model-writtenrun 2338 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May - Jun 2009 · 1 entry
- Robert Haas opened the discussion in 2009 because rounding average rows to integers prevented useful reconstruction of total tuples; Euler Taveira worried about confusing users, Joshua Tolley argued fractions would expose the averaging, and Tom Lane proposed showing decimals only when loops exceeded one, which Robert accepted #1 #2 #4 #5 #6.
Jun - Jul 2022 · 3 entries
- Ibrar Ahmed revived the TODO item in June 2022 with the first implementation, then corrected formatting after David G. Johnston's review #12 #13 #14.
- Amit Kapila and Vignesh C identified unrelated edits and missing regression updates; Ibrar removed those edits, updated partition-pruning expectations and suppressed .00 output to limit test churn #15 #16 #21 #23.
- Greg Stark questioned rounding tiny nonzero averages to zero, while Robert argued that displaying undivided totals would be clearer but that two decimal places were still worthwhile; Naeem Akhter and Hamid Akhtar subsequently reported successful testing #17 #18 #25 #26 #28.
Nov 2022 · 1 entry
- Tom Lane raised format-string, cross-platform stability and documentation concerns, and Andrei Lepikhov advocated scientific notation for very small averages; the patch eventually stalled and was returned with feedback in August 2023 #29 #30 #35.
Dec 2024 - Apr 2025 · 15 entries
- Ilia Evdokimov took over in December 2024 with a narrower implementation showing decimals only for averages below one and also changing loop formatting; Guillaume Lelarge's testing led to repaired test masking and documentation additions #36 #37 #38 #39 #41.
- After a February rebase, Matheus Alcantara questioned fractional loop counts and the below-one restriction, while Andrei Lepikhov and Robert Haas argued for preserving fractional averages above one; Ilia restored integer loop formatting and broadened row formatting, initially retaining suppression of .00 #45 #46 #48 #51 #52 #53.
- Alena Rybakina requested clearer motivation and documentation, and discussion settled on fixed two-digit precision as a practical improvement rather than variable significant digits or a switch to totals #54 #60 #61 #62.
- Ilia split documentation from implementation and proposed a non-equality join example; Andrei suggested replacing floor-based fraction detection with remainder arithmetic, producing an intermediate version that cast counters to int64 #64 #66 #67.
- Robert rejected the casts and extra fractional-part test because of precision and overflow concerns, and Ilia simplified the final submission to nloops > 1 while updating more regression expectations #68 #69.
- Andrei continued advocating output that distinguished varying tuple counts, but Ilia considered additional instrumentation or min/max reporting outside the current scope #70 #71 #72.
- Robert committed the implementation and tests from #69 unchanged as ddb17e387a, leaving out the documentation rewrite because he considered the existing explanation still accurate #73.
- Buildfarm failures immediately exposed nondeterministic formatting when slow parallel workers left nloops equal to one; Robert reported the failures and pushed a temporary test workaround, while Tom confirmed intermittent failures and suggested formatting according to potential looping #74 #75 #77 #78.
- The failures also prompted a separate discussion about weekend reverts, committer workload and pre-commit buildfarm testing, with Robert, Andres Freund and Jelte Fennema-Nio arguing for less disruptive handling of small test-output failures #76 #79 #81 #83.
- Ilia reduced the remaining documentation proposal to a one-line example correction, while Robert proposed always displaying two decimal places to eliminate dependence on parallel-worker timing #84 #86.
- Ilia supplied seven-file documentation updates for Robert's new formatting rule, and Robert agreed to combine their patches #87 #88.
- Robert landed the combined code, tests and documentation from #86 and #87 as 95dbd827f2, removing the temporary test workaround and making actual-row precision unconditional; the separate one-line correction from #84 is absent from this commit's diff #90.
- After the commitfest entry was marked committed, Alena reported confusion over rows=0.50, Matthias van de Meent explained averages per execution, and Alena confirmed that she had initially mistaken the value for a total #91 #92 #93 #95.
- Ilia later questioned integer extraction in the extended-statistics test helper, but Robert saw no immediate need to change it and instead raised fractional estimated rows; Tom explained that the planner itself rounds estimates, and discussion favored retaining the minimum estimate of one #98 #99 #100 #101 #102.
- The thread ends with Ilia's new, uncommitted series preserving fractional planner row estimates, separating integral tuple/page estimates through clamp_tuple_est, changing EXPLAIN estimate formatting and updating tests; community agreement and documentation remain outstanding #104.