FYI - Here are some more counter results with/without the skip patch
[1] applied.
This is the same test setup as before except now using *synchronous* pub/sub.
//////////
Test setup:
- using synchronous pub/sub
- subscription is for the pgbench_history table
- pgbench is run for 10 seconds
- config for all the wal_sender/receiver timeout GUCs are just default values
WITHOUT the skip-first patch applied
=====================================
RUN #1
------
LOG: Total records: 310
LOG: 24: 49 / 131 / 49: 8403 / 9270
LOG: 944: 1 / 0 / 1: 159693904 / 159693904
LOG: 8280: 1 / 2 / 2: 480 / 480
RUN #2
------
LOG: Total records: 275
LOG: 24: 45 / 129 / 46: 8580 / 8766
LOG: 5392: 1 / 0 / 1: 160107248 / 160107248
RUN #3
------
LOG: Total records: 330
LOG: 24: 50 / 144 / 51: 8705 / 8705
LOG: 3704: 1 / 0 / 1: 160510344 / 160510344
LOG: 8280: 1 / 2 / 2: 468 / 468
WITH the skip-first patch applied
=================================
RUN #1
------
LOG: Total records: 247
LOG: 24: 5 / 172 / 44: 3601700 / 3601700
LOG: 8280: 1 / 1 / 1: 1192 / 1192
RUN #2
------
LOG: Total records: 338
LOG: 24: 8 / 199 / 55: 1335 / 1335
LOG: 7597: 1 / 1 / 1: 11712 / 11712
LOG: 8280: 1 / 1 / 2: 480 / 480
RUN #3
------
LOG: Total records: 292
LOG: 24: 4 / 184 / 49: 719 / 719
//////////
As before there is a big % reduction of keepalives after the patch,
except here there was never really much of a "flood" in the first
place.
------
[1] https://www.postgresql.org/message-id/CAHut%2BPtyMBzweYUpb_QazVL6Uze2Yc5M5Ti2Xwee_eWM3Jrbog%40mail.gmail.com
Kind Regards,
Peter Smith.
Fujitsu Australia