model-writtenrun 7514 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - May 2021 ยท 19 entries
- Mark Dilger opened the thread after PostgresNode passed an unsupported initdb -N option to older servers, proposing that nodes obtain their version from pg_config before initialization #1.
- Early discussion also exposed invalid install paths silently falling back to unrelated binaries; Alvaro Herrera requested validation, and Mark's preliminary work explicitly selected the installation's pg_config #5 #6.
- Mark then encountered startup timing differences, prompting Alvaro Herrera and Andrew Dunstan to recommend pg_ctl -w; the first indexed series combined version-dependent behavior with a framework for exercising older installations #13 #14 #15 #18.
- Andrew's review corrected socket-setting compatibility and simplified command construction, while Mark's next revision also diagnosed IPC::Run caching executable paths across PATH changes and introduced explicitly qualified commands #19 #20.
- Jehan-Guillaume de Rorthais introduced his existing subclass-based compatibility approach, arguing that backups and replication would require too many conditional branches; Alvaro favored that design and proposed keeping the subclasses together in PostgresNode.pm #21 #25 #30.
- The participants distinguished implementation structure from behavioral compatibility: Mark wanted reliable cross-version tests, while Andrew and Alvaro favored preserving sensible existing behavior and rejecting unavailable operations instead of forcing identical behavior across releases #44 #45 #46 #48.
- Mark's next series retained conditional version checks while applying Andrew's review of the testing framework: CrossVersion.pm moved into the shared Perl directory, returned installation mappings, and accepted an environment-selected configuration file #22 #49.
- Andrew then replaced conditional compatibility checks with version-specific subclasses and reported basic lifecycle tests working down to PostgreSQL 7.2 #50.
- Jehan-Guillaume independently submitted his broader replication and backup implementation with a separate PostgresVersion class, and Andrew proposed combining their work #51 #52 #53.
- Andrew's combined implementation used a minimal PostgresVersion object and smaller overridable methods to avoid both repeated version tests and duplication of large routines #54.
- Concerned about the invasiveness and maintenance burden of all the compatibility approaches, Andrew proposed landing only executable-path fixes and version detection, with a warning below PostgreSQL 12; Mark agreed to defer broader compatibility until the next development cycle #56 #59.
- Andrew nevertheless preserved another compatibility prototype that used adjust_conf() to modify configuration files after initialization, substantially reducing per-setting overrides while still passing his tests down to PostgreSQL 7.2 #58 #60.
- Discussion left broader packaging and test coverage unresolved: Jehan-Guillaume considered an external module, while Mark advocated a shared implementation in core and suggested merging their cross-version test work #62 #63 #64 #65.
- Andrew committed Mark's IPC::Run workaround as 95c3a1956e, extracting the helper and five call-site changes from #60, then posted the reduced version-awareness remainder #66.
- Michael Paquier requested readable development-version strings and module documentation, which Andrew supplied before committing #68 with a small diagnostic addition and cleanup as 4c4eaf3d19 #67 #72.
- After that landing, Alvaro proposed simplifying version parsing and object construction, while Mark raised invalid-input handling; Andrew supplied an inline parser simplification, with error handling still discussed separately #71 #73 #74 #75.
- Michael then raised beta releases, and the discussion chose to treat all prerelease markers equivalently for comparison; Andrew's follow-up stored the version string separately, Michael tested it, and it landed unchanged as aa271209f6 #77 #79 #80 #81 #82.
- A month later, Amit Kapila reported Windows TAP failures caused by the list-form pipe open introduced with version detection; Andrew identified the older native Windows Perl limitation, proposed qx{} instead, and reported pushing a compatibility fix #83 #84 #85 #88.
- Amit confirmed that the Windows fix worked, and Andrew chose to retain the existing Perl requirements for now; the thread ended discussing buildfarm coverage for older Windows Perl, with the three candidate changes landed but broad old-server compatibility and the cross-version testing framework still deferred #92 #93 #94 #95.