TimeFeatures 2.0.0__tar.gz → 2.1.0__tar.gz

{timefeatures-2.0.0/TimeFeatures.egg-info → timefeatures-2.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: TimeFeatures
-Version: 2.0.0
+Version: 2.1.0
 Summary: Timefeatures add-on for Orange 3 data mining software.
 Home-page: https://github.com/alervgr/Orange-TimeFeatures
 Author: Alejandro Rivas García
@@ -86,12 +86,15 @@ If using Anaconda Python distribution, simply run
     pip install TimeFeatures
 
 **Required Dependencies**:
+
 * numpy>=1.22.4
 * AnyQt>=0.2.0
-* Orange3>=3.34.0
 * PyQt5>=5.15.6
 * PyQtWebEngine>=5.15.6
 * scipy>=1.7.3
+* SQLAlchemy>=1.4.0
+* psycopg2-binary>=2.9.9
+* PyMySQL>=1.0.0
 * Orange3-Network>=1.8.0
 
 Usage

{timefeatures-2.0.0 → timefeatures-2.1.0}/README.md

@@ -42,12 +42,15 @@ If using Anaconda Python distribution, simply run
     pip install TimeFeatures
 
 **Required Dependencies**:
+
 * numpy>=1.22.4
 * AnyQt>=0.2.0
-* Orange3>=3.34.0
 * PyQt5>=5.15.6
 * PyQtWebEngine>=5.15.6
 * scipy>=1.7.3
+* SQLAlchemy>=1.4.0
+* psycopg2-binary>=2.9.9
+* PyMySQL>=1.0.0
 * Orange3-Network>=1.8.0
 
 Usage

{timefeatures-2.0.0 → timefeatures-2.1.0/TimeFeatures.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: TimeFeatures
-Version: 2.0.0
+Version: 2.1.0
 Summary: Timefeatures add-on for Orange 3 data mining software.
 Home-page: https://github.com/alervgr/Orange-TimeFeatures
 Author: Alejandro Rivas García
@@ -86,12 +86,15 @@ If using Anaconda Python distribution, simply run
     pip install TimeFeatures
 
 **Required Dependencies**:
+
 * numpy>=1.22.4
 * AnyQt>=0.2.0
-* Orange3>=3.34.0
 * PyQt5>=5.15.6
 * PyQtWebEngine>=5.15.6
 * scipy>=1.7.3
+* SQLAlchemy>=1.4.0
+* psycopg2-binary>=2.9.9
+* PyMySQL>=1.0.0
 * Orange3-Network>=1.8.0
 
 Usage

{timefeatures-2.0.0 → timefeatures-2.1.0}/docs/changes.rst

@@ -22,6 +22,22 @@ Unreleased
 - 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**
 
@@ -31,6 +47,17 @@ Unreleased
   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
@@ -42,6 +69,16 @@ Unreleased
 
 **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
@@ -68,6 +105,15 @@ Unreleased
 
 **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

{timefeatures-2.0.0 → timefeatures-2.1.0}/docs/widgets/load-from-db.rst

@@ -48,9 +48,21 @@ Controls
        host/db (N datasets)", "Loaded <name> (N rows)"), error
        ("Connection failed: …", "Load failed: …").
    * - Dataset
-     - Combo populated from ``SELECT * FROM datasets ORDER BY
-       datetime DESC``. The most recent upload comes first; the last
-       choice is restored when reopening a workflow.
+     - Searchable combo populated from ``SELECT * FROM datasets ORDER
+       BY datetime DESC``. The most recent upload comes first; the
+       last choice is restored when reopening a workflow. Type to
+       filter the visible items — handy when the registry grows large.
+   * - ↻ (Refresh)
+     - Small button to the right of the Dataset combo. Re-runs the
+       list query without dropping the connection. Useful if another
+       Orange canvas (or a parallel tool) just published a new
+       dataset while this widget was open.
+   * - Delete
+     - Drops the currently selected dataset's table from the database
+       and removes its row in ``datasets``. Gated by a confirmation
+       dialog — there is no Orange-side undo. The operation runs on a
+       background thread, so the canvas stays responsive even if the
+       table is large.
    * - Dataset info
      - Read-only block under the combo: save timestamp, row/column
        counts and the original class column recorded by **Save to DB**.
@@ -60,7 +72,9 @@ Controls
        choice, (2) the ``class_name`` stored in the ``datasets``
        metadata, (3) ``(no class)`` if none of the above apply.
    * - Load
-     - Triggers the actual download.
+     - Triggers the actual download. Skipped automatically on the
+       first connection after a workflow reopen if the persisted
+       dataset name is still available (see *Auto-load* below).
 
 How it Works
 ------------
@@ -88,6 +102,28 @@ While any worker runs, the form controls (database type, connection
 fields, **Connect**, **Load**, dataset and class combos) are
 temporarily disabled and the status label keeps the user informed.
 
+Auto-load
+---------
+
+When you reopen a workflow that already had a Load from DB widget with
+a saved ``selected_dataset``, the widget fires Load automatically the
+first time the dataset listing comes back successfully. The data flows
+out of the **Data** output without a single click, mirroring how Orange
+treats sources like **File** and **Datasets**.
+
+Subtleties:
+
+- Auto-load is a *one-shot* per widget lifetime. Manually clicking
+  **Refresh** clears the pending flag, so a Refresh never surprises
+  the user by loading something behind their back.
+- If the persisted dataset no longer exists on the server (deleted
+  from outside, or the registry was wiped), the flag is cleared and
+  the widget just shows the available list — no error.
+- If a different backend or set of credentials is restored, the user
+  still has to click **Connect** explicitly, just like in earlier
+  versions; auto-load happens *after* the first successful connection
+  + listing.
+
 Workflow Persistence
 --------------------
 

{timefeatures-2.0.0 → timefeatures-2.1.0}/docs/widgets/save-to-db.rst

@@ -59,6 +59,21 @@ TimeFeatures-specific controls:
      - Destination table name. Validated against PostgreSQL identifier
        rules (see *Validation* below); MySQL accepts a superset, so
        the same rule is safe on both.
+   * - Mode
+     - Combo between the Table name and the Email fields:
+
+       - **Create new (fail if table exists)** — default, refuses to
+         touch an existing table or metadata row. Use this for first
+         uploads or when you want a clear error on accidental name
+         collisions.
+       - **Overwrite (drop and recreate)** — drops the existing data
+         table and the matching ``datasets`` row before uploading,
+         then recreates them. Re-running the workflow stops breaking.
+       - **Append (keep existing rows)** — adds rows to the existing
+         table (creating it if it doesn't exist). After the upload,
+         the widget runs ``SELECT COUNT(*)`` and rewrites the
+         ``datasets`` row with the *actual* total row count, so the
+         registry stays accurate across repeated appends.
    * - Email
      - Optional notification address. A summary email is sent once the
        upload finishes, including the table name, row / column counts,
@@ -79,12 +94,32 @@ When **Save** is clicked, the widget:
 
    - converts the dataset to a pandas DataFrame with
      ``Orange.data.pandas_compat.table_to_frame(include_metas=True)``
-     and reorders the columns (class first, then metas in domain order,
-     then attributes) to match the existing schema convention;
-   - creates the ``datasets`` metadata table if it doesn't exist;
-   - inserts one row into ``datasets`` describing this upload;
-   - writes the dataset itself in ``PANDAS_SQL_CHUNKSIZE`` chunks
-     (default ``1 000``) via ``df.to_sql(..., method='multi')``.
+     and reorders the columns (class first, then metas in domain
+     order, then attributes) to match the existing schema convention;
+   - ensures the ``datasets`` metadata table exists;
+   - applies the mode-specific preparation:
+
+     - **create** — fail fast if a ``datasets`` row with this name
+       already exists.
+     - **overwrite** — ``DROP TABLE IF EXISTS`` on the target plus
+       ``DELETE`` of the matching ``datasets`` row.
+     - **append** — leave everything in place; the upload itself will
+       create the table on the first chunk if it doesn't exist.
+
+   - writes the dataset in ``PANDAS_SQL_CHUNKSIZE`` chunks (default
+     ``1 000``) via ``df.to_sql(..., method='multi')`` with the
+     ``if_exists`` value tailored to the mode (see
+     ``_pandas_if_exists`` for the exact table);
+   - runs ``SELECT COUNT(*)`` on the target table and re-inserts the
+     ``datasets`` row with that real row count, so the registry stays
+     consistent even after multiple appends.
+
+Everything happens inside a single SQLAlchemy transaction
+(``engine.begin()``), so a mid-upload error rolls every step back on
+PostgreSQL. MySQL auto-commits DDL (``DROP TABLE``, ``CREATE TABLE``),
+so an Overwrite that crashes during the upload may leave you without
+the original table — there's no way around that without engine-level
+support.
 
 While the worker runs, the widget's progress bar and status label are
 updated through Qt signals; the **Save**, **Connect** and form controls

{timefeatures-2.0.0 → timefeatures-2.1.0}/docs/widgets/time-feature-constructor.rst

@@ -144,6 +144,33 @@ returns the appropriate slice for each chunk, so a call like
 ``shift(x, -20)`` keeps the right value across chunk boundaries even on
 multi-million-row tables.
 
+Chained descriptors
+-------------------
+
+Descriptors may reference each other. For example:
+
+.. code-block:: python
+
+   X1 := shift(price, -1)
+   X2 := X1 + bias
+
+When this happens, the widget topologically sorts the descriptors by
+their dependencies and applies them in cascade — each transform step
+runs against the table state produced by the previous step, so ``X2``
+sees ``X1`` as if it were a regular source column.
+
+- The order you click **New** in does not matter. Define ``X2`` before
+  ``X1`` and the cascade still works.
+- Cycles are detected: writing ``X1 := X2 + 1`` together with
+  ``X2 := X1 + 1`` raises a *Circular dependency between descriptors:
+  X1, X2* error instead of producing garbage.
+- Per-descriptor error reporting: if one expression fails (e.g.
+  ``shift(unknown, -1)``), the error mentions the descriptor name so
+  you know which row to fix.
+- Time-window correctness is preserved through the chain: a chained
+  ``X2`` that reads ``X1`` over a window keeps producing the right
+  values even past Orange's 5 000-row chunk boundary.
+
 Workflow persistence
 --------------------
 

{timefeatures-2.0.0 → timefeatures-2.1.0}/docs/widgets/variable-dependency-graph.rst

@@ -33,9 +33,9 @@ Outputs
      - Description
    * - Network
      - ``orangecontrib.network.Network``
-     - A directed weighted graph. ``network.edges[0].edges`` is the
-       sparse adjacency matrix; its non-zero values carry the edge
-       weights described below.
+     - A **directed** weighted graph. ``network.edges[0]`` is a
+       ``DirectedEdges`` instance whose sparse adjacency matrix carries
+       the per-edge weights described in *Edge Weights* below.
 
 How It Works
 ------------
@@ -118,6 +118,11 @@ downstream styling:
    * - ``var_type``
      - Discrete
      - ``Derived`` (has an expression) or ``Original`` (source feature).
+   * - ``expression``
+     - String
+     - The literal expression text for derived variables; empty for
+       original ones. Pick it as **Label** in Network Explorer to see
+       each derived node's formula directly on the graph.
 
 Controls
 --------
@@ -126,6 +131,15 @@ Controls
   table. The widget also auto-regenerates whenever a valid input
   arrives.
 
+Warnings
+--------
+
+- *Input has no derived variables; the dependency graph is empty.*
+  Fires when every row of the configuration table is an original
+  variable (no ``Expression`` set). The output network has nodes but
+  no edges. Usually means you forgot to attach the second output of
+  **Time Features Constructor** instead of the data output.
+
 Input Requirements
 ------------------
 

{timefeatures-2.0.0 → timefeatures-2.1.0}/setup.py

@@ -41,7 +41,7 @@ setup(name="TimeFeatures",
               "Sphinx>=7.0",
           ],
       },
-      version="2.0.0",
+      version="2.1.0",
       author="Alejandro Rivas García",
       author_email="alejandrorivasgarcia@gmail.com",
       keywords=[

{timefeatures-2.0.0 → timefeatures-2.1.0}/timefeatures/help_html/_sources/changes.rst.txt

@@ -22,6 +22,22 @@ Unreleased
 - 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**
 
@@ -31,6 +47,17 @@ Unreleased
   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
@@ -42,6 +69,16 @@ Unreleased
 
 **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
@@ -68,6 +105,15 @@ Unreleased
 
 **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

{timefeatures-2.0.0 → timefeatures-2.1.0}/timefeatures/help_html/_sources/widgets/load-from-db.rst.txt

@@ -48,9 +48,21 @@ Controls
        host/db (N datasets)", "Loaded <name> (N rows)"), error
        ("Connection failed: …", "Load failed: …").
    * - Dataset
-     - Combo populated from ``SELECT * FROM datasets ORDER BY
-       datetime DESC``. The most recent upload comes first; the last
-       choice is restored when reopening a workflow.
+     - Searchable combo populated from ``SELECT * FROM datasets ORDER
+       BY datetime DESC``. The most recent upload comes first; the
+       last choice is restored when reopening a workflow. Type to
+       filter the visible items — handy when the registry grows large.
+   * - ↻ (Refresh)
+     - Small button to the right of the Dataset combo. Re-runs the
+       list query without dropping the connection. Useful if another
+       Orange canvas (or a parallel tool) just published a new
+       dataset while this widget was open.
+   * - Delete
+     - Drops the currently selected dataset's table from the database
+       and removes its row in ``datasets``. Gated by a confirmation
+       dialog — there is no Orange-side undo. The operation runs on a
+       background thread, so the canvas stays responsive even if the
+       table is large.
    * - Dataset info
      - Read-only block under the combo: save timestamp, row/column
        counts and the original class column recorded by **Save to DB**.
@@ -60,7 +72,9 @@ Controls
        choice, (2) the ``class_name`` stored in the ``datasets``
        metadata, (3) ``(no class)`` if none of the above apply.
    * - Load
-     - Triggers the actual download.
+     - Triggers the actual download. Skipped automatically on the
+       first connection after a workflow reopen if the persisted
+       dataset name is still available (see *Auto-load* below).
 
 How it Works
 ------------
@@ -88,6 +102,28 @@ While any worker runs, the form controls (database type, connection
 fields, **Connect**, **Load**, dataset and class combos) are
 temporarily disabled and the status label keeps the user informed.
 
+Auto-load
+---------
+
+When you reopen a workflow that already had a Load from DB widget with
+a saved ``selected_dataset``, the widget fires Load automatically the
+first time the dataset listing comes back successfully. The data flows
+out of the **Data** output without a single click, mirroring how Orange
+treats sources like **File** and **Datasets**.
+
+Subtleties:
+
+- Auto-load is a *one-shot* per widget lifetime. Manually clicking
+  **Refresh** clears the pending flag, so a Refresh never surprises
+  the user by loading something behind their back.
+- If the persisted dataset no longer exists on the server (deleted
+  from outside, or the registry was wiped), the flag is cleared and
+  the widget just shows the available list — no error.
+- If a different backend or set of credentials is restored, the user
+  still has to click **Connect** explicitly, just like in earlier
+  versions; auto-load happens *after* the first successful connection
+  + listing.
+
 Workflow Persistence
 --------------------
 

{timefeatures-2.0.0 → timefeatures-2.1.0}/timefeatures/help_html/_sources/widgets/save-to-db.rst.txt

@@ -59,6 +59,21 @@ TimeFeatures-specific controls:
      - Destination table name. Validated against PostgreSQL identifier
        rules (see *Validation* below); MySQL accepts a superset, so
        the same rule is safe on both.
+   * - Mode
+     - Combo between the Table name and the Email fields:
+
+       - **Create new (fail if table exists)** — default, refuses to
+         touch an existing table or metadata row. Use this for first
+         uploads or when you want a clear error on accidental name
+         collisions.
+       - **Overwrite (drop and recreate)** — drops the existing data
+         table and the matching ``datasets`` row before uploading,
+         then recreates them. Re-running the workflow stops breaking.
+       - **Append (keep existing rows)** — adds rows to the existing
+         table (creating it if it doesn't exist). After the upload,
+         the widget runs ``SELECT COUNT(*)`` and rewrites the
+         ``datasets`` row with the *actual* total row count, so the
+         registry stays accurate across repeated appends.
    * - Email
      - Optional notification address. A summary email is sent once the
        upload finishes, including the table name, row / column counts,
@@ -79,12 +94,32 @@ When **Save** is clicked, the widget:
 
    - converts the dataset to a pandas DataFrame with
      ``Orange.data.pandas_compat.table_to_frame(include_metas=True)``
-     and reorders the columns (class first, then metas in domain order,
-     then attributes) to match the existing schema convention;
-   - creates the ``datasets`` metadata table if it doesn't exist;
-   - inserts one row into ``datasets`` describing this upload;
-   - writes the dataset itself in ``PANDAS_SQL_CHUNKSIZE`` chunks
-     (default ``1 000``) via ``df.to_sql(..., method='multi')``.
+     and reorders the columns (class first, then metas in domain
+     order, then attributes) to match the existing schema convention;
+   - ensures the ``datasets`` metadata table exists;
+   - applies the mode-specific preparation:
+
+     - **create** — fail fast if a ``datasets`` row with this name
+       already exists.
+     - **overwrite** — ``DROP TABLE IF EXISTS`` on the target plus
+       ``DELETE`` of the matching ``datasets`` row.
+     - **append** — leave everything in place; the upload itself will
+       create the table on the first chunk if it doesn't exist.
+
+   - writes the dataset in ``PANDAS_SQL_CHUNKSIZE`` chunks (default
+     ``1 000``) via ``df.to_sql(..., method='multi')`` with the
+     ``if_exists`` value tailored to the mode (see
+     ``_pandas_if_exists`` for the exact table);
+   - runs ``SELECT COUNT(*)`` on the target table and re-inserts the
+     ``datasets`` row with that real row count, so the registry stays
+     consistent even after multiple appends.
+
+Everything happens inside a single SQLAlchemy transaction
+(``engine.begin()``), so a mid-upload error rolls every step back on
+PostgreSQL. MySQL auto-commits DDL (``DROP TABLE``, ``CREATE TABLE``),
+so an Overwrite that crashes during the upload may leave you without
+the original table — there's no way around that without engine-level
+support.
 
 While the worker runs, the widget's progress bar and status label are
 updated through Qt signals; the **Save**, **Connect** and form controls

{timefeatures-2.0.0 → timefeatures-2.1.0}/timefeatures/help_html/_sources/widgets/time-feature-constructor.rst.txt

@@ -144,6 +144,33 @@ returns the appropriate slice for each chunk, so a call like
 ``shift(x, -20)`` keeps the right value across chunk boundaries even on
 multi-million-row tables.
 
+Chained descriptors
+-------------------
+
+Descriptors may reference each other. For example:
+
+.. code-block:: python
+
+   X1 := shift(price, -1)
+   X2 := X1 + bias
+
+When this happens, the widget topologically sorts the descriptors by
+their dependencies and applies them in cascade — each transform step
+runs against the table state produced by the previous step, so ``X2``
+sees ``X1`` as if it were a regular source column.
+
+- The order you click **New** in does not matter. Define ``X2`` before
+  ``X1`` and the cascade still works.
+- Cycles are detected: writing ``X1 := X2 + 1`` together with
+  ``X2 := X1 + 1`` raises a *Circular dependency between descriptors:
+  X1, X2* error instead of producing garbage.
+- Per-descriptor error reporting: if one expression fails (e.g.
+  ``shift(unknown, -1)``), the error mentions the descriptor name so
+  you know which row to fix.
+- Time-window correctness is preserved through the chain: a chained
+  ``X2`` that reads ``X1`` over a window keeps producing the right
+  values even past Orange's 5 000-row chunk boundary.
+
 Workflow persistence
 --------------------
 

{timefeatures-2.0.0 → timefeatures-2.1.0}/timefeatures/help_html/_sources/widgets/variable-dependency-graph.rst.txt

@@ -33,9 +33,9 @@ Outputs
      - Description
    * - Network
      - ``orangecontrib.network.Network``
-     - A directed weighted graph. ``network.edges[0].edges`` is the
-       sparse adjacency matrix; its non-zero values carry the edge
-       weights described below.
+     - A **directed** weighted graph. ``network.edges[0]`` is a
+       ``DirectedEdges`` instance whose sparse adjacency matrix carries
+       the per-edge weights described in *Edge Weights* below.
 
 How It Works
 ------------
@@ -118,6 +118,11 @@ downstream styling:
    * - ``var_type``
      - Discrete
      - ``Derived`` (has an expression) or ``Original`` (source feature).
+   * - ``expression``
+     - String
+     - The literal expression text for derived variables; empty for
+       original ones. Pick it as **Label** in Network Explorer to see
+       each derived node's formula directly on the graph.
 
 Controls
 --------
@@ -126,6 +131,15 @@ Controls
   table. The widget also auto-regenerates whenever a valid input
   arrives.
 
+Warnings
+--------
+
+- *Input has no derived variables; the dependency graph is empty.*
+  Fires when every row of the configuration table is an original
+  variable (no ``Expression`` set). The output network has nodes but
+  no edges. Usually means you forgot to attach the second output of
+  **Time Features Constructor** instead of the data output.
+
 Input Requirements
 ------------------