pixeltable 0.4.15__py3-none-any.whl → 0.4.17__py3-none-any.whl

pixeltable/share/packager.py

@@ -1,5 +1,4 @@
 import base64
-import datetime
 import io
 import json
 import logging
@@ -13,6 +12,7 @@ from uuid import UUID
 
 import more_itertools
 import numpy as np
+import pgvector.sqlalchemy as sql_vector  # type: ignore[import-untyped]
 import PIL.Image
 import pyarrow as pa
 import pyarrow.parquet as pq
@@ -21,6 +21,7 @@ import sqlalchemy as sql
 import pixeltable as pxt
 from pixeltable import catalog, exceptions as excs, metadata, type_system as ts
 from pixeltable.env import Env
+from pixeltable.exprs.data_row import CellMd
 from pixeltable.metadata import schema
 from pixeltable.utils import sha256sum
 from pixeltable.utils.formatter import Formatter
@@ -109,9 +110,12 @@ class TablePackager:
         assert any(tv.id == base.id for base in self.table._tbl_version_path.get_tbl_versions())
         sql_types = {col.name: col.type for col in tv.store_tbl.sa_tbl.columns}
         media_cols: set[str] = set()
+        cellmd_cols: set[str] = set()
         for col in tv.cols:
             if col.is_stored and col.col_type.is_media_type():
                 media_cols.add(col.store_name())
+            if col.stores_cellmd:
+                cellmd_cols.add(col.cellmd_store_name())
 
         parquet_schema = self.__to_parquet_schema(tv.store_tbl.sa_tbl)
         # TODO: Partition larger tables into multiple parquet files. (The parquet file naming scheme anticipates
@@ -126,10 +130,10 @@ class TablePackager:
         # excessive memory usage. The pyarrow tables are then amalgamated into the (single) Parquet table on disk.
         # We use snappy compression for the Parquet tables; the entire bundle will be bzip2-compressed later, so
         # faster compression should provide good performance while still reducing temporary storage utilization.
-        parquet_writer = pq.ParquetWriter(parquet_file, parquet_schema, compression='SNAPPY')
+        parquet_writer = pq.ParquetWriter(parquet_file, parquet_schema, compression='snappy')
         filter_tv = self.table._tbl_version_path.tbl_version.get()
         row_iter = tv.store_tbl.dump_rows(tv.version, filter_tv.store_tbl, filter_tv.version)
-        for pa_table in self.__to_pa_tables(row_iter, sql_types, media_cols, parquet_schema):
+        for pa_table in self.__to_pa_tables(row_iter, sql_types, media_cols, cellmd_cols, parquet_schema):
             parquet_writer.write_table(pa_table)
         parquet_writer.close()
 
@@ -138,7 +142,7 @@ class TablePackager:
     @classmethod
     def __to_parquet_schema(cls, store_tbl: sql.Table) -> pa.Schema:
         entries = [(col_name, cls.__to_parquet_type(col.type)) for col_name, col in store_tbl.columns.items()]
-        return pa.schema(entries)  # type: ignore[arg-type]
+        return pa.schema(entries)
 
     @classmethod
     def __to_parquet_type(cls, col_type: sql.types.TypeEngine[Any]) -> pa.DataType:
@@ -151,13 +155,17 @@ class TablePackager:
         if isinstance(col_type, sql.Float):
             return pa.float32()
         if isinstance(col_type, sql.TIMESTAMP):
-            return pa.timestamp('us', tz=datetime.timezone.utc)
+            return pa.timestamp('us', tz='UTC')
         if isinstance(col_type, sql.Date):
             return pa.date32()
         if isinstance(col_type, sql.JSON):
             return pa.string()  # JSON will be exported as strings
         if isinstance(col_type, sql.LargeBinary):
             return pa.binary()
+        if isinstance(col_type, sql_vector.Vector):
+            # Parquet/pyarrow do not handle null values properly for fixed_shape_tensor(), so we have to use list_()
+            # here instead.
+            return pa.list_(pa.float32())
         raise AssertionError(f'Unrecognized SQL type: {col_type} (type {type(col_type)})')
 
     def __to_pa_tables(
@@ -165,6 +173,7 @@ class TablePackager:
         row_iter: Iterator[dict[str, Any]],
         sql_types: dict[str, sql.types.TypeEngine[Any]],
         media_cols: set[str],
+        cellmd_cols: set[str],
         arrow_schema: pa.Schema,
         batch_size: int = 1_000,
     ) -> Iterator[pa.Table]:
@@ -176,14 +185,21 @@ class TablePackager:
         for rows in more_itertools.batched(row_iter, batch_size):
             cols = {}
             for name, sql_type in sql_types.items():
-                is_media_col = name in media_cols
-                values = [self.__to_pa_value(row.get(name), sql_type, is_media_col) for row in rows]
+                values = [
+                    self.__to_pa_value(row.get(name), sql_type, name in media_cols, name in cellmd_cols) for row in rows
+                ]
                 cols[name] = values
             yield pa.Table.from_pydict(cols, schema=arrow_schema)
 
-    def __to_pa_value(self, val: Any, sql_type: sql.types.TypeEngine[Any], is_media_col: bool) -> Any:
+    def __to_pa_value(
+        self, val: Any, sql_type: sql.types.TypeEngine[Any], is_media_col: bool, is_cellmd_col: bool
+    ) -> Any:
         if val is None:
             return None
+        if is_cellmd_col:
+            assert isinstance(val, dict)
+            # Export JSON as strings
+            return json.dumps(self.__process_cellmd(val))
         if isinstance(sql_type, sql.JSON):
             # Export JSON as strings
             return json.dumps(val)
@@ -194,6 +210,10 @@ class TablePackager:
         return val
 
     def __process_media_url(self, url: str) -> str:
+        """
+        Process a media URL for export. If it's a local file URL (file://), then replace it with a pxtmedia:// URI,
+        copying the file into the tarball if necessary. If it's any other type of URL, return it unchanged.
+        """
         parsed_url = urllib.parse.urlparse(url)
         if parsed_url.scheme == 'file':
             # It's the URL of a local file. Replace it with a pxtmedia:// URI.
@@ -214,6 +234,21 @@ class TablePackager:
         # For any type of URL other than a local file, just return the URL as-is.
         return url
 
+    def __process_cellmd(self, cellmd: dict[str, Any]) -> dict[str, Any]:
+        """
+        Process a cellmd dictionary for export. This involves replacing any local file references
+        with pxtmedia:// URIs, as described above.
+        """
+        cellmd_ = CellMd.from_dict(cellmd)
+        if cellmd_.file_urls is None:
+            return cellmd  # No changes
+
+        updated_urls: list[str] = []
+        for url in cellmd_.file_urls:
+            updated_urls.append(self.__process_media_url(url))
+        cellmd_.file_urls = updated_urls
+        return cellmd_.as_dict()
+
     def __build_tarball(self) -> Path:
         bundle_path = self.tmp_dir / 'bundle.tar.bz2'
         with tarfile.open(bundle_path, 'w:bz2') as tf:
@@ -409,6 +444,9 @@ class TableRestorer:
         # 2. "rectify" the v_max values in both the temporary table and the existing table (more on this below);
         # 3. Delete any row instances from the temporary table that are already present in the existing table;
         # 4. Copy the remaining rows from the temporary table into the existing table.
+        # 5. Rectify any index columns.
+
+        # STEP 1: Import the parquet data into a temporary table.
 
         # Create a temporary table for the initial data load, containing columns for all columns present in the
         # parquet table. The parquet columns have identical names to those in the store table, so we can use the
@@ -416,7 +454,7 @@ class TableRestorer:
         # e.g., pa.string() may hold either VARCHAR or serialized JSONB).
         temp_cols: dict[str, sql.Column] = {}
         for field in parquet_table.schema:
-            assert field.name in store_sa_tbl.columns
+            assert field.name in store_sa_tbl.columns, f'{field.name} not in {list(store_sa_tbl.columns)}'
             col_type = store_sa_tbl.columns[field.name].type
             temp_cols[field.name] = sql.Column(field.name, col_type)
         temp_sa_tbl_name = f'temp_{uuid.uuid4().hex}'
@@ -432,6 +470,8 @@ class TableRestorer:
             rows = self.__from_pa_pydict(tv, pydict)
             conn.execute(sql.insert(temp_sa_tbl), rows)
 
+        # STEP 2: Rectify v_max values.
+
         # Each row version is identified uniquely by its pk, a tuple (row_id, pos_0, pos_1, ..., pos_k, v_min).
         # Conversely, v_max is not part of the primary key, but is simply a bookkeeping device.
         # In an original table, v_max is always equal to the v_min of the succeeding row instance with the same
@@ -540,6 +580,8 @@ class TableRestorer:
         result = conn.execute(q)
         _logger.debug(f'Rectified {result.rowcount} row(s) in {store_sa_tbl_name!r}.')
 
+        # STEP 3: Delete any row instances from the temporary table that are already present in the existing table.
+
         # Now we need to update rows in the existing table that are also present in the temporary table. This is to
         # account for the scenario where the temporary table has columns that are not present in the existing table.
         # (We can't simply replace the rows with their versions in the temporary table, because the converse scenario
@@ -570,7 +612,9 @@ class TableRestorer:
         result = conn.execute(q)
         _logger.debug(f'Deleted {result.rowcount} row(s) from {temp_sa_tbl_name!r}.')
 
-        # Finally, copy the remaining data (consisting entirely of new row instances) from the temporary table into
+        # STEP 4: Copy the remaining rows from the temporary table into the existing table.
+
+        # Now copy the remaining data (consisting entirely of new row instances) from the temporary table into
         # the actual table.
         q = store_sa_tbl.insert().from_select(
             [store_sa_tbl.c[col_name] for col_name in temp_cols], sql.select(*temp_cols.values())
@@ -579,39 +623,113 @@ class TableRestorer:
         result = conn.execute(q)
         _logger.debug(f'Inserted {result.rowcount} row(s) from {temp_sa_tbl_name!r} into {store_sa_tbl_name!r}.')
 
+        # STEP 5: Rectify any index columns.
+
+        # Finally, rectify any index columns in the table. This involves shuffling data between the index's val and
+        # undo columns to ensure they appropriately reflect the most recent replicated version of the table.
+
+        # Get the most recent replicated version of the table. This might be the version we're currently importing,
+        # but it might be a different version of the table that was previously imported.
+        head_version_md = catalog.Catalog.get()._collect_tbl_history(tv.id, n=1)[0]
+        head_version = head_version_md.version_md.version
+        _logger.debug(f'Head version for index rectification is {head_version}.')
+
+        # Get the index info from the table metadata. Here we use the tbl_md that we just collected from the DB.
+        # This is to ensure we pick up ALL indices, including dropped indices and indices that are present in
+        # a previously replicated version of the table, but not in the one currently being imported.
+        index_md = head_version_md.tbl_md.index_md
+
+        # Now update the table. We can do this for all indices together with just two SQL queries. For each index,
+        # at most one of the val or undo columns will be non-NULL in any given row.
+        # For rows where v_min <= head_version < v_max, we set, for all indices:
+        #   val_col = whichever of (val_col, undo_col) is non-NULL (or NULL if both are, e.g., for a dropped index)
+        #   undo_col = NULL
+        # For rows where head_version < v_min or v_max <= head_version, vice versa.
+        val_sql_clauses: dict[str, sql.ColumnElement] = {}
+        undo_sql_clauses: dict[str, sql.ColumnElement] = {}
+        for index in index_md.values():
+            if index.class_fqn.endswith('.EmbeddingIndex'):
+                val_col_name = f'col_{index.index_val_col_id}'
+                undo_col_name = f'col_{index.index_val_undo_col_id}'
+                # Check that the val column for the index is actually present in the store table. We need to do this
+                # to properly handle the case where the replica represents a table version that was *not* the most
+                # recent version at the time it was published. In that case, it is possible for tbl_md to contain
+                # metadata for indices not known to any version that has been replicated. (However, the converse
+                # *does* hold: all replicated indices must have metadata in tbl_md; and that's what's important.)
+                if val_col_name in store_sa_tbl.c:
+                    assert undo_col_name in store_sa_tbl.c
+                    coalesce = sql.func.coalesce(store_sa_tbl.c[val_col_name], store_sa_tbl.c[undo_col_name])
+                    val_sql_clauses[val_col_name] = coalesce
+                    val_sql_clauses[undo_col_name] = sql.null()
+                    undo_sql_clauses[undo_col_name] = coalesce
+                    undo_sql_clauses[val_col_name] = sql.null()
+
+        if len(val_sql_clauses) > 0:
+            q2 = (
+                store_sa_tbl.update()
+                .values(**val_sql_clauses)
+                .where(sql.and_(tv.store_tbl.v_min_col <= head_version, tv.store_tbl.v_max_col > head_version))
+            )
+            _logger.debug(q2.compile())
+            _ = conn.execute(q2)
+            q2 = (
+                store_sa_tbl.update()
+                .values(**undo_sql_clauses)
+                .where(sql.or_(tv.store_tbl.v_min_col > head_version, tv.store_tbl.v_max_col <= head_version))
+            )
+            _logger.debug(q2.compile())
+            _ = conn.execute(q2)
+            _logger.debug(f'Rectified index columns in {store_sa_tbl_name!r}.')
+        else:
+            _logger.debug(f'No index columns to rectify in {store_sa_tbl_name!r}.')
+
     def __from_pa_pydict(self, tv: catalog.TableVersion, pydict: dict[str, Any]) -> list[dict[str, Any]]:
         # Data conversions from pyarrow to Pixeltable
         sql_types: dict[str, sql.types.TypeEngine[Any]] = {}
         for col_name in pydict:
             assert col_name in tv.store_tbl.sa_tbl.columns
             sql_types[col_name] = tv.store_tbl.sa_tbl.columns[col_name].type
-        media_cols: dict[str, catalog.Column] = {}
-        for col in tv.cols:
-            if col.is_stored and col.col_type.is_media_type():
-                assert tv.id == col.tbl.id
-                assert tv.version == col.tbl.version
-                media_cols[col.store_name()] = col
+        stored_cols: dict[str, catalog.Column] = {col.store_name(): col for col in tv.cols if col.is_stored}
+        stored_cols |= {col.cellmd_store_name(): col for col in tv.cols if col.stores_cellmd}
 
         row_count = len(next(iter(pydict.values())))
-        rows: list[dict[str, Any]] = []
-        for i in range(row_count):
-            row = {
-                col_name: self.__from_pa_value(col_vals[i], sql_types[col_name], media_cols.get(col_name))
-                for col_name, col_vals in pydict.items()
-            }
-            rows.append(row)
+        rows: list[dict[str, Any]] = [{} for _ in range(row_count)]
+        for col_name, col_vals in pydict.items():
+            assert len(col_vals) == row_count
+            col = stored_cols.get(col_name)  # Will be None for system columns
+            is_media_col = col is not None and col.is_stored and col.col_type.is_media_type()
+            is_cellmd_col = col is not None and col.stores_cellmd and col_name == col.cellmd_store_name()
+            assert col is None or is_cellmd_col or col_name == col.store_name()
+
+            for i, val in enumerate(col_vals):
+                rows[i][col_name] = self.__from_pa_value(val, sql_types[col_name], col, is_media_col, is_cellmd_col)
 
         return rows
 
     def __from_pa_value(
-        self, val: Any, sql_type: sql.types.TypeEngine[Any], media_col: Optional[catalog.Column]
+        self,
+        val: Any,
+        sql_type: sql.types.TypeEngine[Any],
+        col: Optional[catalog.Column],
+        is_media_col: bool,
+        is_cellmd_col: bool,
     ) -> Any:
         if val is None:
             return None
+        if isinstance(sql_type, sql_vector.Vector):
+            if isinstance(val, list):
+                val = np.array(val, dtype=np.float32)
+            assert isinstance(val, np.ndarray) and val.dtype == np.float32 and val.ndim == 1
+            return val
+        if is_cellmd_col:
+            assert col is not None
+            assert isinstance(val, str)
+            return self.__restore_cellmd(col, json.loads(val))
         if isinstance(sql_type, sql.JSON):
             return json.loads(val)
-        if media_col is not None:
-            return self.__relocate_media_file(media_col, val)
+        if is_media_col:
+            assert col is not None
+            return self.__relocate_media_file(col, val)
         return val
 
     def __relocate_media_file(self, media_col: catalog.Column, url: str) -> str:
@@ -629,3 +747,14 @@ class TableRestorer:
             return self.media_files[url]
         # For any type of URL other than a local file, just return the URL as-is.
         return url
+
+    def __restore_cellmd(self, col: catalog.Column, cellmd: dict[str, Any]) -> dict[str, Any]:
+        cellmd_ = CellMd.from_dict(cellmd)
+        if cellmd_.file_urls is None:
+            return cellmd  # No changes
+
+        updated_urls: list[str] = []
+        for url in cellmd_.file_urls:
+            updated_urls.append(self.__relocate_media_file(col, url))
+        cellmd_.file_urls = updated_urls
+        return cellmd_.as_dict()

pixeltable/store.py

@@ -321,7 +321,7 @@ class StoreBase:
                     table_row, num_row_exc = row_builder.create_store_table_row(row, cols_with_excs, pk)
                     num_excs += num_row_exc
 
-                    if show_progress:
+                    if show_progress and Env.get().verbosity >= 1:
                         if progress_bar is None:
                             warnings.simplefilter('ignore', category=TqdmWarning)
                             progress_bar = tqdm(
@@ -434,8 +434,7 @@ class StoreBase:
             *[c1 == c2 for c1, c2 in zip(self.rowid_columns(), filter_view.rowid_columns())],
         )
         stmt = (
-            sql.select('*')  # TODO: Use a more specific list of columns?
-            .select_from(self.sa_tbl)
+            sql.select(self.sa_tbl)
             .where(self.v_min_col <= version)
             .where(self.v_max_col > version)
             .where(sql.exists().where(filter_predicate))

pixeltable/type_system.py

@@ -25,6 +25,7 @@ import sqlalchemy as sql
 from typing_extensions import _AnnotatedAlias
 
 import pixeltable.exceptions as excs
+from pixeltable.env import Env
 from pixeltable.utils import parse_local_file_path
 
 
@@ -673,8 +674,9 @@ class TimestampType(ColumnType):
     def _create_literal(self, val: Any) -> Any:
         if isinstance(val, str):
             return datetime.datetime.fromisoformat(val)
-        if isinstance(val, datetime.datetime):
-            return val
+        # Place naive timestamps in the default time zone
+        if isinstance(val, datetime.datetime) and val.tzinfo is None:
+            return val.replace(tzinfo=Env.get().default_time_zone)
         return val
 
 
@@ -760,7 +762,7 @@ class JsonType(ColumnType):
 
     @classmethod
     def __is_valid_json(cls, val: Any) -> bool:
-        if val is None or isinstance(val, (str, int, float, bool)):
+        if val is None or isinstance(val, (str, int, float, bool, np.ndarray, PIL.Image.Image)):
             return True
         if isinstance(val, (list, tuple)):
             return all(cls.__is_valid_json(v) for v in val)

pixeltable/utils/arrow.py

@@ -16,7 +16,7 @@ if TYPE_CHECKING:
 PA_TO_PXT_TYPES: dict[pa.DataType, ts.ColumnType] = {
     pa.string(): ts.StringType(nullable=True),
     pa.large_string(): ts.StringType(nullable=True),
-    pa.timestamp('us', tz=datetime.timezone.utc): ts.TimestampType(nullable=True),
+    pa.timestamp('us', tz='UTC'): ts.TimestampType(nullable=True),
     pa.bool_(): ts.BoolType(nullable=True),
     pa.int8(): ts.IntType(nullable=True),
     pa.int16(): ts.IntType(nullable=True),
@@ -35,7 +35,7 @@ PA_TO_PXT_TYPES: dict[pa.DataType, ts.ColumnType] = {
 
 PXT_TO_PA_TYPES: dict[type[ts.ColumnType], pa.DataType] = {
     ts.StringType: pa.string(),
-    ts.TimestampType: pa.timestamp('us', tz=datetime.timezone.utc),  # postgres timestamp is microseconds
+    ts.TimestampType: pa.timestamp('us', tz='UTC'),  # postgres timestamp is microseconds
     ts.DateType: pa.date32(),  # This could be date64
     ts.BoolType: pa.bool_(),
     ts.IntType: pa.int64(),
@@ -61,7 +61,7 @@ def to_pixeltable_type(arrow_type: pa.DataType, nullable: bool) -> Optional[ts.C
         dtype = to_pixeltable_type(arrow_type.value_type, nullable)
         if dtype is None:
             return None
-        return ts.ArrayType(shape=arrow_type.shape, dtype=dtype, nullable=nullable)
+        return ts.ArrayType(shape=tuple(arrow_type.shape), dtype=dtype, nullable=nullable)
     else:
         return None
 
@@ -92,7 +92,7 @@ def to_pxt_schema(
 
 
 def to_arrow_schema(pixeltable_schema: dict[str, Any]) -> pa.Schema:
-    return pa.schema((name, to_arrow_type(typ)) for name, typ in pixeltable_schema.items())  # type: ignore[misc]
+    return pa.schema((name, to_arrow_type(typ)) for name, typ in pixeltable_schema.items())
 
 
 def _to_record_batch(column_vals: dict[str, list[Any]], schema: pa.Schema) -> pa.RecordBatch:
@@ -106,7 +106,7 @@ def _to_record_batch(column_vals: dict[str, list[Any]], schema: pa.Schema) -> pa
         else:
             pa_array = cast(pa.Array, pa.array(column_vals[field.name]))
             pa_arrays.append(pa_array)
-    return pa.RecordBatch.from_arrays(pa_arrays, schema=schema)  # type: ignore
+    return pa.RecordBatch.from_arrays(pa_arrays, schema=schema)
 
 
 def to_record_batches(df: 'pxt.DataFrame', batch_size_bytes: int) -> Iterator[pa.RecordBatch]:
@@ -192,7 +192,7 @@ def to_pydict(batch: pa.Table | pa.RecordBatch) -> dict[str, list | np.ndarray]:
         col = batch.column(k)
         if isinstance(col.type, pa.FixedShapeTensorType):
             # treat array columns as numpy arrays to easily preserve numpy type
-            out[name] = col.to_numpy(zero_copy_only=False)  # type: ignore[call-arg]
+            out[name] = col.to_numpy(zero_copy_only=False)
         else:
             # for the rest, use pydict to preserve python types
             out[name] = col.to_pylist()

pixeltable/utils/av.py

@@ -3,6 +3,8 @@ from typing import Any
 import av
 import av.stream
 
+from pixeltable.env import Env
+
 
 def get_metadata(path: str) -> dict:
     with av.open(path) as container:
@@ -109,3 +111,66 @@ def ffmpeg_clip_cmd(input_path: str, output_path: str, start_time: float, durati
         ]
     )
     return cmd
+
+
+def ffmpeg_segment_cmd(
+    input_path: str,
+    output_pattern: str,
+    segment_duration: float | None = None,
+    segment_times: list[float] | None = None,
+    video_encoder: str | None = None,
+    video_encoder_args: dict[str, Any] | None = None,
+) -> list[str]:
+    """Commandline for frame-accurate segmentation"""
+    assert (segment_duration is None) != (segment_times is None)
+    if video_encoder is None:
+        video_encoder = Env.get().default_video_encoder
+
+    cmd = [
+        'ffmpeg',
+        '-i',
+        input_path,
+        '-f',
+        'segment',  # Use segment muxer
+    ]
+
+    if segment_duration is not None:
+        cmd.extend(
+            [
+                '-segment_time',
+                str(segment_duration),  # Target segment duration
+                '-break_non_keyframes',
+                '1',  # need to break at non-keyframes to get frame-accurate segments
+                '-force_key_frames',
+                f'expr:gte(t,n_forced*{segment_duration})',  # Force keyframe at each segment boundary
+            ]
+        )
+    else:
+        assert segment_times is not None
+        times_str = ','.join([str(t) for t in segment_times])
+        cmd.extend(['-segment_times', times_str, '-force_key_frames', times_str])
+
+    cmd.extend(
+        [
+            '-reset_timestamps',
+            '1',  # Reset timestamps for each segment
+            '-map',
+            '0',  # Copy all streams from input
+            '-c:a',
+            'copy',  # don't re-encode audio
+            '-c:v',
+            video_encoder,  # re-encode video
+        ]
+    )
+    if video_encoder_args is not None:
+        for k, v in video_encoder_args.items():
+            cmd.extend([f'-{k}', str(v)])
+
+    cmd.extend(
+        [
+            '-loglevel',
+            'error',  # Only show errors
+            output_pattern,
+        ]
+    )
+    return cmd

pixeltable/utils/console_output.py

@@ -1,6 +1,8 @@
 import logging
 from typing import TextIO
 
+from pixeltable import exceptions as excs
+
 
 def map_level(verbosity: int) -> int:
     """
@@ -19,7 +21,8 @@ def map_level(verbosity: int) -> int:
         return logging.INFO
     if verbosity == 2:
         return logging.DEBUG
-    return logging.INFO
+
+    raise excs.Error(f'Invalid verbosity level: {verbosity}')
 
 
 class ConsoleOutputHandler(logging.StreamHandler):

pixeltable/utils/exception_handler.py

@@ -1,32 +1,9 @@
 import logging
-import sys
 from typing import Any, Callable, Optional, TypeVar
 
 R = TypeVar('R')
 
-
-def _is_in_exception() -> bool:
-    """
-    Check if code is currently executing within an exception context.
-    """
-    current_exception = sys.exc_info()[1]
-    return current_exception is not None
-
-
-def run_cleanup_on_exception(cleanup_func: Callable[..., R], *args: Any, **kwargs: Any) -> Optional[R]:
-    """
-    Runs cleanup only when running in exception context.
-
-    The function `run_cleanup_on_exception()` should be used to clean up resources when an operation fails.
-    This is typically done using a try, except, and finally block, with the resource cleanup logic placed within
-    the except block. However, this pattern may not handle KeyboardInterrupt exceptions.
-    To ensure that resources are always cleaned up at least once when an exception or KeyboardInterrupt occurs,
-    create an idempotent function for cleaning up resources and pass it to the `run_cleanup_on_exception()` function
-    from the finally block.
-    """
-    if _is_in_exception():
-        return run_cleanup(cleanup_func, *args, raise_error=False, **kwargs)
-    return None
+logger = logging.getLogger('pixeltable')
 
 
 def run_cleanup(cleanup_func: Callable[..., R], *args: Any, raise_error: bool = True, **kwargs: Any) -> Optional[R]:
@@ -40,20 +17,20 @@ def run_cleanup(cleanup_func: Callable[..., R], *args: Any, raise_error: bool =
         raise_error: raise an exception if an error occurs during cleanup.
     """
     try:
-        logging.debug(f'Running cleanup function: {cleanup_func.__name__!r}')
+        logger.debug(f'Running cleanup function: {cleanup_func.__name__!r}')
         return cleanup_func(*args, **kwargs)
     except KeyboardInterrupt as interrupt:
         # Save original exception and re-attempt cleanup
         original_exception = interrupt
-        logging.debug(f'Cleanup {cleanup_func.__name__!r} interrupted, retrying')
+        logger.debug(f'Cleanup {cleanup_func.__name__!r} interrupted, retrying')
         try:
             return cleanup_func(*args, **kwargs)
         except Exception as e:
             # Suppress this exception
-            logging.error(f'Cleanup {cleanup_func.__name__!r} failed with exception {e}')
+            logger.error(f'Cleanup {cleanup_func.__name__!r} failed with exception {e.__class__}: {e}')
         raise KeyboardInterrupt from original_exception
     except Exception as e:
-        logging.error(f'Cleanup {cleanup_func.__name__!r} failed with exception {e}')
+        logger.error(f'Cleanup {cleanup_func.__name__!r} failed with exception {e.__class__}: {e}')
         if raise_error:
             raise e
     return None

pixeltable/utils/image.py

@@ -0,0 +1,7 @@
+import PIL.Image
+
+
+def default_format(img: PIL.Image.Image) -> str:
+    # Default to JPEG unless the image has a transparency layer (which isn't supported by JPEG).
+    # In that case, use WebP instead.
+    return 'webp' if img.has_transparency_data else 'jpeg'

pixeltable/utils/misc.py

@@ -0,0 +1,5 @@
+from typing import Any
+
+
+def non_none_dict_factory(d: list[tuple[str, Any]]) -> dict:
+    return {k: v for (k, v) in d if v is not None}

pixeltable/utils/object_stores.py

@@ -22,6 +22,7 @@ class StorageTarget(enum.Enum):
     LOCAL_STORE = 'os'  # Local file system
     S3_STORE = 's3'  # Amazon S3
     R2_STORE = 'r2'  # Cloudflare R2
+    B2_STORE = 'b2'  # Backblaze B2
     GCS_STORE = 'gs'  # Google Cloud Storage
     AZURE_STORE = 'az'  # Azure Blob Storage
     HTTP_STORE = 'http'  # HTTP/HTTPS
@@ -63,6 +64,7 @@ class StorageObjectAddress(NamedTuple):
             StorageTarget.LOCAL_STORE,
             StorageTarget.S3_STORE,
             StorageTarget.R2_STORE,
+            StorageTarget.B2_STORE,
             StorageTarget.GCS_STORE,
             StorageTarget.AZURE_STORE,
             StorageTarget.HTTP_STORE,
@@ -218,15 +220,23 @@ class ObjectPath:
             # Standard HTTP(S) URL format
             # https://account.blob.core.windows.net/container/<optional path>/<optional object>
             # https://account.r2.cloudflarestorage.com/container/<optional path>/<optional object>
+            # https://s3.us-west-004.backblazeb2.com/container/<optional path>/<optional object>
             # and possibly others
             key = parsed.path
             if 'cloudflare' in parsed.netloc:
                 storage_target = StorageTarget.R2_STORE
+            elif 'backblazeb2' in parsed.netloc:
+                storage_target = StorageTarget.B2_STORE
             elif 'windows' in parsed.netloc:
                 storage_target = StorageTarget.AZURE_STORE
             else:
                 storage_target = StorageTarget.HTTP_STORE
-            if storage_target in [StorageTarget.S3_STORE, StorageTarget.AZURE_STORE, StorageTarget.R2_STORE]:
+            if storage_target in (
+                StorageTarget.S3_STORE,
+                StorageTarget.AZURE_STORE,
+                StorageTarget.R2_STORE,
+                StorageTarget.B2_STORE,
+            ):
                 account_name = parsed.netloc.split('.', 1)[0]
                 account_extension = parsed.netloc.split('.', 1)[1]
                 path_parts = key.lstrip('/').split('/', 1)
@@ -370,6 +380,11 @@ class ObjectOps:
             env.Env.get().require_package('boto3')
             from pixeltable.utils.s3_store import S3Store
 
+            return S3Store(soa)
+        if soa.storage_target == StorageTarget.B2_STORE:
+            env.Env.get().require_package('boto3')
+            from pixeltable.utils.s3_store import S3Store
+
             return S3Store(soa)
         if soa.storage_target == StorageTarget.GCS_STORE and soa.scheme == 'gs':
             env.Env.get().require_package('google.cloud.storage')