model-writtenrun 7169 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread with differential coverage results comparing REL_16_STABLE against HEAD, identifying gaps across several subsystems, including the bump allocator, and explaining how immediate shutdowns could lose or corrupt coverage data #1.
- Peter Geoghegan explained that targeted B-tree array-scan tests had been omitted because they required relatively expensive, implementation-dependent fixtures, while existing assertions already checked related invariants #2.
- Robert Haas outlined possible WAL-summary and backup tests and identified unused block-reference code; Andres Freund encouraged basic reachability checks and additional coverage of WAL record types and summarization toggling, without a patch landing for those topics in this thread #3 #4.
- David Rowley submitted a large-datum tuplesort test to exercise BumpAllocLarge(), while arguing that intentionally unreachable allocator error callbacks should remain unchanged #5.
- Andres Freund suggested exercising BumpStats() by keeping a sort alive through a cursor and querying pg_backend_memory_contexts, and also raised the absence of memory-context type information from that view #6.
- Jeff Davis explained that Unicode coverage existed under update-unicode; Andres Freund requested ordinary test-suite coverage, while ICU version mismatches complicated exhaustive tests and Tom Lane ruled out internet access from standard tests #7 #8 #9 #10.
- David Rowley extended his patch with the cursor-based statistics test, omitted platform-dependent values, and reduced its second datum to 512 bytes; the resulting three-block allocation exposed inefficient insertion of dedicated large blocks at the list head, and he reported a separate patch for displaying context types #11.
- Andres Freund approved the expanded tests and suggested positive-value checks for size statistics; David Rowley retained the exact block count to support a future allocator regression check and committed the patch with light comment, formatting, and assertion refinements as bea97cd02e #12 #13.
- The Unicode discussion converged on adding ordinary spot checks while retaining exhaustive ICU comparisons for Unicode updates, and Jeff Davis offered to submit a test module; no such attachment appears in this thread #14 #15 #16.
- David Rowley's allocator improvement discussed in #11 and #13 subsequently acquired its matching implementation patch in #560286 in thread #48070, which is the source of 6d2fd66b99; the commit moved dedicated blocks to the list tail, reduced the test's expected block count from three to two, and cleaned up unsupported-callback diagnostics and comments.
- More than a year later, Alvaro Herrera revived the thread after upgrading the coverage server, asking for the differential-report scripts and reporting broken links in the original reports #17.
- Nazir Bilal Yavuz shared automation developed with Andres Freund for comparing master against the latest stable branch, with Meson and Make variants, but reported an unresolved undercount of uncovered lines; the coverage tests from this thread and allocator improvement sourced from thread #48070 had landed, while coverage-report automation remained work in progress at the thread's close #18.