Improve comment on why we need ctid->(cmin,cmax) mapping.

Combocids are only part of the problem. Explain the problem in more detail.

Reviewed-by: Andres Freund
Discussion: https://www.postgresql.org/message-id/1ba2899e-77f8-7866-79e5-f3b7d1251a3e@iki.fi

1 files changed, 9 insertions, 4 deletions

diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 0786bb0ab71..e403feeccda 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -41,10 +41,15 @@
  * transactions we need Snapshots that see intermediate versions of the
  * catalog in a transaction. During normal operation this is achieved by using
  * CommandIds/cmin/cmax. The problem with that however is that for space
- * efficiency reasons only one value of that is stored
- * (cf. combocid.c). Since combo CIDs are only available in memory we log
- * additional information which allows us to get the original (cmin, cmax)
- * pair during visibility checks. Check the reorderbuffer.c's comment above
+ * efficiency reasons, the cmin and cmax are not included in WAL records. We
+ * cannot read the cmin/cmax from the tuple itself, either, because it is
+ * reset on crash recovery. Even if we could, we could not decode combocids
+ * which are only tracked in the original backend's memory. To work around
+ * that, heapam writes an extra WAL record (XLOG_HEAP2_NEW_CID) every time a
+ * catalog row is modified, which includes the cmin and cmax of the
+ * tuple. During decoding, we insert the ctid->(cmin,cmax) mappings into the
+ * reorder buffer, and use them at visibility checks instead of the cmin/cmax
+ * on the tuple itself. Check the reorderbuffer.c's comment above
  * ResolveCminCmaxDuringDecoding() for details.
  *
  * To facilitate all this we need our own visibility routine, as the normal