[DONOTMERGE] PgSQL pool starvation / worker imbalance investigation - #6189

Open

renecannao wants to merge 20 commits into

v3.0from

fix/pgsql-pool-starvation

Commits on Sep 3, 2026

fix(pgsql): shorten poll timeout when starved but demonstrably busy
renecannao
committed
75fca7f
fix(pgsql): bypass the thread-local connection cache while waiters exist
renecannao
committed
0ebd380
fix(pgsql): reschedule a session that failed to get a backend connection
renecannao
committed
79d9b9a
fix(pgsql): drive retry interval and poll interval from one value
renecannao
committed
dadb52d
Revert "fix(pgsql): drive retry interval and poll interval from one value"
renecannao
committed
a534df4
Revert "fix(pgsql): reschedule a session that failed to get a backend connection"
renecannao
committed
5dc7f5b
fix(pgsql): retry failed pool checkouts in a second pass of the same scan
renecannao
committed
b465d37
TEMP: instrument the pool rescan
renecannao
committed
61c7215
TEMP: log which worker each idle thread hands sessions to
renecannao
committed
13a1d3b
EXPERIMENT: split idle-thread handoffs across two consecutive workers
renecannao
committed
9a57c3d
EXPERIMENT: idle thread gives the batch to the less loaded of two workers
renecannao
committed
426f3aa
EXPERIMENT: apply power of two choices to the MySQL idle-thread handoff
renecannao
committed
997e81d
EXPERIMENT: reinstate the B-band sort, rate limited to ~1 pass in 10
renecannao
committed
ab31f05
EXPERIMENT: rate limit the B-band sort by wall clock, not iteration count
renecannao
committed
f321fb4
EXPERIMENT: only fully order the front 10% of the B band
renecannao
committed
4dd829e
EXPERIMENT: drop the B-band sort interval to 20ms
renecannao
committed
9e254d9
EXPERIMENT: sort top 20% of the B band instead of 10%
renecannao
committed
331b843
Revert "EXPERIMENT: sort top 20% of the B band instead of 10%"
renecannao
committed
007df57
EXPERIMENT: back to top-10%, 50ms interval
renecannao
committed
ea8bb5b
EXPERIMENT: sort top 20% of the B band, at the 50ms interval
renecannao
committed
575f58a

Navigation Menu

Uh oh!

[DONOTMERGE] PgSQL pool starvation / worker imbalance investigation - #6189

[DONOTMERGE] PgSQL pool starvation / worker imbalance investigation#6189
renecannao wants to merge 20 commits into
v3.0from
fix/pgsql-pool-starvation

Commits on Sep 3, 2026

fix(pgsql): shorten poll timeout when starved but demonstrably busy

fix(pgsql): bypass the thread-local connection cache while waiters exist

fix(pgsql): reschedule a session that failed to get a backend connection

fix(pgsql): drive retry interval and poll interval from one value

Revert "fix(pgsql): drive retry interval and poll interval from one value"

Revert "fix(pgsql): reschedule a session that failed to get a backend connection"

fix(pgsql): retry failed pool checkouts in a second pass of the same scan

TEMP: instrument the pool rescan

TEMP: log which worker each idle thread hands sessions to

EXPERIMENT: split idle-thread handoffs across two consecutive workers

EXPERIMENT: idle thread gives the batch to the less loaded of two workers

EXPERIMENT: apply power of two choices to the MySQL idle-thread handoff

EXPERIMENT: reinstate the B-band sort, rate limited to ~1 pass in 10

EXPERIMENT: rate limit the B-band sort by wall clock, not iteration count

EXPERIMENT: only fully order the front 10% of the B band

EXPERIMENT: drop the B-band sort interval to 20ms

EXPERIMENT: sort top 20% of the B band instead of 10%

Revert "EXPERIMENT: sort top 20% of the B band instead of 10%"

EXPERIMENT: back to top-10%, 50ms interval

EXPERIMENT: sort top 20% of the B band, at the 50ms interval