Hi hackers,
I'd like to propose a patch that makes the ReScanForeignScan callback
optional for Foreign Data Wrappers, allowing the planner to automatically
handle non-rescannable FDWs.
# Background
Currently, FDWs are expected to implement ReScanForeignScan to support
rescanning in scenarios like nested loop joins. However, some FDWs just cannot
implement rescanning:
- Streaming data sources (Kafka, message queues)
- One-time token-based APIs with expensive re-authentication
- Data sources where re-fetching is prohibitively expensive
- ...
Right now, these FDWs either implement a stub ReScanForeignScan that fails
at runtime, or they buffer all data in BeginForeignScan to support rescan,
which wastes memory when rescanning isn't needed.
# Proposed Solution
This patch introduces a 'rescannable' field in the Path structure that
tracks whether each path supports rescanning. The planner uses this
information to:
1. Auto-detect FDW rescan capability by checking if ReScanForeignScan is
provided
2. Automatically insert Material nodes when non-rescannable paths are used
as inner paths in nested loops
3. Reject parameterized foreign scan paths if the FDW doesn't support
rescan (preventing planning failures)
4. Raise a clear error for correlated subqueries that cannot be handled
(these are planned independently and can't use Material nodes)
# Not only FDWs
Beyond enabling non-rescannable FDWs, this mechanism could also be used
for performance optimization on other nodes. Some operations can technically
support rescan but at significant cost (MergeAppend redo all the sorts,
Aggregation redo all the calculations...). We may could mark such paths as
non-rescannable in some cases to encourage the planner to materialize results
instead.
The patch is attached. If reviewers feel this is a good idea and needs more
discussion or the complexity warrants it, I'm happy to register this for the
next CommitFest. For now, I'm sharing it here to gather initial feedback on the
approach.
--
Adam
