model-writtenrun 5762 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May - Jun 2021 · 3 entries
- Haiying Tang reported that updates failed to reach a logical subscriber when an unchanged primary key was stored externally, reproduced the problem on PostgreSQL 10–13, and found that REPLICA IDENTITY FULL worked around it #1 #3 #5.
- Dilip Kumar identified the missing WAL data: an unchanged external key appeared neither as a complete value in the new tuple nor in the old key, and proposed always extracting it when the tuple contained external data #7.
- Tang's tests exposed unnecessary logging when only a non-key column was toasted and a missing copy-flag reset; Dilip incorporated those fixes, Kuntal Ghosh's request for clearer comments, and the existing regression test's changed output #10 #13 #15 #17.
Jul - Sep 2021 · 3 entries
- Amit Kapila explored avoiding extra WAL by fetching the key on the subscriber, but Dilip explained that finding the row already required that key, and Amit acknowledged that publisher TOAST pointers could differ on the subscriber #20 #23 #24.
- Alvaro Herrera questioned the extraction API, and Amit proposed computing a key_required flag during the existing column scan; Dilip implemented that redesign to avoid constructing and discarding unnecessary key tuples #26 #27 #28.
- Ajin Cherian tested the redesign and asked whether the detoasted key could instead enter the new tuple, but Dilip explained that doing so would interfere with physical WAL replay and retained the separate replica-identity tuple #29 #30 #31.
Jan - Feb 2022 · 8 entries
- After several quiet months, Michael Paquier and Euler Taveira renewed concerns about the helper APIs; Amit defended avoiding extra tuple processing, while Andres Freund clarified that an unchanged key produces no new TOAST chunks for decoding to reconstruct #32 #33 #34 #38.
- Following Amit's refactoring proposal, Dilip moved attribute extraction into the renamed HeapDetermineColumnsInfo and separated refactoring from the fix for review; Amit then merged the patches, corrected whole-tuple and system-attribute bookkeeping, and added documentation #40 #41 #42.
- Amit prepared patches through PostgreSQL 10 and refined helper names, then incorporated Alvaro's clearer explanation of when externally stored old values are logged; Alvaro explicitly supported backpatching #44 #45 #47 #50.
- Andres noticed that the regression output hid the changed new-tuple data behind a 200-character limit, and Tang supplied a test delta using his proposed string redaction; Amit deferred this to a separate development-branch improvement because broader output changes and deduplication deserved separate treatment #48 #52 #58 #59.
- Tang tested the revised fix and suggested consistent parameter naming and skipping repeated external-data checks; Euler and Amit rejected the additional check as unnecessary complexity #55 #56 #57.
- Amit found that system attributes could reach TupleDescAttr incorrectly, and Tang submitted the final branch patches with an explicit system-attribute guard, descriptor reuse, and consistent parameter naming, reporting that all branch tests passed #60 #61.
- Amit accepted the final submission and committed it as 5e01001ffb, crediting Dilip and himself as authors and backpatching through PostgreSQL 10; the available diff and revision descriptions support an unchanged landing of #61 #62 #63.
- The thread closed with acknowledgments after the bug fix landed; Tang's test-output improvement remained deferred, with no landing shown in this thread #59 #64 #65.