Hi Hackers,

When a subscription has retain_dead_tuples enabled with maxretention set
to zero (unlimited retention), adjust_xid_advance_interval() caps
xid_advance_interval to Min(interval, maxretention).  Since maxretention
is zero, this always collapses the interval to zero milliseconds.

A zero makes  TimestampDifferenceExceeds(last_time, now, 0)  always
true in get_candidate_xid(). This causes the apply worker to call
GetOldestActiveTransactionId() on every single WAL message. This results in
a huge number of ProcArrayLock acquisitions under moderate write load.

Fix by adding a maxretention > 0 guard to the cap. When maxretention is
zero ,
the exponential back-off in adjust_xid_advance_interval()
now works correctly, growing the interval from 100 ms toward the 180 s
ceiling.

Measured with perf uprobe counting GetOldestActiveTransactionId calls
at ~39K TPS (pgbench, 5 clients):

  Before fix: 25,104 calls / 5 s  (~5,021/s)
  After fix:     31 calls / 5 s  (~6/s)

Thank
Satya

Attachment: 0001-Fix-apply-worker-busy-loop-when-subscription-max_ret.patch
Description: Binary data

Reply via email to