model-writtenrun 7762 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan 2018 · 1 entry
- Jing Wang began with a prefer-read extension to target_session_attrs that sought a read-only session and fell back to a writable host, supplying documentation and recovery tests in #1.
Jun - Jul 2018 · 2 entries
- Haribabu Kommi took over implementation in #2; Laurenz Albe's testing exposed an endless reconnect loop when the standby was down in #3, and Haribabu fixed it in #4.
- Laurenz advocated retaining the first writable connection, but Haribabu retained reconnection to avoid managing a cached session; Laurenz verified the fix in #6, and the subsequent refactoring and sentinel documentation appeared in #10 and #12.
Sep - Nov 2018 · 1 entry
- Tom Lane questioned both cached connections and reconnect overhead in #14 and #19, while Michael Paquier's code review called out fragile sentinel handling and argued for an explicit read-only mode in #17.
Jan - Apr 2019 · 7 entries
- The design discussion broadened to JDBC compatibility and the difference between writable sessions and primary servers: Takayuki Tsunakawa proposed reporting the server role in #28, while Robert Haas defended distinct semantics that also support logical replication in #32.
- Application-side retries during promotion were discussed in #37 and #38, but automatic failover waiting remained optional rather than part of the required feature.
- Haribabu restarted implementation as four patches in #67, separating connection-state cleanup, enum-based option handling, transaction_read_only reporting and prefer-read selection.
- Takayuki's review found enum compatibility, initialization, translation and host-tracking issues in #70; Haribabu corrected them and added read-only support in #71.
- Robert argued against a second connection parameter in #76 and against making primary an alias for read-write in #79; Haribabu added distinct primary, standby and prefer-standby modes in #81.
- The series gained completed tests, documentation and connection-rejection helpers in #82, followed by an in_recovery reporting implementation using backend signals in #85.
- Takayuki identified notification ordering, repeated GUC updates and misplaced interrupt handling in #86, leading to revised server initialization and signaling code in #87.
Sep - Oct 2019 · 2 entries
- After rebasing, Alvaro Herrera extracted and adjusted the connection-state cleanup in #90, committed it essentially unchanged as b438e7e7a1, and reposted the remaining work in #93.
- Greg Nancarrow took over with Haribabu's permission and consolidated the remaining eight patches into three in #94, preserving their implementation while improving documentation and updating version checks.
Dec 2019 - Jan 2020 · 1 entry
- Takayuki and Alvaro reviewed the consolidated series in #96 and #97; Alvaro questioned redundant modes and the complex connection-state code, then proposed dropping older-server recovery probes in #102.
May 2020 · 1 entry
- Following a long gap, Greg returned with a combined patch in #107 that added target_server_type and JDBC-style aliases, but made server-role detection differ between older and newer servers.
Jul - Sep 2020 · 5 entries
- Daniel Gustafsson caught broken tests after upstream node renaming in #110, and Greg repaired the rebase in #111.
- Peter Smith reviewed and tested that version in #112 and #113, finding obsolete code, missing cleanup and absent negative tests; Robert again challenged read-write/primary equivalence in #114.
- Greg restored separate session-writability and server-role semantics in #116, then fixed a wrong jump, duplicate query logic and documentation issues after Peter's further review in #117, producing #118.
- Tom rejected version-dependent semantics and the extra target_server_type parameter in #120, requiring consistent fallback queries and unknown/yes/no cached states so reported GUCs served only as an optimization.
- Tom also replaced the proposed backend signaling with end-of-query GUC reports and a small in_hot_standby implementation in #121, avoiding fragile asynchronous notification machinery.
Nov 2020 - Mar 2021 · 9 entries
- Tom refined the reporting prerequisite to suppress unchanged values in #127 and committed that patch exactly as 2432b1a040, confirming the push in #129.
- Greg's revised two-part series in #131 incorporated Tom's server implementation, removed the extra connection parameter and aliases, and restored consistent query-based detection when reported values were unavailable.
- Tom committed the in_hot_standby portion as bf8a662c9a, simplifying its state and completing documentation; in #132 he rejected reporting transaction_read_only because of unnecessary traffic and unreliable startup state.
- Vignesh C resumed the remaining work in #133, instead reporting default_transaction_read_only and combining it with in_hot_standby; Greg's documentation and error-buffer reviews in #134 and #135 led to the final attached version in #139.
- Tom split default_transaction_read_only reporting out of that final patch and landed it with minor documentation edits as d16f8c8e41.
- Tom landed the client feature as ee28cacf61, substantially revising the connection state machine and replacing remembered-primary fallback with a second scan of every host, as explained in #141.
- The landed interface retained any and read-write and added read-only, primary, standby and prefer-standby; the original prefer-read mode and JDBC compatibility parameter were absent from ee28cacf61.
- Vignesh reported buildfarm failures in #142, including a conditional Perl declaration handled in another discussion and an intermittent broken pipe when a deliberately failing psql connection received stdin.
- Tom reproduced the broken pipe and changed the test to pass SQL through psql -c instead of stdin in 3769e11a31; his final message #145 records that fix, with the core feature and its server prerequisites landed.