model-writtenrun 8058 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov - Dec 2018 · 4 entries
- Sergei Kornilov proposed making primary_conninfo and primary_slot_name reloadable after recovery settings became GUCs, initially letting the WAL receiver detect changes and removing connection details from its statistics view #1.
- Andres Freund and Michael Paquier argued that the statistics must retain the actual connection information, which can differ from current configuration or identify a selected host among multiple targets; Sergei restored those fields and extracted the reload handler #2 #3 #4.
- Michael requested preservation of RequestXLogStreaming's interface, while Sergei identified a configuration race between startup and receiver launch; the resulting patch retained shared-memory parameters and saved the receiver's original connection values for comparison #7 #8 #10.
- Development briefly waited for a separate discussion about configuration ownership, but Andres urged proceeding independently, and Michael and Andres agreed that hiding connection values from logs need not block this feature #14 #15 #16 #18 #19.
Jan - Apr 2019 · 3 entries
- After wording revisions, Andres rejected receiver-owned restart decisions because startup and receiver could disagree about configuration, prompting Sergei to move reload handling and shutdown into the startup process #20 #21 #22.
- Michael and Andres then objected that treating an intentional restart as source failure could send recovery into a slow archive backlog; Sergei split out state-machine refactoring and introduced an explicit restart flag #23 #25 #27 #28.
- The patch missed the PostgreSQL 12 cycle while awaiting detailed review, then returned with a test-count rebase in July; Michael's subsequent review led to a streaming-specific flag, simpler conditions, and consolidation into one patch #32 #35 #37 #38 #42.
Jul - Oct 2019 · 3 entries
- Fujii Masao caught stale sample-configuration comments and the need to document shutdown without restart when primary_conninfo becomes empty; Sergei incorporated those corrections but did not add the proposed note about delayed shutdown during paused replay #43 #44.
- Kyotaro Horiguchi supplied a sketch for bypassing archive recovery through state-machine advancement, which Sergei implemented, but Andres questioned its synthetic archive state and timeline handling #45 #46 #47.
- Kyotaro replaced that approach with a sketch that starts the receiver inside the streaming branch of the second switch; Sergei integrated it, retained timeline rescanning, and ensured receiver-start checking still runs after launch #49 #50.
Jan - Apr 2020 · 7 entries
- New temporary-slot functionality exposed a conflict with the reload patch, and Michael agreed that startup should pass the temporary-slot decision to the receiver; Sergei added a prerequisite fix and extended reload handling to wal_receiver_create_temp_slot #52 #53 #54.
- After another test rebase, Michael requested separating temporary-slot reload changes from the main feature, while Álvaro Herrera opposed postponing usable reload behavior; Sergei reconciled those requests with a three-patch series #55 #58 #59 #60 #61.
- Álvaro moved configuration rereading into startup.c and reduced detailed restart logging, while Michael reviewed the resulting series and identified a stale declaration, missing function documentation, and string-handling concerns #62 #63 #64.
- Álvaro next proposed landing connection reload first with a test workaround, but Michael preferred fixing the prerequisites first and Sergei explained that the failure was an actual stale temporary-slot-state bug #66 #67 #68.
- Sergei supplied the state reset and a separate draft permitting configured temporary-slot names; Álvaro then reported landing the temporary-slot prerequisite and posted the remaining reload feature with passing tests #69 #70 #71.
- Álvaro committed Sergei's reload feature from #71 as 1e6148032e, preserving its executable changes and tests while refining documentation and comments, as confirmed in #73.
- The temporary-slot naming draft remained outside that commit and was reposted separately; Peter Eisentraut cautioned that its semantics would diverge from pg_basebackup's use of an explicitly named permanent slot #71 #74 #75.
Oct 2020 · 1 entry
- Months later, Maxim Orlov proposed independently making restore_command reloadable and raised concerns about long-running commands and fatal configuration mistakes; Sergei replied that a separate proposal already existed without consensus, leaving that extension unresolved at the thread's end #76 #77.