model-writtenrun 9503 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul 2015 · 1 entry
- Robbie Harwood opened the thread with an eight-patch GSSAPI encryption implementation that automatically upgraded compatible connections, offered controls to require encryption, and supported older peers #1.
Aug - Oct 2015 · 3 entries
- Michael Paquier's testing and documentation review exposed synchronization problems, while Andres Freund identified buffer leaks, header/GUC issues and misplaced encryption logic; Robbie revised the documentation and consolidated the implementation without yet changing its protocol-level design #3 #13 #6 #18.
- Michael Paquier argued that encryption should operate below PostgreSQL messages through dedicated read/write functions, and Andres Freund rejected the original architecture; Robbie stopped pursuing an additional synchronization failure reported by Jeff Janes while preparing that redesign #28 #31 #32 #36 #37.
- The interoperability discussion established that existing Windows SSPI clients must remain usable even though native SSPI encryption support could be deferred #20 #23 #25.
Feb - Apr 2016 · 6 entries
- Robbie returned in February 2016 with GSS encryption alongside SSL in the secure I/O layer, temporarily omitting fallback so reviewers could assess the new architecture #39.
- Review and connection testing led Robbie to separate common helpers, encryption and authentication cleanup, restore compatibility fallback, and explicitly handle the transition from plaintext AUTH_REQ_OK to encrypted buffered data #48 #53.
- David Steele and Michael Paquier's tests drove fixes for login hangs and compatibility in both directions; subsequent revisions corrected receive-buffer handling, relaxed incompatible GSS flag requirements, and improved complete error reporting and documentation #58 #61 #62 #66 #69 #72.
- Michael Paquier supplied cumulative Windows build fixes and found large-result crashes, Alvaro Herrera challenged writev portability, and Tom Lane rejected representing connection negotiation as a USERSET GUC; Robbie incorporated these findings into the next full series #73 #74 #77 #79 #81.
- The crash survived an initial buffer-reset cleanup until Michael identified the memory-context problem; Robbie moved buffers to connection-lifetime storage and simplified writes, after which Michael and David reported successful tests #84 #85 #87 #88 #92.
- Tom Lane then identified that usernames, database names and PGOPTIONS still crossed the network before encryption, requiring an SSL-like negotiation before the real startup packet; Stephen Frost deferred the feature because this required substantial redesign #99 #101 #104.
Jun - Aug 2016 · 1 entry
- Robbie outlined a separate GSS handshake, and discussion weighed strict require/disable modes against opportunistic encryption; Robbie proposed using existing credentials without prompting, while Stephen supported a prefer mode and visible connection-security information #106 #113 #114 #115.
May - Jun 2018 · 2 entries
- After more than two years, Robbie revived the work as one patch with pre-startup GSS negotiation, credential-cache probing, hostgss/hostnogss rules and disable/prefer/require connection modes #116.
- Thomas Munro's build and documentation findings led to SGML and Windows corrections, while Nico Williams's review tightened GSS success checks and clarified partial reads across the next revisions #118 #120 #121 #123 #130 #132.
Aug 2018 · 1 entry
- Stephen Frost explained that GSS encryption let existing Kerberos deployments secure traffic without also administering SSL certificates, then tested compatibility combinations and requested shutdown handling, comments and formatting improvements incorporated by Robbie #136 #142 #145.
Feb - Apr 2019 · 14 entries
- Stephen added an experimental psql security-strength display, and Robbie expanded it toward MIT/Heimdal algorithm inspection after identifying portability limits in the SASL extension #146 #147 #156.
- Andres Freund proposed generic connection callbacks, but Robbie and Stephen retained direct SSL/GSS branches because a general stream framework was unnecessary for the supported encryption methods #152 #153 #154.
- Peter Eisentraut requested automated coverage and clearer separation between GSS authentication and encryption naming, prompting Robbie to add a Kerberos encryption test script #159 #161.
- Stephen took over preparation of the full patch, folded in the side additions and tests, and replaced dynamically growing string buffers with fixed buffers and gss_wrap_size_limit checks to bound memory consumption #162.
- Robbie confirmed that lucid-context export destroyed the active security context, so Stephen removed algorithm and strength introspection, retained a simple encryption-status API, added pg_stat_gssapi, and simplified setup I/O #163 #164.
- Peter's renewed naming review led Stephen to rename the HBA types to hostgssenc/hostnogssenc and the client option to gssencmode, while preserving the established gss authentication method #167 #171.
- Stephen committed the complete patch from #171 unchanged as b0b39f72b9, crediting Robbie's authorship and his own I/O changes, and confirmed that the feature was complete and the buildfarm was happy #173.
- After landing, Michael Paquier identified missing wire-protocol documentation, and Peter Eisentraut reported macOS test failures with a misleading oversized-error diagnostic #176 #179.
- Tom Lane found missing GSS function declarations when building without OpenSSL; Stephen corrected the conditional declarations and reported pushing the fix #181 #184 #187.
- Tom also traced unusually slow Kerberos tests to repeated KDC creation blocking on random data and proposed sharing one setup, while Stephen separately fixed the misleading error diagnostic without yet resolving Peter's underlying checksum failures #190 #192 #196 #197.
- Peter established that the encryption tests passed in isolation, then identified shared directories and overlapping KDC shutdown/startup as test hazards; Michael reproduced failures under parallel execution #202 #213 #214.
- Stephen's corrective patch documented GSS negotiation and framing, merged authentication and encryption tests around one KDC, added server-side pg_stat_gssapi checks, and corrected the libpq option label #215.
- Robbie checked the new documentation, Michael verified more stable parallel testing and requested presentation fixes, and Stephen applied those edits and reported pushing the follow-up and closing the open item #216 #217 #219.
- The thread ended with Michael confirming the committed follow-up looked satisfactory; the feature and reported follow-up fixes had landed, while richer encryption-algorithm reporting remained future work #220 #175.