The Accounts Payable group ran ~9.5 min, 83% of it in four full
reloads. The cost is the GPSERVER linked-server hop, not row volume
(pm30200 managed only ~590 rows/sec), so cutting rows crossing the link
is the whole win. Predicates go inside OPENQUERY to push down to the
remote — a 4-part name would drag the table across before filtering.
pm30200 own DEX_ROW_TS
pm30600 changed set: parent PM30200
pm30300 changed set: PM30200 union PM20000
All three key on (vchrnmbr, doctype), verified exactly unique in
PM30200 at 108,803/108,803. Watermark resolves off gp.pm30200 with a
7-day lookback, so run order within the group must stay pm30200 first
and the lookback must exceed the sync interval.
The two-parent union on pm30300 is load-bearing: 38 rows have a parent
only in the open table, and a PM30200-only join would strand them
permanently. pm30600 needs no union (0 orphans, verified).
pm30700 stays full — 213 of its rows have no parent in either table, so
no changed-set join can reach them, and it only costs 13s. pm20100 and
pm00400 stay full because open and index tables delete rows, which a
delete-by-key incremental structurally cannot propagate (cf. icstt).
391s -> 15s. reconcile.py --super-quick reports IN SYNC for all three.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Paul Trowbridge
b855ab91d7