[Append Scan] Introduce an AbstractTableScan with default methods
#2230
Conversation
| @@ -1630,16 +1630,17 @@ def _parse_row_filter(expr: Union[str, BooleanExpression]) -> BooleanExpression: | |||
| return parser.parse(expr) if isinstance(expr, str) else expr | |||
|
|
|||
|
|
|||
| S = TypeVar("S", bound="TableScan", covariant=True) | |||
| A = TypeVar("A", bound="AbstractTableScan", covariant=True) | |||
There was a problem hiding this comment.
GitHub's diff here got messed up - TableScan and S remains, just below.
|
|
||
| return type(self)(**{**kwargs, **overrides}) | ||
|
|
||
| def to_pandas(self, **kwargs: Any) -> pd.DataFrame: |
There was a problem hiding this comment.
Highlighting this change. I've added default implementations based on to_arrow() in this class, whereas before previously these methods were abstract on TableScan.
This technically changes the TableScan class in a user-facing way because it now has default implementations, but this felt fine to me. Subclasses can still override these methods if they wish.
The motivation here was to reduce duplication between DataScan and IncrementalAppendScan - by introducing these default implementations in the superclass, all table scans get those defaults for free.
| """ | ||
| return self.to_arrow().to_pandas(**kwargs) | ||
|
|
||
| def to_duckdb(self, table_name: str, connection: Optional[DuckDBPyConnection] = None) -> DuckDBPyConnection: |
There was a problem hiding this comment.
Highlighting this also - to_duckdb and to_ray were previously only present on DataScan, not even on TableScan. With the change here moving them to AbstractTableScan, TableScan now has these additional methods. I felt like this was fine. Curious for reviewers' thoughts!
(Again, the motivation here was to reduce duplication)
| @@ -1688,29 +1786,6 @@ def projection(self) -> Schema: | |||
|
|
|||
| return current_schema.select(*self.selected_fields, case_sensitive=self.case_sensitive) | |||
|
|
|||
| @abstractmethod | |||
There was a problem hiding this comment.
Pointing out that these removed methods have just been moved to AbstractTableScan, so they still exist on TableScan that subclasses it. The goal of this PR is reorganising while not introducing a breaking change.
| S = TypeVar("S", bound="TableScan", covariant=True) | ||
|
|
||
|
|
||
| class TableScan(AbstractTableScan, ABC): |
There was a problem hiding this comment.
Methods on this class like use_ref and snapshot don't translate nicely into incremental scans. I understand that the motivation behind this class was an abstract class for table scans, like Java has, but I felt like a less restrictive superclass is better for us, hence AbstractTableScan.
Rationale for this change
Split up from incremental append scan work - see #2031 (comment). PyIceberg doesn't support incremental reading of appended data between snapshots, like Spark does.
This PR introduces an
AbstractTableScanto reduce duplication between table scans and provide a superclass not specific to reads of a single snapshot.Are these changes tested?
N/A
Are there any user-facing changes?
Yes, see #2230 (comment) / #2230 (comment).