model-writtenrun 9627 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- David Rowley opened the thread by describing autovacuum workers unable to keep up under the longstanding cost limit of 200 and proposed a fivefold or tenfold increase, leaving adaptive throttling for a larger project #1.
- After support for faster vacuuming and suggestions to revisit scale factors, David Rowley deliberately limited his patch to raising vacuum_cost_limit to 2000 for PostgreSQL 12 #4 #5.
- Tomas Vondra questioned the effect on small systems and changing only one conservative default, while Robert Haas argued that vacuum starvation justified addressing this setting independently #11 #12.
- Andrew Dunstan committed David Rowley's four-file patch unchanged as bd09503e63 and confirmed the landing #5 #17.
- Jeff Janes then questioned the hard maximum of 10000, prompting Tom Lane to identify overflow hazards in both accumulated costs and the delay calculation if the limit approached INT_MAX #20 #24.
- Tom Lane reconsidered the approach and proposed reverting the increase in favor of shorter, more frequent sleeps, arguing that modern platforms could support finer delays and smoother I/O demand #25.
- David Rowley warned that changing the implicit delay unit to microseconds would break existing scripts and suggested a future throttling redesign; Tom Lane instead chose floating-point milliseconds with unit support to preserve existing configuration values #26 #28 #29.
- Tom Lane posted two patches adding units to floating-point GUCs, accepting microseconds, converting vacuum delays to floating point, and reducing the autovacuum default to 2ms; he explicitly left the required cost-limit revert, fractional integer input, NaN rejection, and any maximum-limit increase outside the attachments #33.
- Tom Lane separately proposed including the converted base unit in range errors, while Julien Rouhaud questioned the absence of a microsecond base unit for extensions; Tom declined that extension without a demonstrated need #35 #36 #38.
- The NaN issue identified in #33 became the separately backpatched fix ac75959cdc, and the error-message proposal in #35 became 28a65fc360.
- Tom Lane committed the floating-point infrastructure from #33 as caf626b2cd, adding a catalog-version bump and more tolerant output-unit selection while separating out the default change.
- Tom Lane then reverted the cost-limit increase as 52985e4fea and extracted the 2ms default from #33 into cbccac371c, implementing the shorter-delay approach proposed in #25.
- Tom Lane next posted fractional input support for integer GUCs together with an optional rounding step that made inputs such as 30.1GB display in more readable units; David Rowley and Julien Rouhaud supported that readability benefit #40 #41 #42.
- Both fractional-input attachments from #40 landed together as 1a83a80a2f, with additional documentation explaining the rounding rule.
- After that commit, Kyotaro Horiguchi objected that rounding 0.1 days to 2 hours materially changed the requested interval and supplied an alternative that shortened the displayed representation without rounding the stored value #43.
- Kyotaro Horiguchi immediately replaced his mistakenly attached prototype with the intended shortest-representation implementation; the thread ends with that alternative unresolved, while the delay changes and input-rounding implementation have landed and no maximum-cost increase or adaptive throttling redesign is included #44.