for the 5 000-row chunk bug (shift(x, -20) over a 12 000-row table).

  • eval-safety tests asserting that __import__ / open raise NameError while sqrt / abs still resolve.

  • Widget-level tests (via Orange.widgets.tests.base.WidgetTest) for descriptor persistence, including an end-to-end settingsHandler.pack_data / stored_settings= round-trip that mirrors what Orange does when saving and reopening a .ows file.

  • Edge-weight tests covering single calls, three-argument families, mixed temporal / non-temporal references, max across multiple calls, and per-dependency independence.