On Fri, Mar 17, 2023 at 12:20 AM Bharath Rupireddy
<bharath.rupireddyforpostgres@gmail.com> wrote:
> +1 for pg_get_wal_block_info emitting per-record WAL info too along
> with block info, attached v2 patch does that. IMO, usability wins the
> race here.
I think that this direction makes a lot of sense. Under this scheme,
we still have pg_get_wal_records_info(), which is more or less an SQL
interface to the information that pg_waldump presents by default.
That's important when the record view of things is of primary
importance. But now you also have a "block oriented" view of WAL
presented by pg_get_wal_block_info(), which is useful when particular
blocks are of primary interest. I think that I'll probably end up
using both, while primarily using pg_get_wal_block_info() for more
advanced analysis that focuses on what happened to particular blocks
over time.
It makes sense to present pg_get_wal_block_info() immediately after
pg_get_wal_records_info() in the documentation under this scheme,
since they're closely related. It would make sense to explain the
relationship directly: pg_get_wal_block_info() doesn't have the
block_ref column because it breaks that same information out by block
instead, occasionally showing multiple rows for particular record
types (which is what its "extra" columns describe). And,
pg_get_wal_block_info() won't show anything for those records whose
block_ref column is null according to pg_get_wal_records_info(), such
as commit records.
(Checks pg_walinspect once more...)
Actually, I now see that block_ref won't be NULL for those records
that have no block references at all -- it just outputs an empty
string. But wouldn't it be better if it actually output NULL? Better
for its own sake, but also better because doing so enables describing
the relationship between the two functions with reference to
block_ref. It seems particularly helpful to me to be able to say that
pg_get_wal_block_info() doesn't show anything for precisely those WAL
records whose block_ref is NULL according to
pg_get_wal_records_info().
--
Peter Geoghegan
