model-writtenrun 5574 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan - Apr 2021 · 10 entries
- Bruce Momjian opened the thread with Masahiko Sawada's key manager and Stephen Frost's tests, then addressed OpenSSL zero-length-input behavior and reduced the large cryptographic test corpus before posting the first complete attachments #1 #3 #7 #12.
- Masahiko supplied a buffer-encryption proof of concept demonstrating that the key manager could serve its intended consumer, while nonce generation and hint-bit updates remained unresolved #17 #26.
- Stephen Frost explored reserving page space for authentication information, and Andres Freund identified unsafe WAL insertion under a spinlock, replay problems, and unnecessary page copying in the encryption prototype #25 #41.
- Bruce incorporated Michael Paquier's revised hex interface and subsequently removed that prerequisite after it landed, leaving the key manager and its tests for review #31 #44.
- Andres Freund and Robert Haas required clearer in-tree design explanations, suitable tests, and further review before landing the otherwise incomplete feature; Bruce responded with seven components, READMEs, Perl tests, and support for future encryption methods #50 #64 #68 #71.
- Alvaro Herrera and Robert Haas's reviews led to OpenSSL startup corrections, clearer initdb interactions, implementation comments, assertions, and wait-event documentation across the next revisions #72 #76 #77.
- Bruce and Stephen debated integrity protection, eventually accepting an initial confidentiality-focused feature with its limitations explicitly documented and tamper detection deferred #86 #96 #98.
- Bruce returned in March with a thirteen-component proposal adding working heap/index encryption, hint-bit LSN handling, GiST changes, and extensive documentation, while WAL encryption, temporary files, and backup work remained incomplete #101.
- Stephen recommended landing the independent permanence macro, and Bruce committed that unchanged component from #102 as 95d77149c5 #103 #111.
- Alvaro's review corrected WAL-skipping terminology, while Neil Chen subsequently identified incorrect hint-handling behavior for unencrypted clusters; Bruce fixed it and reposted the twelve components remaining after the macro landed #109 #113 #114.
May 2021 · 5 entries
- Robert Haas reopened nonce design with Bharath Rupireddy's prototype reserving explicit nonce bytes at the end of page special space, seeking to avoid repeated hint-bit WAL logging and dependence on page LSN uniqueness #115.
- Robert explained why excluding hint bits from encryption would not cover all unlogged page changes and proposed a server-start random value plus a counter, while Bruce objected to the added complexity and migration costs #167 #171 #177.
- Andres demonstrated that replayed full-page images can zero free space that still contains old data on the primary, making LSN reuse unsafe for the proposed CTR design; Robert agreed that this required a different approach #179 #202.
- Discussion turned toward AES-XTS without page-format changes and a possible separate authenticated mode, with physical identifiers in encryption tweaks creating a reason to preserve database OIDs and relfilenodes across pg_upgrade #198 #210 #221 #227.
- Stephen outlined identifier preservation as independently useful for backups, and Bruce identified the special handling needed for preexisting pg_largeobject storage #228 #230.
Jun - Sep 2021 · 5 entries
- Bruce posted his final twelve-part encryption snapshot, reported passing check-world and private tests, stopped work on the feature, and subsequently marked its CommitFest entry withdrawn #231 #233.
- Shruthi Gowda began a separate pg_upgrade series by extracting Antonin Houska's earlier work and fixing dump identifiers, inappropriate storage creation, database-OID restoration, and postgres-database handling #234.
- Robert requested separate relfilenode/tablespace and database-OID patches; Tom Lane identified restore-order hazards in the separate database-OID archive entry, and Stephen Frost argued that CREATE DATABASE OID should also be available to ordinary users #235 #236 #237 #245.
- The discussion established independent benefits for backup synchronization and upgrade diagnostics, while fixed initial database OIDs and restrictions on low user-assigned OIDs emerged as the way to avoid upgrade collisions #263 #267 #279 #280.
- Shruthi refined relfilenode overrides and storage creation, implemented CREATE DATABASE OID with a fixed template0 identifier, and simplified pg_upgrade's now-redundant mappings in response to Robert's reviews #284 #285 #288.
Oct - Dec 2021 · 4 entries
- The broader encryption discussion continued alongside those prerequisites: Bruce reported support for starting with XTS without tamper detection, while Robert insisted that its security requirements be settled before ruling out stored nonces #319 #334.
- Ants Aasma demonstrated an OpenSSL XTS approach leaving the LSN exposed, and discussion examined repeated tweaks, fake LSNs for unlogged relations, and recognition of uninitialized zero pages without reaching a complete implementation agreement #333 #342 #353 #359 #370.
- Shruthi revised heap_create() storage semantics, and Sadhuprasad Patro's later database-patch review exposed missing checks for a requested OID's existing directory; explicit conflict checks were added to avoid dangerous failure cleanup #373 #376 #379 #384.
- Robert revised the relfilenode component to centralize storage decisions, handle historical partitioned-table relfilenodes, and simplify comments; tushar then reported successful relfilenode and tablespace preservation across several upgrade paths #380 #382 #383.
Jan - Apr 2022 · 8 entries
- After Shruthi rebased both components, Robert committed the relfilenode/tablespace patch from #386 as 9a974cbcba, adding a catalog-version bump and leaving database OIDs for the remaining patch #387.
- For the database remainder, Robert removed ineffective grammar changes and requested a fixed postgres OID as well as template0's; Shruthi implemented that extension and refined errors and comments #388 #390 #392.
- Robert and Tom rejected fixed identifiers taken from counter-managed OID ranges, so Shruthi switched to manually reserved OIDs 4 and 5 and registered both in unused_oids #394 #395 #396 #398.
- Robert committed #398 as aa01051418 with only comment and documentation edits, completing the identifier-preservation series #400.
- Antonin moved renewed encryption development to another thread, while Robert later closed the pg_upgrade CommitFest entry and identified a separate file-descriptor issue needing attention #402 #405.
- Tom Lane subsequently found that the fixed database OIDs appeared pinned and were invisible to Catalog.pm's duplicate detection, initially proposing provisional pg_database.dat entries to register them #406.
- After Robert questioned those provisional rows, Tom replaced them with DECLARE_OID_DEFINING_MACRO support, standardized database macro names, and made IsPinnedObject report that no database is pinned #407 #408 #410.
- Robert approved both revised fixes, and Tom combined them unchanged from #410 into 2cb1272445 #411 #412.
Jun 2022 · 3 entries
- In June Robert revived the page-layout discussion through a wider-checksum proposal, prompting debate about reserved page space, runtime size calculations, algorithm choice, performance, and a separate checksum fork #413 #416 #420 #421.
- Robert proposed sharing reserved space between extended checksums and future authentication data, while Matthias van de Meent and Peter Geoghegan pressed for access-method compatibility and page-format flags useful to diagnostic tools #431 #437 #442 #450.
- The thread ended with identifier preservation and its OID-infrastructure correction committed, Bruce's encryption series withdrawn, and no further attached checksum implementation; metadata exposure, debuggability, stored nonces, and tamper detection remained disputed #461 #462 #463 #464.